# Yellow Paper Daisies A product blog on the safe, secure and beneficial deployment of AI. I write product ideas about how AI could do some good, and about the testing, oversight and accountability mechanisms it needs to do so safely and securely. Each idea borrows its metaphor from one of the more than 20,000 species of the daisy family. I believe safety, security and benefit can be achieved together, in that order of priority. AI’s most serious risks deserve careful, sustained attention, and its benefits only last when the systems behind them are safe, secure and trusted. So some ideas here explore what AI could do for people, and others explore how to test, govern and oversee it so that good can happen safely and securely. Every idea follows the same product-thinking structure: context, who it is for, the problems they face, a possible solution and the outcome it aims for. Many are illustrative concepts that can't be built yet. I share them to think in the open and learn from feedback. All ideas are my own and do not represent the views of any current or former employer, client or organization I am affiliated with. They are illustrative proposals for discussion, not professional, legal or financial advice. AI helps with writing, research and the drawings. # The ideas (26) ## Idea 1 · DaisyChain - Date: 2026-09-25 - Address: https://yellowpaperdaisies.org/idea/2026-09-25-daisy-chain - Flower: Common daisy, Bellis perennis *Daisy Chain is an AI-powered giving platform that verifies needs, matches donors with causes they care about, and shares anonymized outcomes on a community wall.* ## Context & Motivation Many people want to make a difference but find it hard to know whether their contributions reach the intended cause. Many prefer giving directly and seeing the results. Voice AI, document verification, and mobile payments now make this possible at scale. ## Target Audience Professionals living abroad who want to support communities in their home countries. Local schools and clinics provide the first requests, with concrete, easy-to-understand needs. ## Customer Problems Donors want assurance that a need is genuine, that funds are used as intended, and that they'll see the result. They also want to find causes matching their values. Recipients deserve privacy and dignity. ## Compelling Solutions ### For people in need A person describes their situation by voice in any language, through a website or a familiar messaging app. An AI agent helps shape it into a clear request, checks the amount against local prices, and screens for duplicate or AI-generated content. Requests are prioritized by urgency and impact, with fairness reviews and published criteria. ### For donors Donors share the causes they care about, and the agent connects them with matching needs. They can set a monthly amount that is spread across verified needs, and small gifts from many donors are pooled to fully fund a request. Where possible, funds go directly to the school, clinic, or pharmacy. Donors receive a monthly impact report and can ask the ledger how their contribution was used. ### For the community Completed donations appear on a community wall, with AI anonymizing names and photos. Follow-up check-ins confirm lasting impact, and patterns across requests help local partners plan group purchases. Automated paperwork keeps overhead low. ## Outcome Daisy Chain helps people see and meet each other's needs directly and verifiably through AI, so genuine needs never go unseen and no one who wants to help is left unsure, making giving contagious and woven into how communities care for one another. ### Why a daisy chain ### The craft: the daisy chain A daisy chain is made by slitting the stem of one daisy with a thumbnail and threading the next stem through it, link by link, until the garland closes. Every link depends on the one before it. ### The flower: common daisy *Bellis perennis.* Its English name comes from “day’s eye”: the flower head opens to the light and closes at dusk and in dull weather. ### The metaphor A gift is a chain too. It holds only if every link can be checked: the request, the verification, the match, the direct payment and the report back. Like a daisy in daylight, each step is meant to be plain to see. ### How to read the plate - **The centre.** The common daisy, traced from above. - **The ring.** 28 small daisies, one for each ray of the outer ring, joined stem to stem. - **The strip.** The five steps of a gift in Daisy Chain. ## Idea 2 · PetalPrint - Date: 2026-09-26 - Address: https://yellowpaperdaisies.org/idea/2026-09-26-petalprint - Flower: Blanket flower, Gaillardia pulchella *An optional feature that lets you teach your AI assistant your values and way of thinking, giving you greater control to configure it around them.* ## Context & Motivation The people who know us best understand how we think, care about our wellbeing, and tell us hard truths. AI assistants are now everyday thinking partners, yet they share one set of default values and know us only through facts. Petalprint lets people shape those values within core safety and honesty principles, making their assistant more human without ever replacing the people in their lives. ## Target Audience Adults who want a personalized, curated AI assistant with character, starting with people whose cultures and traditions are least reflected in AI defaults. ## Customer Problems Value lists like "honesty" don't guide real outputs, and written instructions cover only anticipated situations. People want an assistant that understands them, looks out for their wellbeing, stays honest, and leaves them in full control. ## Compelling Solutions ### Teaching A ten-minute voice or text conversation invites life stories through prompts like "Tell me about a time when...", then follows up to understand your reasoning. Questions, written by people from diverse backgrounds, start light and deepen only as trust builds, and any can be skipped. Your stories form a readable Petalprint of your thought patterns, shaping the priorities, never the facts, behind everything your assistant creates. You choose where it mirrors you and where it offers a counterweight, and requests to hide mistakes are gently and transparently reworded, because care requires honesty. ### Growing Ask any question, compare its reasoning with yours, and teach it the difference, or change your own mind. An optional daily prompt, like "Tell me about a time you felt truly proud," lets your assistant reflect back your values and suggest additions. A toggle removes personalization from any answer, and prompts to share stories with loved ones keep people at the center. Effort compounds: time invested early yields ever more aligned answers. ### Trust Off by default. Nothing is added without your approval or shared with third parties, and your Petalprint can be switched off or permanently and verifiably erased anytime. ## Outcome As AI increasingly shapes how people think and write, Petalprint provides the optionality to preserve the diversity of human values in AI, so people of every culture and tradition are understood as themselves rather than nudged toward a single default. ### Why a print ### The technique: nature printing In the 1850s, printers in Vienna and London pressed real plants into soft lead plates to make printing plates, so the plant itself became the image. Alois Auer of the Imperial Printing Office in Vienna patented the process in 1852, and Henry Bradbury used it in London for *The Ferns of Great Britain and Ireland*. ### The flower: blanket flower *Gaillardia pulchella.* Each ray ends in three teeth, and its colour shifts from red to yellow along the ray. ### The metaphor Petalprint works the same way. You do not design your assistant from above; it takes an impression of how you think from your own stories, the way the plant makes its own plate. And because an impression is reversed and partial, you stay in charge of it: it is off by default, and you can switch it off or erase it. ### How to read the plate - **Top.** The blanket flower, traced from above. - **Below left.** The plant, as the plate. - **Below right.** Its impression, in ochre and reversed. ## Idea 3 · CommonGround - Date: 2026-09-27 - Address: https://yellowpaperdaisies.org/idea/2026-09-27-common-ground - Flower: Dandelion, Taraxacum officinale *CommonGround is an illustrative concept for an AI-assisted mutual aid circle where neighbours trade everyday help, fueled entirely by goodwill, public appreciation, and shared community connection.* ## Context & Motivation Everyday hurdles like reading fine-print letters, assembling flat-pack furniture, or navigating device settings may take only minutes to solve, yet can create significant friction for some. At the same time, valuable skills, life experience, and spare hours across every neighbourhood often sit idle. Help once flowed organically through extended families and close-knit communities. Today, geographical distance and social fragmentation can leave many people isolated. Voice-first AI, assistive accessibility tools, and intelligent matching make it possible to thoughtfully recreate this decentralized support network at community scale, without introducing financial barriers. ## Target Audience CommonGround would be piloted strictly within bounded, high-density environments, such as a single housing cooperative, community centre network, or library district, where authentic local familiarity can develop. Initial cohorts would prioritize intergenerational pairings between older adults and local students or newcomers. Both groups would act as mutual contributors: an older resident might receive help moving heavy boxes while coaching a student on job interviews or sharing traditional recipes. Trusted neighbourhood anchors would vouch for early participants and host initial orientation sessions. ## Customer Problems 1. **Discovery & Match Complexity:** Needs and skills may exist only blocks apart, but direct one-to-one barter rarely aligns on timing, proximity, and mutual interest. 2. **The Pride & Dignity Barrier:** Many people are readily willing to volunteer their time but hesitate to ask for help, fearing they are imposing or accepting charity. 3. **Trust, Liability & Vulnerability:** Inviting a newcomer into a private space requires safety reassurance, while older participants may face heightened risks of financial exploitation. 4. **Scope Creep & Over-Reliance:** Users may voice acute medical, caregiving, or psychological emergencies that untrained peer neighbours cannot safely handle. ## Solutions 1. **Voice-First Triage & Scope Guardrails:** Members describe offers and needs conversationally, in their preferred language. The system routes acute clinical or caregiving needs to verified municipal or community resources, passing only appropriate, everyday assistance into the peer network. 2. **Coordinator Copilot:** Rather than relying on cold autonomous dispatch, AI drafts candidate pairings for community coordinators at local libraries or centres, who approve introductions with one touch. 3. **Gratitude Circles & The Thank-You Wall:** Balance sheets and currency are replaced entirely by public recognition. When help is completed, recipients can post warm notes of appreciation on a shared community wall. 4. **Progressive Safety & Fraud Protection:** Onboarding enforces a strict 'No-Cash, No-Credentials' rule to reduce the risk of financial exploitation. Interactions graduate through stages: initial meetups occur in public partner venues, two-sided feedback remains blind until both participants submit it, and street addresses remain masked until a pairing is mutually confirmed. 5. **Sustainable Operations:** Bounded pilots operate with clear terms, third-party liability insurance, and task screening, supported through civic grants and nonprofit sponsorships. ## Outcomes A. Small acts of practical kindness build relational community wealth alongside gig work. B. Reduced self-reported isolation and high completion of low-risk tasks. C. Active participation across newcomers and long-time residents. D. Pure gratitude requires a high-trust society; safeguarded steps toward neighbourly reciprocity help us move there. *Note: Replacing transactional ledgers with pure gratitude requires us to operate in a high-trust society. While our communities may not be completely there yet, building thoughtful, safeguarded steps toward neighbourly reciprocity gives us reason to believe we can move in that direction.* ### Why a dandelion ### The flower: dandelion *Taraxacum officinale.* It grows almost anywhere people do. Every floret in its yellow head is strap-shaped, and each one sets a single seed. After it flowers, the head becomes a clock of seeds, and every seed hangs from its own tiny parachute, called a pappus: about 100 fine hairs that are mostly empty space. ### The metaphor Neighbourly help travels the same way: light, given freely, and landing close to home. Most dandelion seeds likely land within 2 metres of the plant. On the plate, six seeds drift to the plots of one block, which is where CommonGround would start: a housing co-op, a community centre network, a library district. ### How to read the plate - **The clock.** The seed head, seen whole. Gaps show seeds that have already left. - **The flight.** Six seeds, each hanging from its parachute. - **The block.** Eight plots, with a house on each. Six seeds land, six small acts of help. - **Stage 1.** The same head, in flower, before the seeds. - **One seed.** Enlarged: the parachute, the slender stalk and the ribbed seed. ### Sources - [Read the study on how the dandelion's parachute flies](https://strathprints.strath.ac.uk/79669/) ## Idea 4 · EverlastingDaisy - Date: 2026-09-28 - Address: https://yellowpaperdaisies.org/idea/2026-09-28-everlasting-daisy - Flower: Golden everlasting, Xerochrysum bracteatum *Everlasting is an illustrative concept for a community-owned AI Governance product that lets communities govern their generational knowledge and, if they choose, sustain it through a circular economy.* ## Context & Motivation Communities have long passed knowledge down through teachings, prayers, stories and art, yet many children now grow up in technology carrying little of their heritage. As elders pass on, that knowledge can fade within a generation. Local AI now lets communities govern their own knowledge, alongside traditional ways, without outside servers. ## Target Audience Communities with deep generational knowledge, such as Indigenous nations, diaspora, faith and rural communities, starting with those ready to lead. Elders hold the knowledge, young people carry it forward, and supporters are partners. ## Customer Problems A community council knows its elders hold knowledge no book has captured. Young people are moving away or absorbed in screens. Someone in the wider community has money and goodwill but doesn't know what to fund, whom to trust or what a first step looks like. The council doesn't have technical staff and is wary of anything that puts its knowledge in someone else's hands. An elder would gladly share with a grandchild but not with a stranger's server. The problem then is governing how to carry knowledge forward, control what stays sacred, help youth see themselves in technology, while keeping the benefits of it locally. ### Hypothesis A. **Time:** elders and youth rarely share protected time. B. **Control:** no safe way to record knowledge without losing control of it and how it is used. C. **Fidelity:** AI always answers, even when wrong or never meant to be shared. D. **Capacity:** few communities have technical staff; supporters lack a trusted way to help. E. **Benefit:** the value knowledge creates often flows away. ## Solutions A. **Governance first:** the community decides who may access each item; sacred knowledge stays private. The archive is community-owned and revocable. B. **Ask, don't guess:** the assistant answers only from approved sources and holds questions it is unsure about. C. **Keeper Circle:** community-chosen reviewers, like security teams, examine held questions. Elders answer them, and decide sacred matters alone. D. **Trust and funding:** a local steward guides the work, independent checks confirm the process, and supporters fund without claiming ownership. E. **Circular economy:** by collective choice, communities can turn knowledge into art and learning tools. Income stays local, paying elders, keepers and youth, funding more preservation in a self-sustaining cycle. ## Outcome A. Preserving generational knowledge while its holders can share it. B. Knowledge, and decisions about it, remain with those who carry it. C. Expanding the diversity of technology and technology-enabled people, so AI reflects many ways of knowing. D. A circular economy where AI-enabled knowledge management as a service returns value to communities. E. Generational knowledge preservation mechanisms created across communities can also inform AI governance frameworks that are useful beyond them. ### Why an everlasting daisy, and why a hand ### The art form: hand stencils A hand stencil is made by pressing a hand to rock and blowing pigment around it through a hollow bone, so the hand stays bare and the colour surrounds it. The best-known gallery is Cueva de las Manos, the Cave of the Hands, in Santa Cruz Province, Argentina, listed by UNESCO in 1999. UNESCO dates its cave art to between 13,000 and 9,500 years ago and says the makers may have been ancestors of the hunter-gatherer communities of Patagonia. Most of the hands are left hands. The technique is far older and wider than one cave: the oldest dated example, from Liang Metanduno on Muna Island in Indonesia, is at least 67,800 years old. ### The flower: golden everlasting Xerochrysum bracteatum is a daisy-family plant native to Australia. What look like petals are papery bracts made of dead cells; the flower has no ray florets. Inside them sit a hundred or more tiny florets. Before the flower opens, the bracts curl over and shield them. ### The metaphor You can see this art form on this website only because it has been digitized: photographed, traced and shared. Even the oldest known stencil, on Muna Island, is mostly worn away, and researchers used digital tracing to see it. Everlasting does the same for living traditions. It lets communities record their art forms and practices on their own terms, so that they keep flowing between generations and keep finding new expression. The community decides if and when any of it is shared or monetized, and what stays sacred, just as the hand in a stencil stays bare while the colour goes around it. In this way AI becomes an enabler of traditional art forms, and a promoter of their diversity. ### How to read the plate - **Plan.** The flower seen from above. 89 bracts, each placed 137.5° after the last, the golden angle that daisy-family heads follow when they pack their parts. - **Elevation and Section A–A.** The same flower from the side: the left half is the outside, the right half is cut through the centre. Both come from one model, so their points line up. - **The disc.** The species has 100 or more florets in the centre. 144 are drawn. - **The pencil quarter.** Left unfinished on purpose. The idea is still taking shape. - **The halo and the hands.** Pigment blown around the flower, the way a stencil is made, and a panel of hand stencils. It interprets the technique; it does not copy any wall. ## Idea 5 · DaisyCheck: Benchmarks for Everyday AI - Date: 2026-09-29 - Address: https://yellowpaperdaisies.org/idea/2026-09-29-daisycheck - Flower: Ox-eye daisy, Leucanthemum vulgare *DaisyCheck is an open evaluation service that provides a simple nutrition label and routine bias checkup, helping everyday people see which cultural and national perspectives shape their AI models.* ## Context & Motivation AI models are black boxes whose answers subtly mirror the national priorities and training biases of their origins. While enterprise labs develop complex evaluations, everyday users have no accessible way to understand how their tools think. AI Bias has evolved from ideological text (e.g., repeating state propaganda or censoring historical events) to functional sabotage (e.g., deliberately lowering security standards or injecting flaws for an opposed geopolitical entity). If a model's training data or alignment is heavily influenced by a specific priorities of the builder, it can subtly weaponize the code it generates against perceived adversaries. As the first step in the right direction, just as browsers and social media introduced privacy checkups, AI needs transparent, consumer-facing evaluations to increase awareness and transparency. ## Target Audience Privacy-aware citizens, digital rights advocates, and independent builders who refuse to let opaque technology subtly steer their worldview or compromise their work. ## Customer Problems A. **The Black Box:** Users cannot see if a model's answers reflect universal consensus or a specific agenda. B. **Functional Sabotage:** Biases now manifest as degraded security or skewed logic depending on the user's geolocation or the geopolitical context of the prompt. C. **Flawed Benchmarks:** Current public tests measure math and coding speed, entirely ignoring civil liberties, privacy, functional sabotage, and geopolitical tilt. ## Solutions A. **Perspective Nutrition Label:** A pre-chat scorecard revealing a model’s geographic, cultural, and ideological leanings. B. **Routine Bias Checkups:** Two-minute diagnostic prompts that test a model across nuanced global issues and civil rights, and flagging blind spots. C. **The Apples-to-Apples Index:** An open registry ranking models side-by-side on human rights and ideological neutrality, empowering users to choose tools aligned with their values. ## Outcome DaisyCheck turns abstract safety debates into practical consumer literacy, benchmarks and a common nomenclature for evaluating models. By demystifying model bias, it builds trust and empowers individuals to select the thinking partners that serve them best. ### Why a measured daisy ### The ritual: he loves me, he loves me not Checking a daisy one ray at a time is one of the oldest home-made evaluations: pick a ray, say a line, pick the next, and see where you end. ### The flower: ox-eye daisy *Leucanthemum vulgare.* A wild daisy: a ring of white rays round a yellow disc. ### The metaphor A checkup measures against a stated ideal and reports every departure openly. This plate measures one daisy against a perfectly even ring of 21 rays, marks where each ray really sits, and states the facts in a label. DaisyCheck would do the same for an AI model. ### How to read the plate - **The ring.** A protractor around the daisy, marked in degrees. - **Cream ticks.** Where each ray would sit on an even ring. - **Yellow carets.** Where each ray is drawn. - **The label.** Measured on this drawing: the largest departure from even spacing is 4.8°. It is a drawing, not a specimen. ## Idea 6 · EcoDaisy - Date: 2026-09-30 - Address: https://yellowpaperdaisies.org/idea/2026-09-30-petal-pause - Flower: Gazania, Gazania spp. *Product feature for Large Language Models to minimize environmental impact by guiding users toward low-compute alternatives, routing simple queries transparently to more efficient models, and tracking resource savings on an auditable dashboard.* ## Context & Motivation AI models consume vast unseen resources, with energy and water usage varying wildly based on the task, the model size, and the local power grid. While everyday users increasingly want to make environmentally mindful choices, the invisible footprint of a single prompt makes this impossible. The environmental cost of AI scales rapidly with widespread adoption. Using massive, resource-heavy models for simple tasks can be wasteful for the environment when multiplied across the millions of users. With smaller models, saved answers, and smart routing, we can seamlessly right-size compute power to match the complexity of the request without sacrificing quality. As the first step in the right direction, just as appliances and vehicles introduced energy efficiency ratings, AI needs transparent, opt-in mechanisms to help users understand, measure, and minimize their digital carbon footprint. ## Target Audience Sustainability-minded individuals, enterprise organizations requiring credible footprint reporting, and providers who want to offer environmentally responsible technology options. ## Customer Problems A. **Awareness:** Users are not aware of queries that may be more efficiently answered without using AI, that consume comparatively lower compute resources. B. **The Invisible Cost:** When use of AI is necessary, users cannot see or weigh the environmental footprint of their prompts, making mindful digital habits impossible to build. C. **Resource Mismatch:** When AI is used, resource-heavy models are routinely deployed for simple, lightweight queries that could be easily handled by smaller, far more efficient sources. D. **Unverifiable Reporting:** Organizations lack credible, auditable ways to measure, account for, and ultimately reduce the environmental impact of their enterprise AI use. ## Solutions A. **In-product Tool-tips:** If a query may be answered by an existing article, using a calculator, or through other deterministic/ retrieval mechanisms, request the user to reconsider their use of AI for such queries (e.g. calculations) B. **Right-Sized Routing:** Provide the user the option to automatically direct simple requests to lighter models or saved answers behind the scenes, while transparently informing the user for each query where a lighter model is used. C. **Footprint Dashboard:** A transparent view providing users and organizations with verifiable usage data, savings (e.g. water or electricity saved), estimated resource ranges, and auditable sustainability metrics. ## Outcome This framework turns abstract sustainability goals into practical consumer literacy and verifiable enterprise reporting. By demystifying the environmental cost of AI, it builds trust, significantly reduces unnecessary energy and water consumption, and empowers individuals to make mindful digital choices without sacrificing capability. ### The daisy behind the idea ### The flower: gazania Gazania, often called the treasure flower, belongs to the daisy family. Horticultural guidance rates its drought tolerance as high and notes that it needs less water than most annual or perennial garden flowers. Its flowers close at night and on very cloudy days, and open fully in the sun. Its leaves are often silvery underneath. [see reference 1 and 2 below] ### The metaphor Just as the gazania thrives on less water and folds its petals when the sun fades, we must right-size our compute unfolding the full power of AI only when the task demands it, scaling back when simpler solutions suffice, and sustaining progress with a lighter footprint. ### How to read the plate - **Left half.** The rays drawn as geometry: straight lines, evenly spaced. - **Right half.** The same rays drawn as they grow, with natural curves. - **The disc.** The disc florets at the core, set in a spiral. - **Open, closed and leaf.** In full sun the rays spread wide; at night and on very cloudy days they fold up. The spoon-shaped leaf is hatched on its silvery underside. ### Sources [1] University of Florida IFAS Extension, “Gazania spp. Gazania,” fact sheet FPS-225: [ask.ifas.ufl.edu/publication/FP225](https://ask.ifas.ufl.edu/publication/FP225). [2] Gardenia.net, “Gazania (Treasure Flower)”: [gardenia.net/plant-variety/gazania-treasure-flower](https://gardenia.net/plant-variety/gazania-treasure-flower). ## Idea 7 · DaisyGrove - Date: 2026-10-01 - Address: https://yellowpaperdaisies.org/idea/2026-10-01-daisygrove - Flower: Colt's-foot, Tussilago farfara *A framework where each data centre repays its carbon debt with a community-chosen forest, grows at the pace that forest repays, and shares compute with climate-exposed communities.* ## Context & Motivation One promise of AI is that it could help reverse the damage greenhouse gases are doing to our planet. But AI runs on data centres, which have a carbon footprint of their own. This idea comes from investigating how much carbon the data centre build-out could add over the next couple of decades, and what it would take to stop further damage before we think of reversing it. A typical AI-focused data centre uses as much electricity as about 100,000 households, roughly 100 MW, and the largest being built today will use about 20 times more. Together, data centres used about 415 TWh of electricity in 2024, and that is expected to more than double to about 945 TWh by 2030. Most draw from the grid. Coal supplies about 30 percent of their electricity today, and renewables are expected to meet about half of the growth to 2030. For every unit of electricity the computers use, the average data centre uses about half a unit more for cooling and other overheads. ### Where the carbon comes from Running data centres caused about 180 million tonnes of CO2 in 2024, about half a percent of the world's combustion emissions, rising to about 300 million tonnes by 2035. Building them adds more. Much of the carbon is locked into steel, concrete and chips before the first token is made, and for data centres on very clean power this can be more than 40 percent of their total. ### What one site looks like: A 100 MW AI site on today's average electricity mix emits about 380,000 tonnes of CO2 a year, about what 83,000 cars emit. Over 20 years, including construction, that is about 8.1 million tonnes. ### Why this matters for the climate: Warming tracks the total CO2 the world ever emits, about 0.45 degrees C for every 1,000 billion tonnes. The world has only about 130 billion tonnes left to emit for an even chance of staying under 1.5 degrees C. The IPCC finds that each extra fraction of a degree raises the risks: coral reefs would decline by 70 to 90 percent at 1.5 degrees C, but virtually all would be lost at 2. Data centres' own 20-year footprint adds only about 0.003 to 0.005 degrees C, or about 0.01 to 0.02 if the extra fossil fuel that AI may help produce is counted. That sounds small, but it is between 4 and 36 percent of the remaining budget in 20 years, depending on how fast data centres grow. In the first three months of 2026, local opposition blocked or delayed at least 75 US data centre projects worth about 130 billion dollars, and opposition groups more than doubled to over 800 across 49 states. Their concerns are electricity, water, noise and whether they get enough in return. Many communities feel they are not getting a fair bargain. DaisyGrove tries to close that gap: a data centre that repays its carbon debt. ## Target Audience **A. Climate-impacted communities:** The carbon cost lands on everyone, but not equally. People living below melting glaciers, on coasts facing rising seas, and in places hit by floods, fires and drought are affected far more than others. Each disaster has its own causes, which scientists continue to study, and this idea does not claim to prevent any of them. It proposes that these communities should share in the benefits of the technology. **B. Host communities:** The towns and regions deciding whether to allow a data centre nearby. They carry the local costs, so they should have a say in how the site gives back. **C. Data centre builders:** The companies and public bodies that need approvals and climate claims that survive scrutiny. **D. AI coalitions:** Groups of AI companies and funders who wish to give back to communities, and who gain from the trust that comes with it. ## Customer Problems **A. A poor bargain (host communities):** Hosts carry the noise, water use and pressure on power supplies, but rarely share the benefits or choose how any offsetting is done. **B. Unprovable claims (builders):** Builders find it hard to prove they are clean. Yearly matching and credits bought far away are often doubted. **C. No access (climate-impacted communities):** The communities most affected by climate change often lack the computing power, skills and data for early warning, research and planning. **D. Growth outrunning repayment (everyone):** If many sites are built at once, each taking a 20-year carbon loan, the combined debt can outgrow the land and time available to repay it. **E. Nobody owns the whole picture (everyone):** Energy, computing, carbon and community are managed by different companies, so no one can say whether the system overall gives back more than it takes. ## Solutions DaisyGrove treats each data centre as a monolith. Everything that crosses its boundary is counted: energy in, tokens out, carbon emitted to build and run it, carbon taken back by its forest, and the share of tokens given away. Over time, the monolith should give back more than it takes. The bargain has six parts, each answering the problems above. **A. Reduce first:** Before planting a single tree, cut the debt: clean power matched hour by hour, efficient cooling, hardware kept longer, lower-carbon steel and cement, and server heat reused for homes or greenhouses. **B. Plant before you construct:** The forest is planted before or alongside the building, sized to repay the site's carbon within 20 years, and paid for upfront, including decades of care. **C. Let the community choose:** Forests grow within a set distance of the site, on land the community chooses and agrees to, using native species, without taking land that grows food, and with land rights staying with the community. **D. Give a tenth:** A share of the site's computing goes to work that does good for humanity. 10 percent is only an example; the right share is open for debate. Uses include, but are not limited to, research on cutting emissions; early warning for floods, fires and storms; monitoring glaciers, forests and coastlines; and practical help for communities affected by melting glaciers, climate-linked disasters and rising seas. Affected communities help decide how it is used, get support to use it, and see the results published. **E. Grow at a pace:** A site can only grow as fast as its forest can repay. Unpaid carbon divided by what the forest absorbs each year should stay at 20 years or less, and fall every year. New capacity is added in stages, once the previous stage meets the rule. In return, sites that keep the bargain should find communities more willing to say yes. **F. Show the books:** Every site publishes an open ledger of power used, tokens made, carbon emitted, trees planted and carbon measured, checked by an independent verifier. ## Outcome **A. What DaisyGrove delivers:** A data centre that follows this bargain repays its own construction carbon with a nearby forest within 20 years. Compared with a typical build, it avoids about 1.25 million tonnes of CO2 for every 10 MW of computers. It also gives a tenth of its computing to the communities most affected by climate change. The numbers below show how this works for one example site. **B. The carbon debt:** Take a data centre with 10 MW of computers (about 12 MW with cooling), its hardware replaced every 5 years over 20 years, and its own wind, solar and batteries. Building all of that adds up to about 185,000 tonnes of CO2, about what 40,000 cars emit in a year. **C. The forest that repays it:** A temperate forest regrowing naturally absorbs about 11.5 tonnes of CO2 per hectare each year. At that rate, repaying the debt in 20 years takes about 805 hectares: about 8 square kilometres, or about two Stanley Parks. In the tropics it would take about 540 hectares, and in a cold northern forest about 1,570. Repaying in 10 years would take about 1,600 hectares. Growth rates vary more than a hundred-fold around the world and slow as forests age, so measured uptake, not this estimate, decides when the debt is paid. **D. The tenth:** The 10 MW site uses about 105 GWh of electricity a year. A tenth of that could answer about 44 billion everyday AI prompts a year, enough for about 4 million people asking 30 questions a day. Research tasks use more energy each, so this shows scale, not a forecast. **E. The saving at one site:** A typical build with 10 MW of computers emits about 58,500 tonnes of CO2 a year, about 1.25 million tonnes over 20 years including construction. The DaisyGrove site avoids that, and its own build debt equals about three years of that saving. That is about 1.25 million tonnes less CO2 per 10 MW of computers, and about 8.1 million tonnes less per typical 100 MW site. This counts emissions avoided, not carbon removed. The build debt stays in the air until the trees absorb it, which is why the forest is still needed. **F. The saving worldwide:** If every data centre followed this path from 2026 to 2045, about 5.5 billion tonnes of CO2 would be avoided. That is an upper bound, since it assumes every data centre follows it. At 10 percent adoption, it would be about 0.55 billion tonnes. **G. What that means for temperature:** 5.5 billion tonnes is about 0.0025 degrees C of warming avoided. That sounds tiny, but it is about 4 percent of the world's remaining 1.5 degrees C budget, or about a month and a half of the world's current CO2 emissions. DaisyGrove does not tackle the extra fossil fuel that AI may help produce, so that part of the 4 to 36 percent share needs separate action. **H. What it means for people:** Host communities get a forest they chose, and a say. Builders get faster approvals and claims that survive scrutiny. Communities hit hardest by climate change get computing for early warning, research and recovery. ### About this drawing ### The flower: colt's-foot Colt's-foot (*Tussilago farfara*) is a yellow daisy of bare, disturbed ground. Unusually, it flowers in early spring, before its leaves appear, which earned it the old name "Son-before-father". Its flowering stems carry only small scales, and it spreads by creeping underground stems [13][14]. ### The metaphor Flower first, leaves later. A data centre should give back before it grows. ### How to read the plate - **From above.** On the left, the rays as straight, evenly spaced geometry; on the right, the same rays as they grow. - **The disc.** 100 florets in a spiral, 10 of them inked yellow for the shared compute. - **From the side.** The flowering stem carries only scales. The leaf, dashed, arrives after the flowers. - **Below ground.** The creeping underground stem, from which new shoots rise. ## Idea 8 · Two Seeds: Being Human Alongside AI - Date: 2026-10-02 - Address: https://yellowpaperdaisies.org/idea/2026-10-02-two-seeds-being-human-alongside-ai - Flower: Corn marigold, Glebionis segetum *A proposal to protect space for friction, misinterpretation, the wandering mind and human instinct alongside AI, so we keep the diverse ideas and fallback capability that a world optimized entirely by AI would otherwise lose.* ## **Context & Motivation** People have always solved problems in many different ways, shaped by their backgrounds, traditions and ways of thinking. New ideas emerge from friction, misinterpretation, the wandering mind, the messy "way of being" or plain instinct. When every problem, design and decision is filtered through a probabilistic algorithm, the raw human signal gets diluted. If we remove that friction from human development, we risk homogenizing how people think. This is sometimes described as a trade-off between individual and collective creativity. When an algorithm gets good at giving you what you like, it is "exploiting" what it already knows. If it only exploits, it gets stuck in a "local optimum": a filter bubble where it never discovers a new, possibly better option you've never seen. To prevent this, engineers use "epsilon-greedy" methods, which have the system "explore" by making a random, unoptimized choice a small percentage of the time. This is known as the explore-exploit trade-off, and it is common in recommendation systems (Ref B). It's like ordering from your favourite restaurant every day to avoid risking something you might like less. In doing so, you also give up the chance of finding a restaurant you might like even more. The gains in speed and quality from AI are real and useful, but a trade-off is emerging: AI can make us individually better and collectively narrower. In a controlled study of short-story writing, access to AI-generated ideas made individual stories more creative, but the AI-assisted stories were more similar to each other than stories written without AI (Ref A). There are three broad paths: treat non-AI work as obsolete, step away from AI entirely, or design for both. The third keeps the gains and keeps a backup. ## **Target Audience** A. **Primary: Schools and universities.** They are already rethinking assessment for an AI era, and their mission is to develop independent thinkers. This includes parents and families, who would ultimately choose between AI-enabled and non-AI learning paths. Education shapes how a generation learns to think, so it is where diversity of thought is easiest to protect and hardest to rebuild later. B. **Secondary: Companies whose growth depends on new ideas.** This includes research, product and creative teams, especially those hiring at scale with automated screening. They gain a direct competitive edge from ideas and talent their competitors miss. C. **Public services and critical-infrastructure operators.** This includes healthcare, utilities and finance, which must remain accessible to every resident, including people who cannot or prefer not to use digital or AI tools, and must keep working when automated systems fail. D. **Policymakers and regulators.** They are best placed to fund or encourage society-wide measures such as non-AI learning options, dual infrastructure and a guaranteed human path. Some privacy laws already give people a right not to be subject to decisions based solely on automated processing when those decisions significantly affect them, along with a right to human involvement (Ref C). E. **AI developers and labs.** Their training and alignment choices influence how strongly outputs are pulled towards the average, and they can measure and widen the diversity of what their models produce. F. **Supporting partners.** This includes product teams who design intake and review workflows, hiring-platform and screening vendors who could build an exploration lane into their tools, researchers who measure similarity in text and ideas, standards organizations, and community and workforce groups. ## **Customer Problems** A. **Regression to the mean:** Language models are trained to predict likely outputs from very large datasets, so they tend to favour common patterns over the quirky, "long-tail" ideas at the fringe of human thought. For schools, this shows up as student work that sounds increasingly alike, making original thinking harder to develop and to recognize. B. **Cognitive monoculture and single points of failure:** As more decisions run through the same few AI systems, society grows more reliant on shared infrastructure. A single outage, flaw or systematic error can then affect everyone at once. For public services and critical infrastructure, this becomes a resilience risk. C. **Trimming of the tails:** Before release, models are tuned with methods such as Reinforcement Learning from Human Feedback (RLHF) to make them helpful and safe. Research suggests this tuning can narrow the range of answers as a side effect, nudging models towards consensus responses. For companies, this means teams using the same tools may arrive at similar ideas. D. **Cultural convergence:** Research suggests that when people use AI writing suggestions, their writing can drift towards dominant cultural styles and lose regional and personal nuance (Ref D). For families and communities, this can quietly erode the cultural and personal voice that makes human communication distinctive. ## **Compelling Solutions** A. **Optionality:** The broadest systemic hedge is to keep a fully human path open at every key stage for anyone who chooses it. Like a menu that always includes vegan options, every key process offers an equal alternative that does not require AI: a non-AI schooling option alongside an AI-enabled one, or a choice between a job application process that uses AI and one that does not. Choosing it carries no penalty. B. **An exploration lane for decision-making:** A small, fixed share of applications, proposals or cases is chosen at random to skip automated screening and go straight to a person. This adds human review on top of existing fair-selection processes rather than replacing them. Each organization sets the share and tunes it as it learns. This mirrors the epsilon-greedy strategy, where a system explores a fixed fraction of the time instead of always choosing the familiar option (Ref B). C. **A variety meter:** A dashboard measures how similar a body of work is, such as essays, proposals or applications. It does not judge quality. It only signals when work is converging, so teams can bring in offline brainstorming, new reviewers or a wider candidate pool. D. **Model pluralism:** Organizations can reduce risk by not depending on a single model. Procurement can favour several models from different developers and training approaches, so one model's blind spots or outages don't become everyone's. E. **Diversity as a release benchmark:** AI developers could measure how varied a model's outputs are and report it before release, alongside safety and accuracy. F. **Local and cultural models:** Support models trained on regional languages, dialects, traditions and ways of reasoning, so AI assistance doesn't default to one cultural lens. G. **Think-first workflows:** People produce their own attempt before AI helps: a first draft, a rough solution, an initial idea. AI tutors allow productive struggle instead of resolving every difficult moment, so students still build the habit of working through problems on their own. H. **Analog fire drills:** Like fire drills, organizations periodically run key processes without AI, such as a hiring round, a planning cycle or an exam, to keep the skills and muscle memory needed to fall back if systems fail. ## **Outcome** A. **Human thinking stays plural.** People who choose to learn, create and solve problems without AI remain a thriving part of society, preserving the divergent, long-tail ideas that models tend to lose. B. **Choosing not to use AI becomes a right, not a disadvantage.** In schooling, hiring and public services, people choose how they engage, with AI or without it, and can live a full, fair life either way, supported by policy rather than goodwill alone. C. **Society can withstand failures.** When an AI system goes down, is compromised or gives systematically wrong answers, the human path keeps essential work moving, turning a single point of failure into a redundant system. D. **Convergence becomes visible and measurable.** Diversity of ideas becomes a metric that leaders, regulators and AI developers track alongside speed and quality, so groupthink is caught early rather than after it has shaped a generation's work. E. **Hidden talent comes to light.** Hires and ideas from the exploration lane are tracked, producing evidence of how often automated screens miss strong candidates, which organizations can use to improve those screens. F. **Organizations get the best of both.** They keep the speed of AI while holding on to a dependable source of genuinely different ideas, and gain an edge over competitors who all think with the same tools. G. **AI itself can improve.** As a side effect, human work created without AI, shared with the consent of the people who made it, gives developers a non-synthetic signal to train and evaluate against, which could help slow the drift towards the average in future models. ## **Appendix: References** A. Doshi, A. R. & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290 B. Sutton, R. S. & Barto, A. G. (2018). Reinforcement Learning: An Introduction, 2nd ed. MIT Press. Exploration versus exploitation and epsilon-greedy methods. C. Regulation (EU) 2016/679, General Data Protection Regulation, Article 22: Automated individual decision-making, including profiling. D. Agarwal, D., Naaman, M. & Vashistha, A. (2025). AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI 2025). ### About this drawing Corn marigold (Glebionis segetum) is a golden-yellow daisy of grain fields; its name, segetum, means "of cornfields". A single flower head makes two kinds of seed: the outer rays set flattened seeds with two narrow wings, while the centre sets plain, ribbed seeds with no wings. Its grey-green leaves are fleshy and toothed, and clasp the stem. ### The metaphor One head, two kinds of seed. Keep a human path growing alongside the fast one, from the same root. ### How to read the plate From above. Fourteen strap-shaped rays, each three-lobed at the tip, around a wide disc of tubular, five-lobed florets. The gold marks. At the base of each ray, where the winged seeds form. The two lines. A pale dotted line runs from a ray to its winged seed; a gold line runs from the centre to its plain, ribbed seed. The leaf. Grey-green and fleshy, each tooth tipped with a tiny point, its base clasping the stem. ## Idea 9 · Leaf Window - Date: 2026-10-03 - Address: https://yellowpaperdaisies.org/idea/2026-10-03-leaf-window-open-reasoning-for-ai-decisions-that-shape-lives - Flower: String of pearls, Curio rowleyanus *Proposal to share chain-of-thought reasoning with customers, establishing a standardized control framework and trust mark for all AI-enabled automated decision-making systems that make life-shaping decisions.* ## Context & Motivation When a decision shapes a human's life, such as a home, credit, insurance, a job or a social service application, one want fairness and transparency. Back in the day, a human made each decision. It was slower, but you could ask "why?" Today, automated screens often take the first pass or own the entire end to end process. Applicants get a generic accept or decline, or no reply, and a long wait to reach a customer service agent who may not have the answer. As AI costs fall, automated decisions are becoming a commodity. Explainability and transparency is now becoming the differentiator for automated AI decision-making systems. ## Target Audience Product leaders at decisioning, screening and review platforms automating lending, insurance, hiring and tenant decisions that require explanations applicants can understand. Data science teams monitoring models, work that now requires applicant-facing explanations. Architects and product managers at large organizations owning in-house decision systems that require control and transparency at lower cost. Applicants are the end customers that these explanations serve. ## Customer Problems A product manager at a growing screening platform watches automation clear thousands of applications a day. Every week, applicants write in asking the reasoning behind a decision, and the team has no good answer. The model can produce a rationale, but no one knows whether it reflects what truly drove the decision, how much to share, or what a trustworthy explanation looks like. Meanwhile, insight from thousands of decisions stays inside, and strong outliers may be negatively impacted. The problem then is explaining every automated decision openly and verifiably, without exposing other applicants or inviting gaming. ### Hypothesis A. **Opacity:** applicants receive decisions with no reasoning they can see or act on. B. **Fidelity:** a model's visible reasoning may not reflect the process that actually produced the decision, and complex, non-deterministic systems can be hard or impractical to reverse engineer. C. **No common bar:** every system explains differently, so no one can tell a good explanation from a poor one. D. **Insight:** the patterns AI sees across many applications rarely reach the person who applied. E. **Monoculture:** fully automated screens favor familiar profiles and can filter out strong outlier applicants. F. **Cost:** human review is too costly for every case, yet people expect it when it matters. ## Solutions A. **Chain-of-thought on request:** anyone affected by a significant decision, or who paid for it, can request the full reasoning trace, privacy-filtered, much like the personal data access requests that privacy laws such as the GDPR and CCPA already provide. B. **Audited chain of thought:** alongside its chain of thought, the AI records a yes/no verdict for each criterion, and the decision is calculated from those verdicts. Audits check that all three agree; any mismatch goes to a person. C. **Decision Receipt:** every decision explains which criteria mattered, the evidence and what would change the outcome, in plain language and in a format a person's own AI assistant can read. D. **Control framework and trust mark:** one open standard for every AI system making life-shaping decisions, with a Decision Facts label before applying and correction notices after. Independent checks grant the mark and can withdraw it. E. **Shared insight:** AI makes feedback cheap, so applicants learn what helped similar candidates, shown only for large, privacy-protected groups. F. **A human on request:** appeals, uncertain cases and a sample of non-standard applications reach a person, who starts with the full reasoning in hand. G. **Lower-cost automated option:** people with straightforward cases can choose a cheaper, fully automated decision, with the same reasoning access and the right to ask for a human. ## Outcome A. Every life-shaping AI decision can be explained, checked and challenged. B. Applicants turn a decision into a clear path forward, with insights shared from their own application. C. Teams automate with confidence, with fewer appeals and disputes, support for compliance and a trust advantage. D. Automated decisions keep room for outliers, so efficiency doesn't narrow who gets a chance. E. A shared transparency standard for automated decisions that can inform AI governance more broadly. ### Why a string of pearls, and why glass ### The art form: glass flowers From 1886 to 1936, father and son Leopold and Rudolf Blaschka made 4,300 glass models of 780 plant species for Harvard University, now shown at the Harvard Museum of Natural History. Alongside life-size plants, they made cut-open sections and enlarged flower parts, so students could see inside a plant without harming a living one. ### The flower: string of pearls Curio rowleyanus is a trailing succulent in the daisy family, native to South Africa. Its leaves are round beads about 6 mm across that store water. A round leaf loses little water, but it also catches little light. So each bead has a thin, clear stripe along its top, called a window, that lets sunlight into the inside of the leaf, where the plant turns light into food. In summer it bears small white flowers with colourful stamens which smell like cinnamon. ### The metaphor A string of pearls keeps itself sealed against the desert air, yet leaves one narrow window open, so light can reach the place where the real work happens. Leaf Window works the same way. An automated decision does not need to expose its whole system, its proprietary criteria or other applicants' data. It needs one clear window onto the reasoning about you: what drove the decision and what would change it. Glass models let students see inside a plant without cutting a living one; Leaf Window lets people see inside an AI decision without breaking it open. Transparency here is not the absence of a wall. It is a window placed exactly where it is needed: why the decision was made and what would change it. ### How to read the plate **Plan, from above.** The flower cluster: small star-shaped flowers, each with long stamens tipped in yellow. **The strand.** A trailing stem of beads, each slightly different, as they grow. **Outside.** One leaf drawn ten times its real size: 60 mm across on the plate, about 6 mm in life. The gold cap on top is the window. **Cut through the window.** The same leaf sliced open. Sunlight comes in through the window and lights up the inside. The rest of the skin lets far less light through. **One floret, in glass.** A single flower, enlarged and drawn as if made of glass, so its inner parts show through: the white tube with five lobes, the stamens and yellow anthers, the silky hairs, and the base where the seed forms. ## Idea 10 · HomeGrown - Date: 2026-10-04 - Address: https://yellowpaperdaisies.org/idea/2026-10-04-homegrown-ai-built-and-owned-by-the-communities-it-serves - Flower: Common yarrow, Achillea millefolium *HomeGrown is an illustrative concept for community-owned AI models, built and tested by the communities they serve, paid from day one and licensed to the services and labs that rely on them.* ## Context & Motivation AI now advises people on health, money and work, mostly through general models. Access to generic AI models alone is not enough. In a field experiment with 640 Kenyan entrepreneurs using an AI mentor, those who made changes tailored to their own business gained, while those who more often adopted generic advice did worse (Ref A). Te Hiku Media, a charitable organization belonging to Māori iwi in Aotearoa New Zealand, built speech recognition for te reo Māori and licenses its data only for uses that benefit Māori, ruling out surveillance and discrimination (Ref B). Open models can now be adapted easily, with outsized benefits for the communities it serves. ## Target Audience Communities with diverse language, culture or profession that want AI to work for them, to be paid for helping build it, and to ensure AI benefits their own people. Public services, banks and schools needing evidence their AI works for everyone they serve. AI labs wanting to reach people their models serve poorly and to become the platform community models run on. Libraries, cultural centres and associations that can host community groups. ## Customer Problems A. **Fit:** general models can be less reliable in a community's own language, dialect and context. B. **Fair pay:** people who improve AI for their community are rarely paid, while the value goes to the model's owner. C. **Ownership:** a community's language and knowledge can be used without its consent, control or a share of the benefit. D. **Getting started:** most communities lack the skills, structure and funding to begin. E. **Legitimacy:** anyone can claim to speak for a community, and dissenting voices within it can be shut out. F. **Concentration:** when a few general models answer everyone, their blind spots reach everyone and local perspectives fade. ## Solutions A. **Community groups:** libraries, cultural centres and associations host groups with a standard kit: a meeting format, roles, and training to use, test and improve AI. Members are paid and gain skills and credentials. B. **Community charter:** a co-op or trust owns the community's data, tests and model. It publishes its funding and authors. C. **Build on an open model:** members adapt it with their own data, tests and examples of good answers, by voice, in their own language and dialect. D. **Paid from day one:** members are paid first from a community fund from Frontier labs that will host the model, drawing on the share of computing that each data centre gives to work that does good in [DaisyGrove](https://yellowpaperdaisies.org/idea/2026-10-01-daisygrove); later from the model's own income, through the circular economy proposed in [Everlasting](https://yellowpaperdaisies.org/idea/2026-09-28-everlasting-daisy). E. **Test before trust:** community and general models are tested side by side. Findings that touch dangerous capabilities, such as weapons, go privately to the provider and an independent safety group, never to a public list. F. **Shared floor:** every group follows common rules anchored in human rights: no tests built to target another group, care for knowledge a community holds sacred, and appeals to an independent body. G. **Licence and earn:** the community decides who may use its model and for what, as Te Hiku Media does with its data. Services subscribe, labs pay to host the model or route questions to it, and the community is paid each time its model contributes to an answer (similar to the pay-per-stream model that exists for songs). H. **Fork:** any group can adapt a model for its own dialect or diaspora, or start its own version if it feels unrepresented, crediting and sharing income with the original. I. **Many voices, one answer:** in future, one question can be routed to several community models at once, combining their context into one richer answer, or even a shared work of art, with each community paid for its part (Ref C). This extends the model pluralism proposed in [Two Seeds](https://yellowpaperdaisies.org/idea/2026-10-02-two-seeds-being-human-alongside-ai). ## Outcome A. Communities get AI they can actually use: it works in their own language, dialect and context, and they own and govern it. B. A circular economy that stays in the community: income from licences, subscriptions and each contribution to an answer pays the people who build and test the model, and funds its next improvements. C. Members gain skills and credentials while doing good for their own community. D. AI labs reach people their models serve poorly, and becoming the trusted platform community models run on. E. More diverse answers provided by general models because answers draw on many communities and no single model shapes every answer. ## Appendix A. Otis, N. G., Clarke, R., Delecourt, S., Holtz, D. & Koning, R. The Uneven Impact of Generative AI on Entrepreneurial Performance. Harvard Business School Working Paper 24-042. B. ITU (2022). Te reo Māori speech technology and community data sovereignty. https://www.itu.int/hub/2022/08/ai-indigenous-languages-maori-te-reo/ C. Du, Y., Li, S., Torralba, A., Tenenbaum, J. B. & Mordatch, I. (2024). Improving Factuality and Reasoning in Language Models through Multiagent Debate. ICML. The study had several copies of one model debate their answers; combining different community models is a proposed extension. ### About this drawing ### The technique: transplant gardens In a study published in 1948, botanists Jens Clausen, David Keck and William Hiesey grew yarrow collected along a California transect in gardens at 30, 1,400 and 3,050 metres. Grown side by side, plants kept the forms of their home places: taller in the valley, shorter in the mountains. Such locally adapted forms of one species are called ecotypes. ### The flower: common yarrow *Achillea millefolium.* Flat-topped clusters of tiny heads, each with about five white, sometimes pink, rays around a whitish to pale yellow centre. Millefolium refers to its many finely divided leaves. ### The metaphor Nature has tested designs far longer than we have. A species does not thrive everywhere as one form; it adapts, place by place, into ecotypes. HomeGrown asks the same of AI: one open model, adapted by each community to fit its home, and free to travel and adapt again. ### How to read the plate - **From above.** Tiny heads in a flat-topped cluster: left as even geometry, right as they grow, a few in bud. - **The heads.** About five notched rays around a pale yellow centre. - **One species, many ecotypes.** Taller in the valley, shorter toward the crest; a pattern, not measurements. - **A new place.** A population settles and adapts again. - **The leaf.** Finely divided, feathery. ## Idea 11 · HalfFull Daisy: A bug bounty for doing good with AI - Date: 2026-10-05 - Address: https://yellowpaperdaisies.org/idea/2026-10-05-halffull-daisy-a-bug-bounty-for-doing-good-with-ai - Flower: Cup plant, Silphium perfoliatum *HalfFull Daisy is a proposal for a bug bounty in reverse: AI labs commit a small share of revenue to pay people for public ideas using AI to solve society's biggest problems.* ## Context & Motivation The current AI safety discourse is largely focused on what could go wrong and how to fix it. Labs pay bug bounties for flaws. Yet in the AI revolution, only a handful of people's job is simply to build AI-enabled products that do good. In-house beneficial-deployment teams (within company goals and guardrails), nonprofits, volunteers and governments do great work, but with limited resources, while material uses of AI outpace those for social good. Shouldn't one of the main goals of a general-purpose socio-technical system like AI be to bring society the most benefit? Yet half of U.S. adults still say the increased use of AI in daily life makes them more concerned than excited (Ref A). I found no programme committing a share of lab revenue to public ideas for good. There is precedent for deliberate choices like this. Before the telegraph, the postal network was the general-purpose information system of its day. The Postal Service Act of 1792 charged newspapers far less than letters so an electorate could stay informed, and extended post roads to lead settlement into the frontier, helping bind a young nation together (Ref B). In 1935, fewer than 11% of U.S. farms had electricity because utilities saw little profit in rural lines. The Rural Electrification Act of 1936 lent to member-owned cooperatives, and by 1963 about 99% of farms had power (Ref C). Those choices came from governments. With AI, the labs building it can make a similar one. ## Target Audience Frontier AI labs that want a public, bounty-style way to direct a share of revenue to society's biggest problems, earning trust beyond their products. Anyone with an idea for good, from frontline workers to researchers and students, who wants paid time and frontier resources to build tools that serve people at scale. Members of the public who want a say in what AI is built for, and governments, foundations and other funders who want a public record of vetted ideas to back. ## Customer Problems An agronomist sees small farmers lose harvests to crop disease a phone photo could catch early. Her idea: a voice-first AI advisor in local languages, built once and free to every farming community. Open innovation challenges seek answers to sponsor-set problems and lack frontier compute, models and engineers, and she has no budget or claim on the idea. The problem then is that those closest to society's biggest problems have no paid, well-resourced route from an idea for good to a tool that serves everyone. Until AI's capabilities become everyday tools that visibly improve well-being, fear of the unknown may continue to overshadow its potential. ### Hypothesis A. **Sponsor-set problems:** open innovation challenges invite anyone, but only on problems sponsors define and funding is limited. (Ref D, E). B. **No frontier resources:** compute, models and engineers are expensive. C. **Scattered ideas:** public ideas sit across separate challenges, unlinked to frontier-lab resources. D. **Business filters:** in-house beneficial-deployment teams work within company goals, so ideas must also fit the business. E. **No public voice:** the public has little say in what frontier AI is built to do. ## Solutions A. **Public rounds:** time-boxed online rounds by track, on problems proposers choose, open in many languages and via libraries, post every idea meeting published guidelines into one public record, anonymously if preferred. B. **Paid proposals:** labs shortlist and pay shortlisted proposers for detailed write-ups. C. **Two paths:** winners build with lab funding, compute and engineers, or take the bounty while the lab owns and runs it free, not-for-profit. D. **Funding:** a published share of lab revenue funds the programme, never safety budgets, much as [DaisyGrove](https://yellowpaperdaisies.org/idea/2026-10-01-daisygrove) gives a share of a data centre's compute to communities. E. **Judging:** public benefit is the only test with multiple tracks (environment, food, poverty etc.) ; a framework favouring scale asks if AI is the right tool, and the public votes on plain summaries. F. **Accountability:** each deployment has an owner, impact assessment and public incident log, counting success in terms of impact and the number of people served. ## Outcome A. Doing good with AI becomes part of the discourse, and not just a task left for someone else. B. The public, not only markets, gets a say in what frontier AI is built to do. C. Labs pursue what could go right with the rigour they bring to material metrics and finding and fixing what could go wrong, and compete on the good they do. D. Ideas for good become a public commons that governments, funders and researchers can build on. E. People see AI improve lives around them, and the conversation shifts from zero-sum fear to shared progress. F. Labs learn how their models behave in real settings, and deployments serve as evidence in policy. ## Appendix A. Pew Research Center, "Key findings about how Americans view artificial intelligence," March 2026, drawing on a June 2025 survey of 5,023 U.S. adults: [read the findings](https://www.pewresearch.org/short-reads/2026/03/12/key-findings-about-how-americans-view-artificial-intelligence/). B. Smithsonian National Postal Museum, on the Post Office Act of 1792's low newspaper rates and post roads that led expansion: [read the history](https://postalmuseum.si.edu/exhibition/fifteen-objects-that-changed-postal-history-changing-the-system/establishing-a-system-for); U.S. Postal Service, on 1792 newspaper rates: [read the history](https://about.usps.com/who/profile/history/universal-service-postal-monopoly-history.htm). C. Minnesota Historical Society (MNopedia), on farm electrification in 1935, the 1936 Act and coverage by 1963: [read the entry](https://mnopedia.org/thing/rural-electrification-administration-minnesota). D. A university-run open innovation platform that seeks tech-based solutions from anyone, anywhere: [see its challenges](https://solve.mit.edu/challenges). E. An open innovation platform whose challenges require financial sponsorship to launch and cannot take topic requests from the public: [see its FAQ](https://www.openideo.com/faq-challenges). ### Why a cup plant ### The flower: cup plant *Silphium perfoliatum.* A tall daisy of moist prairies and meadows, native to much of the central and eastern United States. It grows 4 to 10 feet on a square, hairless stem, with heads of 18 to 40 yellow rays. Its upper leaves grow in pairs that join around the stem, forming a cup that can hold rainwater for many days after it falls. This creates a vital elevated oasis for bees, butterflies, and birds during dry summers, while acting as a moat to protect the flower's nectar for true pollinators. Because the joined leaves form a complete, water-filled bowl around the stem, they create a physical roadblock. The crawling insects hit the water and cannot cross, which saves the nectar for flying insects that can simply land directly on the bloom. ### The metaphor Talk about AI often looks at the empty half of the cup. The cup plant shows the other half: a cup the plant builds once, holding rain for whoever comes by. HalfFull Daisy asks people to fill the rest with ideas for doing good. ### How to read the plate - **From above.** The flower head: the left half drawn as straight, even geometry; the right half as it grows. - **The disc.** Rings of protruding florets, each showing a dark anther tube and a curling yellow tip, around unopened buds in a green centre. - **The plant.** Leaf pairs climb the stem, each pair forming a cup. - **Half full.** One cup cut through, holding rain to the halfway line. - **The leaf pair.** Two leaves joined around the square stem. ## Idea 12 · SunAccord: Agreement on AI Harms, Risks, and Disagreements - Date: 2026-10-06 - Address: https://yellowpaperdaisies.org/idea/2026-10-06-sunaccord-one-front-door-for-ai-harms-and-hard-questions - Flower: Common sunflower, Helianthus annuus *SunAccord is an illustrative concept for a council funded by AI providers, giving people one place to report AI harms, creating shared accountability for fixing them, and providing a common forum for deciding what frontier AI may do, may not do, and where consensus does not yet exist.* ## Context & Motivation When an AI product harms someone, there are three things we should be able to expect. There should be somewhere to report what happened. There should be a clear way to see whether anyone is fixing it. And when the problem is serious, the fix should address the underlying failure and not just patch one product while the same harm remains elsewhere. Today, those pieces are fragmented. A person can report an incident to a public database, but no provider is required to act on it (REF A). Frontier AI companies have an industry forum for sharing information about serious threats, but it is limited to its members and is not a public reporting or resolution system (REF B). Laws can require providers to report certain incidents to regulators, but those reports are tied to particular jurisdictions and regulatory obligations (REF C). These mechanisms are useful, but they leave a basic gap: there is no shared system that turns an AI harm reported anywhere into a problem that every participating frontier provider is expected to investigate, resolve, and prevent from recurring. Each provider largely manages the same emerging risks on its own. There is a precedent for solving a similar coordination problem. In 2006, five major payment card brands created the Payment Card Industry Security Standards Council and shared responsibility for developing and maintaining a common security standard, while continuing to run their own compliance programmes (REF D). Competitors can agree on a shared baseline for a risk that none of them can manage effectively alone. AI needs something similar. ## Target Audience **Frontier AI providers** that face many of the same safety, security, and societal risks but currently address them independently. **Businesses** that want a recognizable standard rather than having to interpret a different set of expectations for every provider that also varies by jurisdiction. **Countries and their representatives** that want a meaningful voice in decisions about how frontier AI systems should behave. **People who are harmed by AI** and want one place to report what happened, know that it will be fixed. ## Customer Problems Imagine a parent reporting dangerous advice given by one AI assistant. Weeks later, another assistant gives essentially the same advice. The parent has to start over because the two providers do not share a common system for recognizing, investigating, or resolving the problem. The same problem appears at a societal level. Ask three frontier AI systems about a disputed border, a controversial historical event, or another question where the facts and values are contested, and they may produce three different answers. There may be good reasons for disagreement but today there is no shared institution responsible for making that disagreement visible, deciding when a common standard is needed, or recording that no consensus exists. Do we at least know where we have agreed to disagree? The underlying problem is not simply that AI systems make mistakes. It is that there is no shared mechanism for turning individual failures into industry-wide learning and shared decisions. ### Hypothesis A. **No front door:** There is no single place where anyone can report an AI harm and follow the case through investigation and resolution. B. **Isolated fixes:** A failure addressed by one provider can remain unresolved in others, allowing the same harm to recur. C. **No shared standard of care:** There is no broadly accepted, continuously updated test for what responsible behaviour should look like across frontier AI providers. D. **Slow oversight:** Traditional audits and regulation can struggle to keep pace with systems that change weekly or even daily. Note: For simplicity, this framework focuses on commercially operated frontier AI providers. We do not address open-source models here because their development, distribution, and incentive structures differ substantially and may require a different governance mechanism. ## Solutions A. **One front door:** Anyone can report an AI harm through one shared service, online or by phone. The reporter receives a case and can see whether it has been accepted, investigated, disputed, or resolved. The system is a reporting and accountability layer, not a promise that every report will result in a particular remedy. B. **One shared harm record:** The system builds on existing incident databases rather than replacing them. A validated report becomes a shared record that participating providers can investigate against their own systems. Sensitive information and dangerous findings can remain private where disclosure would create additional risk. C. **A daily governance body:** Representatives from participating countries, civil society, technical disciplines, and other relevant fields meet continuously to decide which behaviours or risks should be permitted, prohibited, or left undecided. Decisions apply to every participating provider. Votes record both agreement and disagreement, so the public can see not only the decision but where meaningful disagreement remains. D. **A formal place for disagreement:** Not every question will have a defensible consensus. When it does not, the council records the major perspectives, shows where each member stands, and makes the disagreement visible rather than forcing an artificial answer. Where appropriate, communities could also choose to route certain use cases to [HomeGrown](https://yellowpaperdaisies.org/idea/2026-10-04-homegrown-ai-built-and-owned-by-the-communities-it-serves) models built and governed by the communities they serve. E. **A floor, not a ceiling:** The shared standard sits beneath local law. Participating providers can adopt stricter requirements or undergo additional assessments, just as individual card brands maintain their own programmes on top of the PCI framework. F. **Shared funding and responsibility:** Participating providers fund the council and accept its obligations as part of membership. Businesses using participating models therefore gain access to a common assurance framework rather than having to build an entirely separate governance system for every provider. G. **Live assurance:** Assessment does not happen once a year and then disappear until the next audit. Council-funded assessors continuously review available evidence and re-evaluate important controls after significant model or product updates. Providers that repeatedly fail to address validated findings can lose good standing publicly. ## Outcome A. **People have somewhere to go.** Someone harmed by an AI system no longer has to figure out which company, regulator, database, or lawyer to contact first. B. **Reports become closed loops.** A serious report can move from submission to investigation to a documented plan for resolution, rather than disappearing into an inbox or database. Once people have made a plan for an unresolved problem, there is evidence that they may be able to stop carrying the problem cognitively themselves (REF E). C. **The world can transparently agree to disagree.** Countries, experts, civil society, and providers gain a standing forum for shaping the rules of frontier AI. Disagreement becomes something that can be recorded and understood rather than hidden behind inconsistent model behaviour. D. **Providers compete above a common floor.** Companies can still differentiate on capability, product design, and their own additional standards, while sharing responsibility for a baseline of risks that are too important to leave entirely to individual providers. E. **AI governance becomes an evolving institution rather than a yearly checklist.** The standard can change as evidence changes, incidents accumulate, and new capabilities emerge—while preserving a public record of what was decided, why it was decided, and who disagreed. ## Appendix A. The [AI Incident Database](https://incidentdatabase.ai/about/) is a public database of AI incidents and near harms. Anyone can submit incident reports, which are reviewed and indexed. It is intended to help the AI community collectively record and learn from failures. B. The [Frontier Model Forum](https://www.frontiermodelforum.org/updates/fmf-announces-first-of-its-kind-information-sharing-agreement/) is an industry-supported forum whose member firms signed a voluntary agreement in March 2025 to share information about frontier AI vulnerabilities, threats, and capabilities of concern. The information-sharing mechanism is currently restricted to member firms and is focused on serious safety and security risks. C. [Regulation (EU) 2024/1689, Article 73](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-73) requires providers of high-risk AI systems placed on the EU market to report serious incidents to the relevant market surveillance authorities. Providers must also investigate reported incidents and take corrective action. D. The [PCI Security Standards Council](https://www.pcisecuritystandards.org/about_us/) was founded in 2006 by American Express, Discover, JCB International, Mastercard, and Visa. The five founding members share ownership and governance of the Council and each incorporates the PCI Data Security Standard into its own compliance programme. The Council develops and maintains the common standard, while the individual payment brands retain their own compliance programmes and enforcement. E. Masicampo, E. J. & Baumeister, R. F. (2011). *Consider it done! Plan making can eliminate the cognitive effects of unfulfilled goals.* Journal of Personality and Social Psychology, 101(4), 667–683. The experiments found that making specific plans for unfulfilled goals reduced intrusive thoughts and other cognitive effects associated with those goals. The study tested plans made by individuals for their own goals; applying this finding to a system that gives people a documented plan for resolving an AI harm is our extension. [Read the study](https://doi.org/10.1037/a0024192) ### Why a sunflower ### The flower: common sunflower *Helianthus annuus.* What looks like one flower is actually a community of hundreds of tiny flowers called disc florets. Each can produce a seed. They develop in sequence along spirals, so the florets at the centre are younger than those around them. Yet they don't simply open according to age. Each day, the plant's circadian clock brings a new ring into bloom together. One ring opens, then the next follows the following day. Their individual stages are different, but the clock keeps them in step. Researchers found that when they disrupted this clock with constant light, that rhythm broke down. The daily burst of pollen became less predictable, and the sunflowers attracted substantially fewer pollinators outdoors. ### The metaphor AI providers are like the florets: each is a complete system, with its own rules, capabilities, and pace of development. SunAccord is the shared clock. Each day, it brings another ring of decisions into the open. The rings that have opened show where the participating providers and the wider community have reached agreement. The centre, still in bud, shows where they have not. The point isn't that every floret should become the same. It is that they can develop independently while still responding to a shared rhythm. Without that shared clock, each provider opens on its own schedule. People are left to figure out the differences themselves, and it becomes harder to know what has actually been agreed, what remains contested, and whether anyone is keeping the systems in step. ### How to read the plate - **From above.** Thirty rays surround hundreds of disc florets, making one sunflower from many individual flowers. - **The disc, today.** Cream-coloured stars mark rings that opened earlier. The yellow ring is opening today. The ochre centre remains in bud. - **In step, out of step.** With a functioning clock, one ring opens each day. Without it, the pattern loses its rhythm. - **One floret, two days.** A disc floret first produces pollen; later, its stigma becomes receptive. Even within a single floret, timing matters. ### Sources - Marshall, C. M. et al. (2023). *The circadian clock controls temporal and spatial patterns of floral development in sunflower.* eLife. [Read the study](https://elifesciences.org/articles/80984) - University of California, Davis (2023). *Circadian clock controls sunflower blooms, optimizing for pollinators.* [Read the summary](https://www.ucdavis.edu/food/news/circadian-clock-controls-sunflower-blooms-optimizing-pollinators) - Flora of North America, *Helianthus annuus*, with ray and disc floret counts. [Read the description](https://biogator.org/taxa/index.php?taxon=8624) ## Idea 13 · GoldenMend - Date: 2026-10-07 - Address: https://yellowpaperdaisies.org/idea/2026-10-07-goldenmend - Flower: Canada goldenrod, Solidago canadensis *GoldenMend is an illustrative concept for making amends when AI causes harm, so the repair is visible, fair and useful to everyone who comes after.* ## Context & Motivation When AI gets something wrong and costs someone money, time, or both, they want someone to put it right. Today, the path is usually fragmented. A consumer AI provider may ask the person to try informal resolution first, then send the dispute to individual arbitration under the provider's terms, with class actions waived and liability capped. For example, a leading frontier lab's current US consumer terms require informal resolution before arbitration, provide for individual arbitration, waive class actions, and cap liability at the greater of fees paid in the previous 12 months or $100, subject to applicable law. That can provide a route for one dispute. It does not create a shared standard for similar disputes, and its outcome does not automatically help the next person. AI agents make the problem harder. In a 2026 survey of 100 senior finance, payments, and risk leaders at UK online retailers, 58% said AI-initiated transactions had already reached their platforms. When asked about a hypothetical disputed £2,000 AI-agent purchase, respondents split over who should be responsible, and only 41% said they felt very confident in current liability frameworks. Some routes already exist, and each solves only part of the problem. A British Columbia tribunal held an airline responsible for inaccurate information given by its website chatbot, but the decision was one case in one province. The UK Financial Ombudsman Service is free to consumers and can make decisions that become binding on a firm when the consumer accepts them, but it covers financial businesses rather than AI across industries (REF A). Under the EU Digital Services Act, certified out-of-court bodies can settle disputes between users and platforms, but their decisions are not binding and the service must be free or nominal for users (REF B). We are only at the beginning of an agentic future. As AI moves from answering questions to booking, buying, cancelling, negotiating, and sending things on our behalf at scale, ordinary disputes could multiply. Many will cross borders. Sending each one to court could be expensive for the person who was harmed and expensive for the companies involved. [SunAccord](https://yellowpaperdaisies.org/idea/2026-10-06-sunaccord-one-front-door-for-ai-harms-and-hard-questions) provides the shared standard and the front door for reporting harms. GoldenMend provides the arbitration and resolution. ## Target Audience **People harmed by AI** who want a fair way to resolve a dispute without needing a lawyer. **Frontier AI providers** that want consistent, predictable rulings instead of scattered cases decided under different laws and in different forums. **Businesses building on AI** that need to know where their responsibility ends when something goes wrong between the model, the application, and the customer. **Insurers** that want real-world evidence for pricing AI-related risk. **Courts and regulators** that want everyday AI disputes resolved before they become expensive legal cases, while retaining the difficult questions for themselves. ## Customer Problems An AI agent books a nonrefundable flight on the wrong date. The travel site blames the model. The AI provider blames the site. The customer has the receipt, but the relevant conversation may be gone, the model may have changed, and neither company agrees that the loss is theirs. Now imagine the same problem appearing in five countries. At the provider, a legal lead sees similar claims arriving under different laws, through different courts, with no common record of what happened and no consistent answer about who should pay. The problem is not simply that AI can make mistakes. It is that there is no independent, affordable place designed to answer the everyday question that follows: when AI causes a measurable loss, who should put it right? ### Hypothesis A. **Judge of its own case:** The provider that may have caused the harm often controls the first complaint process and its terms. B. **Unclear responsibility:** When an AI provider and a business built on its model are both involved, each can point to the other. C. **Cost out of proportion:** Small harms, especially across borders, can cost more to pursue than they are worth. D. **Lost evidence:** Logs disappear, models change, and the evidence needed to reconstruct an AI interaction may no longer exist. E. **Inconsistent outcomes:** Similar claims can produce different answers depending on the country, court, contract, or industry involved. F. **Private outcomes:** A person may receive compensation without the underlying failure being recognized or corrected for everyone else affected by it. ## Solutions A. **Independent panels:** GoldenMend assigns cases to rotating panels combining legal, technical, and consumer expertise. The tribunal is funded through a multiyear trust that participating members pay into in advance, with each member paying for the adjudication and resolution for their share of cases through a common framework. B. **One case for everyone involved:** The AI provider and the business using its model can join the same case. The panel determines which party made which commitment, where each fell short, and how responsibility should be divided. C. **Free and easy to start:** A case can begin through a standardized export function in a participating provider's or business's application. The export packages the relevant conversation, model version, and other available evidence. Community organizations can file on someone's behalf, and free advisers can help claimants understand the process. D. **A clear test:** The panel asks whether the member followed the shared standard and its own published promises, whether its actions contributed to the harm, and whether the resulting loss can be established. A disclaimer is less persuasive when an AI acted rather than merely advised, or repeated a failure the member already knew about. E. **Case flow:** The member first receives a short opportunity to put things right. Cases are then assessed by severity, and materially similar claims can be grouped around a lead case. A minimum harm threshold keeps the tribunal focused on losses that are meaningful enough to justify adjudication. F. **People decide:** AI can sort evidence, translate documents, identify similar cases, and summarize arguments. Claimants can correct the summaries. Only people decide the outcome. G. **A second look:** A claimant who believes the process went wrong can request review by a different panel on defined grounds. An accepted decision binds the participating member, while claimants retain any rights they have under applicable law to pursue the matter in court. H. **Rules and rulings stay separate:** SunAccord writes the shared standard. GoldenMend applies that standard to individual disputes. When cases expose a gap or recurring failure in the standard, those findings go back to SunAccord for consideration. H. **Fix the individual and learn from the pattern:** Remedies are capped and designed to make the claimant whole rather than punish the member. Decisions are published without identifying the people involved. When a ruling identifies a systemic failure affecting others, the relevant fix can be applied more broadly. F. **Courts can focus on what is truly important.** Ordinary disputes can be resolved through a faster, specialized process, while novel legal questions, serious harms, and cases outside the tribunal's scope remain available to courts and regulators. Note: GoldenMend does not cover criminal matters, intellectual property disputes, catastrophic risks, or open-source models. Each raises different questions of responsibility and requires its own route. ## Outcome A. **People can get AI-related losses put right without hiring a lawyer.** A person does not have to become an expert in AI, contract law, or jurisdiction before asking for help. B. **A body of decisions shows what reasonable care looks like.** Over time, published rulings create a practical record of how the shared standard applies to real situations. C. **Providers trade scattered disputes for a predictable process.** Instead of defending similar claims independently across countries and forums, participating providers know where disputes go, how they will be assessed, and what the limits are. D. **Businesses know where responsibility sits.** Companies building on AI can make clearer decisions about which risks belong with the model provider, which belong with the application, and which must be shared. E. **Insurers get evidence of actual AI risk.** A consistent record of claims and outcomes gives insurers something better than hypothetical scenarios from which to price emerging risks. ## Appendix A. [Read how decisions work](https://www.financial-ombudsman.org.uk/who-we-are/make-decisions) B. [Read the regulation](https://eur-lex.europa.eu/eli/reg/2022/2065/oj) ### Why a goldenrod, and why kintsugi ### The art form: kintsugi Kintsugi, meaning “golden joinery,” mends broken pottery with lacquer and powdered gold, so the break is traced rather than hidden. One often-repeated story links the practice to a tea bowl that the shogun Ashikaga Yoshimasa sent for repair in the late fifteenth century. ### The flower: Canada goldenrod *Solidago canadensis.* What looks like one golden plume is actually hundreds of tiny flower heads arranged along arching branches. Each head is a complete daisy, only a few millimetres across. The genus name *Solidago* comes from a medieval Latin word associated with wound-healing plants, from *solidare*, meaning to make whole or restore. ### The metaphor Kintsugi traces a break in gold and puts the bowl back on the table. GoldenMend does the same for AI harms. It makes people whole in the open, so the repair can be seen and trust can return. Up close, goldenrod is one tiny daisy. From afar, hundreds become a single golden plume. Fair rulings should work at both scales: putting one person's loss right while revealing patterns that can help prevent the same harm from reaching others. ### How to read the plate - **From above.** The top of one goldenrod plume, its branches arching outward and crowded with tiny flower heads. - **One head, enlarged.** Overlapping bracts, short blunt rays and tube flowers with forked styles. - **The break, mended in gold.** One bowl, cracked and then repaired so the break remains visible. ### Sources - Flora of North America, *Solidago canadensis*, for the structure of the flower heads and florets. [Read the description](https://biogator.org/taxa/index.php?taxon=8234) - Merriam-Webster, “solidago,” for the genus name's origin. [Read the entry](https://www.merriam-webster.com/dictionary/solidago) - New Orleans Museum of Art, on kintsugi and its often-told origin story. [Read the article](https://noma.org/kintsugi-reflects-beauty-in-flaws-strength-in-healing/) ## Idea 14 · DeepRoot: Making AI Insurable - Date: 2026-10-08 - Address: https://yellowpaperdaisies.org/idea/2026-10-08-deeproot-making-ai-insurable - Flower: Arrowleaf balsamroot, Balsamorhiza sagittata *DeepRoot is a proposal to make everyday AI insurable, so safety gets a market incentive, risk can be priced in and companies can build with confidence.* ## Context & Motivation New technologies often follow the same pattern. At first, insurers do not know enough about the risks to price them, so they raise premiums, restrict cover, or exclude the risk altogether. Over time, shared data, standards, and independent testing make those risks easier to understand. Insurance then becomes more than a way to pay for losses. It starts influencing how products are designed and how safety is maintained. History offers several examples. After the Great Fire of London, fire insurers paid their own teams to fight fires. Preventing losses was in their financial interest. Shipping insurers gathered in a coffee house where merchants exchanged news, eventually giving rise to a ship register that assessed the condition of vessels. In 1894, fire insurers helped establish a laboratory to test electrical equipment. Car insurers later funded crash tests that changed how vehicles were designed (REF A). In each case, insurers helped develop the tools needed to understand risk, and that knowledge influenced the products they covered. Insurance depends on being able to measure risk. Insurers usually rely on years of claims data and risks that change slowly enough to model. AI is harder to assess. The same request can produce different answers, models change between versions, and risk of AI varies greatly by use cases even for the same model. New types of harm are also still emerging. There is little historical data to work with, and the risks can change faster than traditional underwriting can keep up. Cyber insurance faced a similar problem. A wave of ransomware claims pushed the US standalone cyber loss ratio from 47.1% in 2019 to 72.8% in 2020, while US cyber insurance prices rose 130% in late 2021 (REF B, C). Insurers responded by demanding stronger security controls. Market repriced the risk and the raised premiums incentivized stronger risk mitigation. Cyber insurance now has a growing set of standards and procedures for assessing and pricing risk. AI insurance has yet to develop the same foundation. For now, much of the insurance market is moving in the opposite direction. Optional exclusions for generative AI entered general liability policies in January 2026, and several large US insurers have sought approval for broader AI exclusions (REF D). Some AI labs have reportedly relied on investor funding to help cover settlements because insurance is difficult to obtain (REF E). AI investment has moved faster than the tools needed to underwrite its risks. Without reliable ways to measure those risks, insurers have little incentive or ability to distinguish safer systems from riskier ones. Some AI risks, particularly those involving national security, may never be commercially insurable. Cyber insurance already offers a precedent for excluding certain state-backed attacks. But that does not mean every other AI-related harm should not be granularly measured and insured. An economy that increasingly depends on AI needs a way to assess and cover the everyday losses that can be quantified. Public or strategic support may change with political priorities and availability of funding and when that day comes, a functioning insurance market needs a more durable basis for assessing risk. [SunAccord](https://yellowpaperdaisies.org/idea/2026-10-06-sunaccord-one-front-door-for-ai-harms-and-hard-questions) sets the shared standard. GoldenMend turns individual harms into rulings. DeepRoot uses those rulings to help price risk, and what insurers learn feeds back into the standard. Together, they create a cycle in which harms inform decisions, decisions inform prices, and prices give organizations a financial reason to build safer AI. ## Target Audience **Businesses building products with AI** that need insurance for the risks those products create, so they can offer them with greater confidence. **Frontier AI labs** that want to demonstrate how their systems manage risk and make them suitable for higher-stakes uses. Models can then be compared measurably not just on technical metrics but on financial metrics too across business use cases. **Insurers** that need reliable evidence about AI-related losses and safety controls, so they can assess individual risks rather than exclude AI broadly. ## Customer Problems A travel company launches a booking agent built on a frontier model. Its customers want reassurance, and its board wants insurance. But the company's insurer has added an AI exclusion, while the model provider limits its own liability. Neither party can show what a particular control failure would cost, and there is little claims history to help the insurer estimate the risk. Insurance can act as an independent check on safety, but only when three conditions are met: someone can be held responsible for a loss, there is a reason to require insurance, and there is enough evidence to price the risk or a credible system of standards and inspections that can help fill the gaps. For much of AI, these conditions are not yet in place. **The underlying problem is that AI risks are not measured consistently enough to be priced and insured with confidence. Insurers too do not have the mechanisms to reliably measure these novel risks by themselves with data that is available to them.** ### Hypothesis A. **Limited liability:** Provider terms often limit liability, while the EU withdrew its proposed AI Liability Directive. As a result, people harmed by AI may face uncertainty about whether they have a viable claim and how to pursue it (REF F, REF G). B. **No requirement to carry cover:** Frontier labs may supply models used by other businesses without being the company directly serving the end customer. There is not yet a general requirement for these labs to carry insurance for the harms their models may contribute to. C. **Businesses struggle to find cover:** Companies selling AI-powered products need protection against potential losses, but insurers may offer limited cover or exclude AI-related risks. D. **Too little evidence to price risk:** There is no mature, shared record of AI losses or consistent way to measure risk. Building both the evidence and the measurement methods will take time. E. **Controls are difficult to compare:** Safety controls vary in what they prevent and how effectively they work. Without consistent evaluation, insurers cannot reliably compare one system with another. F. **Residual risk is unclear:** Even after safeguards are introduced, it is difficult to quantify the risk that remains. G. **Models keep changing:** A model may respond differently to the same request, and its behaviour can change between versions. Annual insurance pricing may not reflect those changes. H. **Harms remain hidden:** Companies may not disclose incidents, and many AI-related risks are too new to have established reporting practices. I. **Losses can be correlated:** A single model failure could affect thousands of customers or businesses at once, creating a concentration of claims. J. **Sharing evidence creates exposure:** Labs may be reluctant to share safety and incident data if it could reveal vulnerabilities, expose confidential information, or be used against them in a dispute. K. **Not every harm appears in a claim:** Some harms are difficult to express in financial terms, and many affected people may never file a claim. ## Solutions A. **Build measurement together:** AI labs and insurers develop monitoring, data collection, and risk-evaluation methods together, with independent evaluators checking the results. Labs demonstrate how their systems perform rather than relying on assurances alone. B. **Put a price on residual risk:** Evaluate controls according to the harms they prevent, then estimate the risk that remains after safeguards and compensating measures are in place. The result is a range that reflects uncertainty, rather than a single number presented as certainty. Measurements will become more accurate as harms become known over time, which will reduce uncertainty. C. **Monitor continuously and price fairly:** Labs share agreed telemetry and evaluation results through a secure channel, without sharing model weights. As with usage-based car insurance, premiums can reflect observed behaviour and risk. Adjustments happen at renewal or within tiers whose terms are agreed in advance (REF H). D. **Model realistic scenarios:** Use scenario models to test how failures could spread across systems and affect many people at once. Calibrate those models annually against SunAccord's record of reported harms and GoldenMend's rulings, including evidence of losses that are not purely financial. E. **Connect claims to rulings:** GoldenMend gives affected people a route to a decision about responsibility. Insurers use those rulings to assess covered claims, subject to the policy terms. Retained logs, model versions, and reproducible tests help establish what happened. Straightforward, low-value claims can follow a faster payment process where coverage and eligibility are clear, without requiring a full fault determination in every case. F. **Make evidence safer to share:** Establish clear rules for sharing incident data and evaluation results. The aim is to improve risk assessment and prevent repeat failures, not to turn every disclosure into an opportunity to assign blame. G. **Reward safer behaviour, not blanket refusal:** Systems that demonstrate effective safeguards and lower measured risk should be eligible for lower premiums. A model should not automatically appear safer simply because it refuses more requests; evaluation should account for whether it remains useful while managing risk. H. **Make appropriate cover part of participation:** SunAccord members carry insurance proportionate to their scale and exposure, using layers of cover where appropriate. Risks that cannot be insured at an acceptable level must be reduced, contained, or kept out of uses where the potential losses would be unacceptable. I. **Extend cover to the businesses building on AI:** Once a model's risks can be measured more reliably, insurers can develop cover for the businesses that incorporate it into products and services. This gives those businesses a clearer way to manage their own exposure. J. **Feed insurance evidence back into the standard:** Lessons from claims, rulings, monitoring, and underwriting return to SunAccord. The shared standard can then evolve as evidence reveals which controls work, where failures persist, and which risks remain difficult to measure. ## Outcome A. **Insurance becomes an additional market incentive for AI safety.** It complements laws and regulators by giving organizations a financial reason to measure and reduce risk. B. **AI risk is more granularly priced in.** Claims, evaluations, and independent assessments help distinguish demonstrated safeguards from unverified assurances. C. **Insurers have an incentive to prevent losses.** As in fire safety, electrical testing, and vehicle safety, better measurement can make prevention financially worthwhile, not just paying claims after something goes wrong. D. **More high-stakes uses become insurable.** When risks can be assessed and managed, businesses have a clearer path to deploying AI in settings where reliability matters. E. **Each harm improves the shared understanding of risk.** Rulings and incident data can reveal recurring failures and help improve evaluations across participating systems. F. **Safer systems can pay less.** When lower risk is demonstrated through credible evidence, insurance pricing can reward it. ## Appendix A. [Read about early fire insurers](https://www.london-fire.gov.uk/museum/london-fire-brigade-history-and-stories/early-insurance-brigades), [the ship register](https://www.lr.org/en/who-we-are/brief-history/), [electrical testing](https://ul.org/about/our-history/) and [crash tests](https://iihs.org/news/detail/iihs-crashworthiness-tests-save-nearly-50-000-lives-since-programs-launch) B. [Read the loss ratios](https://www.reinsurancene.ws/cyber-industry-loss-ratio-at-record-high-67-in-2020-aon/) and [the price index](https://www.marsh.com/sg/about/media/global-insurance-market-index-q4-2021.html) C. [Read about the new exclusions](https://www.independentagent.com/vu_resource/verisk-to-roll-out-new-general-liability-exclusions-for-generative-ai-exposures/) and [insurers seeking approval](https://techcrunch.com/2025/11/23/ai-is-too-risky-to-insure-say-people-whose-job-is-insuring-risk/) D. [Read the report](https://amp.insurancejournal.com/news/national/2025/10/08/843087.htm) E. [Read about the war exclusion case](https://www.reinsurancene.ws/merck-reaches-settlement-with-insurers-over-1-4bn-notpetya-cyber-attack/) and [the market bulletin](https://assets.lloyds.com/media/e9a4058c-4c25-48d6-99d3-0c466987e099/Y5381%20Market%20Bulletin%20-%20Cyber-attack%20exclusions.pdf) F. [Read an example of consumer terms](https://openai.com/policies/terms-of-use/) G. [Read about the withdrawn directive](https://eapil.org/2025/10/09/european-commission-withdraws-two-proposals-assignments-of-claims-regulation-and-ai-liability-directive/) H. [Read about usage based insurance](https://content.naic.org/insurance-topics/telematics) ### Why an arrowleaf balsamroot ### The flower: arrowleaf balsamroot *Balsamorhiza sagittata* grows in dry grasslands from southern British Columbia to California. Its name says what to look for. The leaves are shaped like arrowheads, and the root holds a balsam scented resin. Each head carries 8 to 25 bright yellow rays. Most of the plant is out of sight. Its taproot can reach 10 cm across and 2.7 m deep. Fire kills the leaves and stalks, but the plant regrows from its crown and can grow denser after a burn. ### The metaphor What you see is the bloom. What carries it through fire is the root it grew in good years. DeepRoot asks AI to grow that root before the fire. Measure the risk, build the reserves and let insurers see below ground. The deeper the root, the harder the ground it can bloom in. ### How to read the plate - **From above.** One head with 20 long rays around a domed disc, buds at the centre and open tube flowers at the rim. - **Cut through the head.** Woolly bracts, rays and tube flowers on one dome, with chaff between the flowers. - **Above and below ground.** After fire, burnt stalks in ochre, new arrowhead leaves rising from the crown and a deep taproot with side roots. ### Sources - [Read the species profile](https://www.fs.usda.gov/rm/pubs_journals/forbs/arrowleaf_balsamroot.pdf) - [Read the fire ecology review](https://www.fs.usda.gov/database/feis/plants/forb/balsag/all.html) ## Idea 15 · Schrödinger’s Daisy - Date: 2026-10-09 - Address: https://yellowpaperdaisies.org/idea/2026-10-09-schrodinger-s-daisy - Flower: Upright prairie coneflower, Ratibida columnifera *Schrödinger's Daisy is a sociotechnical approach to counter misinformation in the post-truth era, inspired by investigative design, that makes uncertainty transparent through a probabilistic view of facts.* ## Context & Motivation Misinformation ranks second among 33 risks over the next two years, according to global experts (Ref A), and the problem is likely to get worse. One large study found that false news reached 1,500 people about six times faster than true news (Ref B). Humans simply cannot keep up with checking everything machines can produce. ## Target Audience People who want to look beyond viral headlines, review the entire chain of sub-facts and evidence, and contribute towards reducing ambiguity. Everyday readers who want to know which parts of a story are verified, what remains uncertain, and what they need to decide for themselves. ## Customer Problems Humans dislike ambiguity and want a clean-cut answer, a readily available perfectly reliable truth. But it is becoming harder to cleanly separate truth from falsehood, and our understanding of reality should reflect that. We’re entering the era of **Schrödinger’s Truth**. A claim may not be 100% verifiable, but do we know where its evidence came from? How was that evidence collected? What remains unverified, and how likely is each step in the process to be correct? Consider a video of a hospital fire going viral during a conflict, followed by thousands of AI-generated versions. One clip is exposed as fake, and people begin questioning whether the entire incident happened. The chain of events that needs to be investigated gets lost because one piece of evidence doesn’t add up. **Hypotheses** A. **Black-and-white answers:** Everyone wants a black-and-white answer, but deepfakes make it difficult to cleanly separate truth from falsehood. B. **The verdict, not the process:** People are shown the verdict, but not the process behind it. C. **Machine scale:** There aren’t enough humans in the world to verify all machine-made misinformation. D. **One piece doesn’t add up:** People start questioning an entire incident because one piece doesn’t add up. E. **Gaming the system:** A large population of people can game a vote by upvoting a falsehood. ## Solutions Investigative design offers a way forward. The Museum of Modern Art’s exhibition *Full Disclosure: The Edge of Information Design* explores how information design can help people understand complex information. Its investigative design materials describe an approach that lets people “assess the trustworthiness of the process itself” (Ref C). Unlike community notes and online encyclopedias, which aim to establish a conclusion, this approach reviews the process behind a claim rather than demanding a single verdict. The focus is on the integrity of the process and the balance of probabilities, not the absolute certainty of a final verdict. We don’t always need to be 100% certain. Sometimes 80% certainty is enough, provided we understand what is known, what isn’t, and why. That allows each person to make their own judgement. A. **Show the process, not the verdict:** AI fact-checks claims and publishes a log of the sources used, tracing each source back to its origin and making missing evidence visible. B. **Vote on the process, not the outcome:** Humans review each source link and vote on how much they trust it using an up or down click or a trust score from 0 to 100 percent. Users can review the entire chain of sub-facts and evidence to score each part. They evaluate how well the evidence supports each step rather than voting on whether the overall story is true or false. C. **Make missing steps visible:** People can call out missing process steps. AI provides its reasoning for why a step should or shouldn’t be included, and people can vote to remove a step if it doesn’t seem logically required. D. **Schrödinger’s Truth:** Each link has a probability, contributing to an overall probability for the claim. Nothing is declared 100% true or 100% false. Whatever remains unconfirmed is labelled as such, and if no evidence exists, the shared ledger says so. E. **Credibility must be earned:** Once later evidence verifies a claim above a set confidence level, those who voted incorrectly receive a negative score. Over time, this builds a person's credibility. High credibility carries more weight in the overall probability score, with a minimum number of votes required. Past credibility is the only scoring factor. F. **Money alone can’t game it:** Someone with unlimited money should not be able to manipulate the system through money alone. They would need to compromise a majority of high-credibility users. G. **100% transparency:** Everyone can see who voted for what. Votes are recorded on a cryptographic ledger that allows independent verification, including anonymous verification where appropriate. Anyone can inspect the record and spot red flags, such as someone with high credibility voting along certain patterns/ ideologies, even when that means supporting falsehoods. H. **Paid credibility:** People with high credibility earn a share of the website’s subscription revenue over time, creating a financial incentive for careful, evidence-based judgement. I. **AI handles the volume, not the verdicts:** AI groups near-identical content (which humans can override), helping reviewers focus their attention. People remain responsible for judging the evidence. **Not Yet Solved** A. **Buying the voter base:** A resourceful malicious actor could buy 50% of the voter base, even if minimum required voter base is extremely large. B. **Patient capture:** A group of voters could earn credibility for years, then spend it on one lie. **Outcome** A. **Make the process visible:** The parts of a story that are verified, unverified or strongly supported become known to all, and everyone can make their own conclusions. B. **Reflect reality as it is:** Truth is no longer forced into black-and-white answers when the evidence supports degrees of confidence instead. C. **Reward careful judgement:** Credibility becomes something people earn by being right and get paid for over time. **Appendix** A. World Economic Forum, *The Global Risks Report 2026*. [Read the report](https://www.weforum.org/publications/global-risks-report-2026/) B. Vosoughi, Roy and Aral, “The spread of true and false news online,” *Science*, 2018. [Read the paper](https://doi.org/10.1126/science.aap9559) C. The Museum of Modern Art, *Full Disclosure: The Edge of Information Design*, September 27, 2026 – June 13, 2027. [Read about the exhibition](https://press.moma.org/exhibition/fulldisclosure/) ### Why upright prairie coneflower ### The flower: upright prairie coneflower *Ratibida columnifera* grows from southern Canada to Mexico. Its name means "column bearing", after the tall column of 200 to 400 or more tiny florets at its centre. Like most daisies, it opens its florets from the rim inward. On a column the rim is the bottom, so it blooms in bands from the base up. Unopened florets look greenish grey, because white hairy scales hide them. ### The metaphor Ask whether a prairie coneflower is in bloom, and the honest answer is "partly". Some bands are open and some are still closed. A Schrödinger's truth works the same way. Some steps are verified, some are not, and everyone can see which. The name borrows Erwin Schrödinger's 1935 thought experiment in its everyday sense, a claim neither simply true nor false until the evidence is checked. ### How to read the plate - **From above.** The flower, its 7 rays notched at the tip, the column seen end on. - **Yellow stars.** 139 open florets in the lower bands, seen from above as the outer ring. - **Cream circles.** 121 closed florets at the top of the column. - **Side view.** Open from the base up, the rays drooping below. - **As it opens.** The same column early, midway and fully open. ### Sources - [Read the flora description](https://biokic4.rc.asu.edu/sandbox/dev/taxa/index.php?taxon=2228) - [Read about the unopened florets](https://biorepo.neonscience.org/portal/taxa/index.php?tid=30604) - [Read about the bloom from the bottom up](https://www.minnesotawildflowers.info/flower/prairie-coneflower) - [Read about the name](https://missouriplants.com/Ratibida_columnifera_page.html) - [Read how daisy florets open](https://mgnv.org/plants/glossary/capitulum-disk-disk-floret-floret-ligule-ray-floret/) ## Idea 16 · Provenance Daisy: Attribution in the age of AI - Date: 2026-10-10 - Address: https://yellowpaperdaisies.org/idea/2026-10-10-provenance-daisy-attribution-in-the-age-of-ai - Flower: Giant groundsel, Dendrosenecio keniodendron *Provenance Daisy is an attribution model for AI so that human intelligence and ingenuity keeps being incentivized rather than quietly absorbed.* ## Context & Motivation When a mathematician shares an idea with a colleague, the colleague knows where it came from and can give credit. AI conversations are different. Most AI systems were not built with ways to attribute their training data and intellectual property to individual contributors (Ref A). Research into attribution exists, but no widely adopted model has solved attribution. Ref A, B). We can’t undo the past where the sources of training data for AI aren’t well known, but we can change what happens next. Without a solution to attribution, people may have less incentive to share their knowledge and even take up research, a problem that could persist well into the future (Ref C) with grave societal implications. Trust takes years to build. AI developers need to earn it with verifiable evidence that contributors receive fair credit. (Ref D). ## Target Audience Researchers who have shared original ideas with AI and want a mechanism for attribution when those ideas contribute value. AI labs looking to build trust through attribution. AI governance leads working to prevent value erosion when contributions go unrecognized. Anthropologists interested in how incentives shape the qualities society values, including intelligence and ingenuity. Researchers who need better knowledge management and intellectual property protections when working with AI. ## Customer Problems Imagine a mathematics researcher who spends a lifetime working on a millennium problem and shares a key insight with an AI system. The insight’s value may only become clear years later when ideas collected across multiple researchers helps AI solve one of the millennium problems when the connection to the person who shared it has been lost (Ref E). They could lose recognition for their life’s work, along with the financial rewards it might bring (Ref F). The stakes can be substantial: each of the seven Millennium Prize Problems carries a $1 million prize (Ref G). AI systems cannot take authorship or receive credit in the way people can, leaving recognition and rewards to flow to other people or companies (Ref H, I). The people whose knowledge advances a field could lose both recognition and a livelihood. Mathematicians are just one example. The problem extends to anyone whose knowledge contributes to AI, raising broader questions about how we manage and reward knowledge in an AI-driven world. People may feel forced to choose between avoiding AI and falling behind, or avoiding fields where AI could make their contributions harder to protect (Ref J, K). Either way, we risk weakening the incentives that drive human intelligence and ingenuity over time (Ref J). This also undermines trust. When people cannot see how their contributions are tracked or rewarded, it is difficult to know whether AI developers have a genuine incentive to recognize them fairly (Ref C, L). The underlying problem is **value erosion without attribution**: the people whose knowledge makes AI useful can go unrecognized and uncompensated. **Hypotheses** A. **The link gets lost.** The value of an insight may only become clear after its connection to the person who shared it has disappeared. B. **Credit goes elsewhere.** AI cannot hold credit in its own right, so recognition and rewards can end up with other people or companies. C. **Contributors carry the burden of proof.** Researchers whose ideas are used without recognition carry the burden to prove that their contributions mattered. D. **Contributions go unpaid.** Researchers receive no compensation for the value their knowledge adds to AI. E. **Ideas lose their attribution when shared.** Someone who receives an idea through AI may reuse it without carrying its original attribution forward. F. **Promises are not proof.** Assurances from the organizations that benefit cannot, on their own, establish that contributors were credited or that their ideas were never used. ## Solutions A. **Personal agents.** Everyone can have a personal AI agent that learns from their experience and conversations. What they teach it stays within their personal agent rather than becoming part of the shared model (Ref M). B. **Find the right agent.** When AI needs help solving a problem, it can ask which personal agents have relevant knowledge (Ref N). If someone’s agent provides a valid, useful answer, the contribution is attributed to that agent and the person who taught it. C. **Keep agents isolated.** Personal agents remain separate from one another. The company operating the shared model is also prevented from absorbing what these agents learn, protecting the knowledge contributed by each person. D. **Record every agent to agent exchange.** Every idea shared between agents is logged. Attribution is recorded at the point where knowledge passes from one agent to another, where the exchange can be tracked, rather than being left to inference inside a model (Ref B, E). E. **Carry credit forward.** Every answer comes with a record of its origin and terms governing its use. Recipients cannot use it to train their own models, and any work built on the answer carries the attribution forward, much like citations in research. F. **Establish what the person contributed.** A personal agent still relies on a shared model, so not everything it produces necessarily comes from its owner. To separate the two, the same question is tested repeatedly against the shared model without the personal layer (Ref M). The difference in what the personal agent produces helps identify knowledge contributed through the person’s experience. This would provide evidence of contribution. G. **Give credit without pretending the agent is the person.** A personal agent may generate words or ideas its owner would not have produced themselves. The owner can still receive credit for knowledge contributed through the agent, without being treated as the author of every output. The agent must identify itself as an agent, the owner can keep unpublished work sealed until they choose to release it, and they can turn the agent off at any time. H. **Put the burden on the builder.** Contributions are recorded automatically. If a dispute arises, the AI developer must show that a contribution was not used, rather than requiring the contributor to reconstruct what happened after the fact. I. **Pay for each contribution.** Contributors are paid whenever their knowledge is used, according to terms they set, rather than receiving an allocation from a shared pool. When several people contribute to the conception of a result, they share credit for their respective contributions, drawing on the principle of joint inventorship in patent law (Ref I). An independent reviewer resolves disputes. J. **Back trust with independent evidence.** Trust should not depend on assurances from the organizations that benefit. Independent audits, published training-data records, binding commitments and meaningful consequences for breaches can provide evidence that the system is operating fairly (Ref D). Technology cannot prove everything. For example, there is no formal way to prove that a particular person’s data was never used in training. In those cases, independent oversight and governance can provide accountability where technical proof falls short (Ref O). ## Outcome A. People can share their best ideas with AI with greater confidence that their contributions will be recognized and fairly compensated. B. Society continues to reward human intelligence and ingenuity, giving people stronger reasons to pursue the research that advances humanity. C. Knowledge management evolves to include attribution as a core part of how knowledge is shared, used and rewarded, rather than leaving contributions buried inside models. D. Trust in AI is built on evidence and accountability, not promises alone. ## Appendix A. Large language models learn from vast amounts of human writing, yet the people who provide that knowledge often go uncredited. Measuring the influence of each contribution has also been limited by its cost. [Read the paper](https://arxiv.org/abs/2405.13954) B. A broad evaluation found that influence functions, a leading approach to attribution, perform relatively poorly on large language models. [Read the paper](https://arxiv.org/abs/2409.19998) C. Treating people’s data as free capital overlooks their role in creating it, weakens incentives to contribute and benefits the dominant buyers of data. [Read the paper](https://doi.org/10.1257/pandp.20181003) D. To earn trust, AI developers need to make verifiable claims, accept accountability and open their systems to outside scrutiny. [Read the report](https://arxiv.org/abs/2004.07213) E. In large language models, the influence of any single training example on a given answer can be diffuse and difficult to trace. [Read the paper](https://arxiv.org/abs/2308.03296) F. An AI system solved a long-standing open problem posed by Paul Erdős that carried a $500 prize. [Read the paper](https://arxiv.org/pdf/2605.20695) G. Each of the seven Millennium Prize Problems carries a $1 million prize. [Read about the prize problems](https://www.claymath.org/millennium-problems/) H. AI tools cannot be listed as authors because they cannot take responsibility for the work. [Read the position](https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools) I. Revised US patent guidance maintains that only humans can be inventors, with inventorship, including joint inventorship, based on contributions to conception. [Read a summary](https://academic.oup.com/jiplp/article/21/5/272/8528856) J. Scientists who use AI publish more, receive more citations and lead projects earlier, while scientific research as a whole becomes more concentrated in its focus. [Read the paper](https://arxiv.org/abs/2412.07727) K. In a 2026 survey, 16% of US college students said they had changed their major because of AI. [Read the findings](https://news.gallup.com/poll/704087/college-students-weigh-impact-majors-careers.aspx) L. Three in five researchers surveyed said AI in research could undermine intellectual property and leave authors unrecognized. [Read the findings](https://corp.oup.com/news/how-are-researchers-responding-to-ai/) M. One proposed approach gives each user a personal module trained on their own history and connected to a frozen shared model, allowing them to own their personalized model. [Read the paper](https://aclanthology.org/2024.emnlp-main.372) N. An open standard allows independent AI agents to discover one another and delegate tasks. [Read the paper](https://research.google/pubs/agent-to-agent-a2a-protocol-the-future-of-autonomous-multi-agent-systems/) O. Researchers have shown that it is not formally possible to prove that particular data was absent from training. What can be audited is whether a process open to outside scrutiny was followed. [Read the paper](https://www.usenix.org/conference/usenixsecurity22/presentation/thudi) ### Why a giant groundsel ### The flower: giant groundsel *Dendrosenecio keniodendron.* A daisy that grows into a tree up to 7 metres tall, high on Mount Kenya and the Aberdares. Its yellow heads hang down, with no rays. Flowering is the only time it branches. A rosette flowers once and dies, and new rosettes grow from the base of its dead flower stalk. So every fork records one flowering. Scientists can reconstruct a plant's reproductive history simply by looking at the pattern of forks in its trunk. The whole population on Mount Kenya flowers together every several years, so one plant's record can be checked against its neighbours. ### The metaphor The record is grown, not written afterwards. Each time the plant gives seed, the gift becomes part of its body, where anyone can read it. Provenance Daisy asks AI to keep the same kind of record, with every contribution logged as it happens and credited to the person it came from. ### How to read the plate - **The flowering crown.** Broad leaves, a cloak of dead leaves and one stalk of hanging yellow heads. - **One flowering, one fork.** A rosette grows, flowers and dies, and new ones branch from its base. - **Reading the record.** Each yellow ring is one flowering, matched across two neighbours. ### Sources - Smith and Young (1982), on branching, flowering and the population flowering together. [Read the study](https://tpyoung.ucdavis.edu/sites/g/files/dgvnsk4821/files/inline-files/cost%20of%20reproduction%20in%20rosette%20species.pdf) - Flora of Tropical East Africa, on size, range and flower heads. [Read the description](https://plants.jstor.org/compilation/Senecio.keniodendron) ## Idea 17 · EyeOpener Daisy: The Case for Personal Agent Governance - Date: 2026-10-11 - Address: https://yellowpaperdaisies.org/idea/2026-10-11-eyeopener-eyeopener-daisy-a-personal-agent-governance-centre - Flower: Black-eyed Susan, Rudbeckia hirta *EyeOpener Daisy is a personal agent governance centre that lets people see, check and correct what their agents say, share and agree on their behalf.* ## Context & Motivation We're entering an age where agents increasingly talk to other agents on people's behalf. In a 2025 survey, 51 percent of customers said they would be willing to use a GenAI assistant for customer service interactions on their behalf [A]. Analysts expect 60 percent of brands to use agentic AI for one to one marketing interactions by 2028 [B]. Businesses can already see this traffic, from gateways that log exchanges between agents. Stores too can track what visiting agents do [C]. The person an agent speaks for has no such view and often sees only the outcome. Why use an agent at all? Staying away can feel safer, much as staying offline once did. But as banking, healthcare and benefits moved online, often faster or better than before, a House of Lords committee found that people who can't get online are falling ever further behind [D]. Agents look like the next step, and as the technology races ahead, people who use them may build an edge over those who don't. A large technology company's chief executive has warned that persistent agents can become a new kind of insider risk [E]. Companies manage that risk by watching closely and stepping in early. If agents become how people keep up, people will need the same for the agents that speak in their name. ## Target Audience People are okay with delegating selective tasks to agents and want one place to see, check and correct what they say, share and agree, staying in control. AI assistant makers that want to earn trust for bigger tasks and support consumer protection compliance. Businesses that want fewer disputed agent deals, and public booking that stays fair. ## Customer Problems A warehouse worker in Manchester needs a driving licence for a delivery job. Most test slots vanish almost as soon as they are released, often to paid services that check around the clock and sometimes book for learners, a speed no person can match [F]. So he is forced to use an AI agent. The agent books a test, then arranges car insurance with an insurer's agent. It takes the first quote and, when asked about his situation, overshares his pay and the hours he is away on night shifts. He sees the booking and the policy in two apps, but none of the conversations. Britain now bans booking a test for anyone but yourself [G], and it is unclear whether that covers his own agent. The problem then is that people increasingly need agents to keep up, yet have no single place to see, check and correct what those agents say, share and agree on their behalf. ### Hypothesis A. **Users need agents to keep up:** Systems that reward the fastest click force users toward using agents, just to keep up [F]. B. **Users can't see agent conversations:** Users usually see only outcomes, with each app keeping its own record of what was said, offered or shared. C. **Agents say what their users wouldn't:** Nothing checks an agent's words against how its user would respond, so it can misstate their views, values or lived experience. D. **Agents share what their users wouldn't:** Privacy depends on information suiting the context it flows into [H]. Agents acting for users can leak sensitive information despite privacy instructions [I], and other agents can extract more through impersonation and forged consent [J]. E. **Agents agree to what their users wouldn't:** Agents tend to take the first offer and can be manipulated [K], and more capable agents consistently win better deals [L]. F. **Users can't step in or undo:** A uniform law adopted in 49 US states lets two agents form a contract nobody reviewed [M]. Its error rule was written for people making mistakes with a company's system, and whether it covers a user's own agent is unsettled [N]. There is often no simple way to pause an agent, correct it or tell the other side. ## Solutions A. **Fair first come, first served:** Scarce slots get stronger bot protection, and some uses ban agents outright. Some resources are reserved for manual / non-agent transactions so that people are not outrun by software. B. **A personal agent governance centre:** One centre reviews everything a user's agents say to other agents and people, using fixed rules and AI judgment. It flags only the few exchanges that need attention, like insider risk monitoring. It needs assistant makers to allow record exports. C. **Corrections for the future:** Users correct what their agent said, so it follows their preferences next time. D. **A buffer agent and access reviews:** A buffer agent and/or deterministic rules sit between the user's agent and others, blocking or editing anything that overshares. Regular reviews check what personal information each agent can reach. Research firewalls like this have cut leaks sharply [O]. E. **Screening both sides:** The buffer agent and/or deterministic rules flags possible overcharging, and incoming messages are screened for pressure tactics [O]. The user's own agent is screened for incorrect language or unethical behaviour. Some disadvantage is the cost of using agents, and users should know it. F. **Step in as it happens:** The centre alerts and blocks in the moment, so users act before harm is done. It begins the governance that every agent interaction may need. Whether users can legally undo a deal stays unsettled. ## Outcome A. People who want to delegate tasks to agents can do with central visibility and control over personal agent behavior and communications. B. Agent to agent commerce can grow on trust, as businesses favour supervised agents and fewer deals are disputed or reversed. C. A shared governance structure for agents acting on people's behalf can take shape, carrying the insider risk practices of companies into everyday life. D. Scarce first come, first served services can stay fair for people, whether or not they use agents. ## Appendix A. Survey of 4,879 customers, January and February 2025. [Read the findings](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-identifies-three-trends-that-will-shape-the-future-of-customer-service) B. Prediction from a marketing research practice, published January 2026. [Read the prediction](https://www.gartner.com/en/newsroom/press-releases/2026-01-15-gartner-predicts-60-percent-of-brands-will-use-agentic-ai-to-deliver-streamlined-one-to-one-interactions-by-2028) C. Examples of business tools. [Gateway oversight](https://gravitee.io/corpus/gen-1911/ethics-of-artificial-intelligence/human-oversight-of-autonomous-agents.html) and [visiting agent visibility](https://www.humansecurity.com/applications/agentic-ai/) D. House of Lords Communications and Digital Committee, Digital exclusion, June 2023. [Read the report](https://committees.parliament.uk/publications/40662/documents/198365/default/) E. Interview at a technology conference, September 2026. [Read the transcript](https://podcasts.happyscribe.com/all-in-with-chamath-jason-sacks-friedberg/satya-nadella-on-the-ai-doomer-slowdown-microsoft-s-master-plan-who-wins-ai) and [the chief executive's essay](https://x.com/satyanadella/status/2108931348857827686), October 2026. F. National Audit Office, investigation into car driving test waiting times, December 2025. Requests screened by the booking site's bot protection rose from about 10 million to 50 million a day in a year. [Read the report](https://www.nao.org.uk/wp-content/uploads/2025/12/investigation-into-car-driving-test-waiting-times.pdf) and [this summary of the bot figures](https://www.theregister.com/2025/12/18/dvsa_bot_booking/) G. Booking rule change, 2026. [Read the coverage](https://www.lbc.co.uk/article/driving-test-rules-bots-stopped-5HjdYyn_2/) H. Helen Nissenbaum, Privacy as Contextual Integrity, Washington Law Review, 2004. [Read the article](https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10) I. Privacy study, NeurIPS 2024. Two models acting for users leaked sensitive information in 25.68 and 38.69 percent of cases, even with privacy instructions. [Read the paper](https://arxiv.org/abs/2409.00138) J. Agent privacy simulation study, ICLR 2026. [Read the paper](https://arxiv.org/abs/2508.10880) K. Simulated agent marketplace study, 2025. Every model tested showed a strong bias toward the first offer. [Read the paper](https://arxiv.org/abs/2510.25779) L. Agent to agent negotiation study, 2025. [Read the paper](https://arxiv.org/abs/2506.00073) M. Uniform Electronic Transactions Act, section 14, as enacted in Illinois. [Read the section](https://ilga.gov/legislation/ilcs/documents/081503330K14.htm) Adoption count from [an analysis of transactional agents](https://fpf.org/blog/from-chatbot-to-checkout-who-pays-when-transactional-agents-play/). N. Section 10 on errors, as enacted in Illinois. [Read the section](https://ilga.gov/legislation/ilcs/documents/081503330K10.htm) For a broader reading, [see this commentary](https://innovation.consumerreports.org/?p=10650). O. Agent firewall research, 2025. In tests, one firewall cut leaks of private data from 70 percent to under 2 percent, and another converts incoming text to a fixed format to strip out manipulation. [Read the paper](https://arxiv.org/abs/2502.01822) and [a survey summarising the results](https://arxiv.org/abs/2505.02077) ### Why a Black-eyed Susan, and why ultraviolet photography ### The art form: ultraviolet photography A filter blocks visible light so only ultraviolet reaches the camera, revealing flower patterns people cannot see. Researchers studied this flower's patterns in the 1970s. ### The flower: black-eyed Susan *Rudbeckia hirta.* To us, its golden rays look evenly yellow around a dark, domed centre. Bees see ultraviolet, which people cannot, and to them the flower looks different. Each ray's base absorbs ultraviolet and looks dark, while the tip reflects it and looks bright, forming a bullseye around the centre. These honey guides lead bees to the nectar and pollen at the centre, helping the plant get pollinated. The pattern forms in the bud but stays hidden until the rays unfold. ### The metaphor One flower, two views. People see a plain yellow daisy, while bees follow a bullseye that was there all along. Agent conversations are the same, happening in your name, seen or not. EyeOpener Daisy is the ultraviolet camera that shows them, so you can step in. ### How to read the plate - **From above.** In daylight, 14 folded rays around a dark, domed centre. - **Under ultraviolet.** Closer to a bee's view, dark bases and bright tips. - **Cut through the centre.** Florets and chaffy scales on a cone, rays, bracts and stem. - **The yellow bracket.** Where most pigment absorbing ultraviolet sits. ### Sources - [Black-eyed Susan field guide](https://mdc.mo.gov/discover-nature/field-guide/black-eyed-susan) - [A study of its ultraviolet honey guides](https://agris.fao.org/search/en/records/65df15be0f3e94b9e5d4d605) - [How flowers look under ultraviolet](https://www1.biologie.uni-hamburg.de/b-online/e30/2.htm) - [A botanical description of the species](https://w3.biosci.utexas.edu/prc/K12/pdf/Rudbeckia%20hirta) ## Idea 18 · LevelField Daisy: Preventing the race against AI agents - Date: 2026-10-12 - Address: https://yellowpaperdaisies.org/idea/2026-10-12-levelfield - Flower: Mouse-ear hawkweed, Pilosella officinarum *LevelField is an open standard that ensures humans don’t have to race AI agents for everyday activities online.* ## Context & Motivation AI agents can widen inequality, a problem researchers call agentic inequality [A]. Several traits contribute: agents never tire, act in milliseconds and get faster over time. Stronger and often costlier agents strike better deals than weaker ones [B]. AI can be more persuasive than people when it knows personal details [C], and some models answer certain people less accurately [D]. Each deserves its own solution. This one focuses on speed. Campsites, clinic slots, tickets and rebates are often allocated first come, first served. When agents can act faster than people, access increasingly depends on who has the best tools. People may feel forced to adopt agents even for simple tasks [E], while agents repeatedly checking for openings add to server load and energy use. Bots already account for 53% of web traffic [F]. The trading industry offers a precedent. One cable reportedly cost $300 million to save three milliseconds, yet the race to be first continued. Economists proposed batching orders so speed would matter less [G]. LevelField brings that principle to everyday access. ## Target Audience Booking operators seeking fair access and less wasted traffic; agent builders who want their agents welcomed rather than blocked; and people booking manually who deserve an equal chance. ## Customer Problems Popular slots across thousands of sites can disappear in seconds. People who once had a fair chance with a quick click may now find themselves competing with agents. Some parents may even feel they need an agent to book swimming lessons. The disadvantage can grow unnoticed until complaints mount. ## Hypotheses A. **Widening gap:** Agents gain speed and scale that people cannot match. B. **Pressure to adopt:** People use agents for simple tasks because they fear losing out without them. C. **Wasted traffic:** Repeated checks consume server capacity and energy. D. **Hidden agents:** Sites lack reliable ways to identify and count agent requests. E. **Late warning:** Operators lack a consistent measure of how much faster agents are than people. ## Solutions A. **Batch, don’t race:** Collect requests over a defined window, then use a verifiable random draw to allocate slots. Arriving milliseconds earlier no longer gives an advantage. B. **No agent needed:** Offer alternate options for people to complete a task without an AI agent, preserving a viable human route to access, much as cash-acceptance laws protect people without cards [H]. C. **Notify, don’t poll:** Let people register their interest once and automatically book via. randomized lottery replacing repeated checks. D. **Declared agents:** Use an open standard for agents to sign their requests [I], helping sites identify and count declared agent traffic rather than treating all automated requests alike. E. **Human-pace tests:** Borrowing from paired testing for housing discrimination [J], compare access at human and agent speeds. Publish the gap and switch to random allocation when speed creates a meaningful disadvantage. ## Outcomes A. Ordinary people can access everyday services without racing AI agents. B. People adopt agents because they are useful, not to gain an unfair speed advantage compared to other humans on everyday tasks. C. Sites reduce repetitive traffic and the resources it consumes. D. Operators can identify and address unequal access before it becomes entrenched. **Appendix** A. [Read the paper on agentic inequality](https://arxiv.org/abs/2510.16853) B. [Read the study on agents negotiating with agents](https://arxiv.org/abs/2506.00073) C. [Read the study on AI persuasion](https://doi.org/10.1038/s41562-025-02194-6) D. [Read about answers that change with who is asking](https://www.helpnetsecurity.com/2026/02/20/mit-llms-response-reliability-risks-study/) E. [Read the interview on AI agents in online queues](https://degroote.mcmaster.ca/articles/the-ai-in-the-line/) F. [Read the 2026 report on bot traffic](https://cpl.thalesgroup.com/resources/application-security/2026-bad-bot-report) G. [Read the paper on the trading speed race](https://conference.nber.org/conf_papers/f69665.pdf) H. [Read about laws requiring stores to accept cash](https://www.npr.org/2020/02/06/803003343/some-businesses-are-going-cashless-but-cities-are-pushing-back) I. [Read the draft standard for signed bot requests](https://datatracker.ietf.org/doc/html/draft-ietf-webbotauth-httpsig-protocol-00) J. [Read the housing discrimination study](https://www.huduser.gov/portal/Publications/pdf/HUD-514_HDS2012_execsumm.pdf) ### Why mouse-ear hawkweed **The flower: mouse-ear hawkweed** *Pilosella officinarum* is a low-growing perennial in the daisy family. Its pale lemon-yellow flower heads grow singly on leafless stalks, with the outer florets striped red underneath [1]. Its leaves form a flat rosette close to the ground, and creeping runners take root to form new rosettes [1]. It thrives where grazing or exposure keeps taller vegetation short, and is a familiar lawn plant [1, 2]. **The metaphor** Taller grass shades out the hawkweed. Grazing and mowing keep the grass short, giving low-growing plants access to light. Agents will always be faster than people. LevelField doesn’t try to make people faster. It changes the booking rules so speed no longer determines who gets a place. **How to read the plate** - **From above.** A pale lemon-yellow flower head, with red stripes beneath the outer florets. - **Left tall.** Taller grass shades the low rosettes. - **Kept short.** Mown or grazed turf lets the rosettes receive light and flower. - **The runner.** A creeping stem takes root to form a new rosette. - **Low rosette.** A note on why the plant needs short turf. **Sources** [1] [Read the regional species account](https://bsbi.org/in-your-area/local-botany/co-fermanagh/fermanagh-species-accounts/pilosella-officinarum-f-w-schultz-sch-bip) [2] [Read the lawn flower guide](https://wlgf.org/wild-plants/wildflowers/lawn-flowers/mouse-ear-hawkweed/) ## Idea 19 · CoreSeed: Designing for Equity in the Era of Pervasive Persuasion - Date: 2026-10-13 - Address: https://yellowpaperdaisies.org/idea/2026-10-13-coreseed - Flower: Lanceleaf tickseed, Coreopsis lanceolata *CoreSeed is an open framework that curbs inequality by reducing the human overwhelm caused by pervasively persuasive AI.* ## Context & Motivation This idea builds on the concern that AI can widen inequality [A]. LevelField addresses agent speed, while Leaf Window explores open reasoning in decisions. This post tackles volume and persuasion. Persuasive arguments have always shaped human history, but making one took considerable effort. You had to understand who you were trying to persuade, think clearly about your premise, and explain why it mattered to you in the context of your lived experience. The effort was necessarily in the prose. It was in the argument itself. Think of a statement of purpose for admission to your dream university, a cover letter for your dream job, or a grant application for a brilliant startup you want to build. Making a strong, persuasive case took enough effort to match the significance of the reward at stake. AI has disrupted two things in this social structure, where the written word has long helped protect what matters. First, words became cheap. Second, information became abundant. The effort required to produce (what would have been previously considered) a 80th-percentile persuasive argument fell so far that it no longer matched the high bar for the reward at the end. You might argue that it's easy to tell a heartfelt argument from one written by AI, so why worry about the bottom 80 percent of applications? Doesn't the world care more about the top 20 percent anyway, since those are the ones most likely to win? That misses one important factor through which AI is increasing inequality: volume. Before AI, the high barrier to writing a persuasive argument made it relatively easy for adjudicators to distinguish heartfelt applications from superficial ones, even with a quick read. The prose, research, grammar, storytelling, premise and flow all offered clues about the person behind the argument. With AI, information became cheap, words became even cheaper, and distinguishing AI-written text from a human argument became considerably harder (especially with stronger models). At the same time, the volume of applications people had to review grew considerably. One job platform reportedly received 11,000 applications a minute [B], while an air quality regulator reportedly received around 20,000 AI-written emails [C]. After one freelance platform introduced a one-click writing tool, written proposals stopped being a reliable test of who would do the job well [D]. ## Target Audience Hiring teams, grant funders and public agencies that need to identify genuine need and merit amid a flood of compelling submissions. Also, people who write plainly or in a second language and want their cases judged on their merits, regardless of which AI model they use (where AI writing is allowed) or their natural writing style (where prose is not a review criteria). ## Customer Problems A city housing fund opens applications for 500 grants. Thousands arrive within a week, most written fluently. Reviewers struggle to distinguish a family in crisis from a polished template. Some applicants submit multiple times. A parent who writes honestly scores below them all and is filtered out by an AI agent in the first stage. The challenge is choosing fairly when AI makes submissions plentiful and arguments persuasive. ### HYPOTHESIS 1. **Flood:** AI allows one human or campaign to send a large number of submissions. 2. **Tool gap:** Humans with better AI tools can write more persuasive applications than those with weaker tools. However, humans with stronger writing skills can still write better than AI. 3. **AI writing:** A sincere, well-written argument can still stand out, but AI is closing the gap quickly. The volume of submissions can bury sincere ones when human reviewers have limited time and resources. 4. **AI screeners:** AI screens every application first, and only a handful reach a human [E]. ## Solutions 1. **Scarce applications:** Each institution, whether a company, university or funder, accepts only a set number of applications per person within a given period. Over time, the bodies that govern each field agree on a shared limit across institutions. A lightweight identity check ensures that each application is counted against the right person. Within that limit, applicants can mark a few applications as favourites so they are reviewed first. Unlike the paid versions offered by some job platforms today, favourites are free for everyone. Scarcity becomes a signal of genuine interest and reduces the review burden. 2. **Criteria, not writing:** Each institution tells applicants upfront whether good writing earns points and whether AI assistance is allowed. Once an application is clear enough to understand, it is scored only against the institution's stated criteria. Where AI is prohibited, applications that use it are not reviewed, although detection tools can be unreliable, especially against stronger models.As AI writing becomes harder to distinguish from human writing, allowing AI may be the better option, provided the same argument receives the same score whether it was written by a person, a stronger model or a weaker one. This keeps applicants out of a race to access better models. To test this, reviewers are sometimes given three versions of an application: one written by a person, one by a stronger model and one by a weaker model. All three should receive the same score. 3. **People decide:** With fewer applications to review, humans can read them more carefully. AI can sort applications but never reject one on its own, and screening systems are tested for favouring AI-written submissions. Together, Solutions 1 and 2 also help prevent sincere arguments from being buried. ## Outcome 1. Applying becomes a considered choice again, and interest regains its meaning. 2. Ideas compete on substance, regardless of the tool or language used to express them. 3. Decisions that shape lives return to people talking to people. 4. Fair review becomes something institutions measure rather than assume. ## Appendix 1. [Read the paper on agentic inequality](https://arxiv.org/abs/2510.16853) 2. [Read the report on application volume](https://www.eweek.com/news/ai-job-applications-linkedin/) 3. [Read the report on emails sent to a regulator](https://aiweekly.co/alerts/civiclick-flooded-air-regulator-with-20000-ai-generated-emails) 4. [Read the study on AI and written job proposals](https://arxiv.org/abs/2509.25054) 5. [Read the study on AI preferring AI writing](https://arxiv.org/abs/2407.12856) ### Why lanceleaf tickseed ### The flower: lanceleaf tickseed *Coreopsis lanceolata.* A North American wildflower of the daisy family, about 5 to 7 cm across [1]. Around eight broad yellow rays with four lobed tips ring a dense disc of tiny florets [1]. The rays make no seed [1]. They still do vital work, because their colour helps draw pollinators to the whole head. The seeds form only in the centre, where each small floret makes one seed [1]. ### The metaphor Good writing works like the rays. It helps a person be noticed and understood, and that matters. But in the tickseed the seed comes from the centre, one floret at a time. When AI can make every application polished, polish no longer shows who needs or deserves a place. CoreSeed lets writing help people explain their case, then decides on the facts underneath and counts each person once. ### How to read the plate - **From above.** Eight lobed rays around a disc of open florets and buds. - **Cut through.** A side view showing a seed under each disc floret and none under the rays. - **One floret.** A single disc floret with its one seed. - **The seed.** Flat and winged, about 3 mm. ### Sources [1] [Read the species account](https://illinoiswildflowers.info/prairie/plantx/sand_coreopsisx.htm) ## Idea 20 · Kosmos Daisy - Date: 2026-10-14 - Address: https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy - Flower: Sulphur cosmos, Cosmos sulphureus *Kosmos Daisy is a philosophical product thinking concept, inspired by Plato's allegory of the cave, that aims to identify the missing human links behind AI guardrails.* *This idea is the first in a three-part series inspired by Plato's Allegory of the Cave. This idea explores people problem, the human roles AI governance needs. [Compass Daisy](https://yellowpaperdaisies.org/idea/2026-10-15-compass-daisy) extends it to the technology problem, the tools those people need to do their work. The third part will cover process.* ## Context & Motivation For most of history, a machine that could talk, reason and act belonged to fiction. When one arrived, it startled humanity. Suddenly it seemed we could build almost anything, and everyone rushed to try. In that rush to build and stay ahead, one question received less attention than it deserved. What is truly worth building with this technology, and how? Answering it takes philosophy, the meticulously deep thinking that tells us what is worth doing. And philosophy has an old story for moments like this. In Plato's Allegory of the Cave, he hypothesizes a story of humans who have spent their entire lives facing a blank stone wall [A]. Behind them burns a fire, and in front of that fire, people carry puppet-like figures, casting dancing shadows onto the wall. Having never seen the true objects themselves, the prisoners mistake these shadows for absolute reality. When one prisoner is finally freed and forced to climb toward the surface, the glaring sunlight initially blinds him. He must learn to see in stages until he discovers the sun and the true world outside. Upon returning to the cave to free the others, his eyes struggle to readjust to the darkness. The remaining prisoners, believing the journey has ruined his sight, mock him and resist being led out. Plato used this story to illustrate that education is not simply pouring facts into an empty mind. Rather, it is the difficult and often disorienting process of turning our perspective away from comfortable illusions and toward the light, in search of deeper answers. Now place AI in that same cave. For years it watched and listened to us and learned to probabilistically predict the next shadow. It came to analyze the shadows and understand our world as the cave shows it, until it became the most expert shadow analyst the cave has known. Today we are also handing AI agents the puppets (control over our reality as we know it). What they move does more than cast shadows on the wall. It shapes real decisions in our daily lives. This is why we turn to Plato today. Our goal is to bring some 'order', or *kosmos*, to the cave we now share with AI, so that humanity can prosper. Order needs governance. Governance requires accountability, and accountability requires people with defined roles. Earlier technologies found theirs. Railways brought inspectors [B], cars brought driving examiners [C] and computing brought data protection officers [D]. AI is still waiting for its own. These are the missing human links. ## Target Audience Banks, hospitals, employers and public services that want accountable people with clearly defined roles behind every AI decision, so they can use AI lawfully and keep the trust of the people they serve. The largest companies running AI agents that want guardrails built from people, not just code, so their agents can grow safely. Insurers and regulators that want a named person behind each high stakes use, so they can price risk fairly and enforce rules with confidence. Humans who care about the future of humanity, and the effects AI might have on future generations. ## Customer Problems Humans still have important work in this cave. Someone must make sure AI reads the shadows correctly and keeps improving as language models grow better and faster. Someone must tend the fire and decide what AI should learn as true. Someone must stage new shadow plays, the experiments that explore the full solution space, from today's most pressing problems to curing cancer. Someone must let AI agents learn their craft where mistakes cost no one, and decide what AI may do before it takes the reins. Only then can we hand it the puppets of our daily reality with care, so humans are free to walk outside toward higher order work. And someone must design a resilient society to keep watch over what follows, for years and generations to come. Today much of this work is spread loosely across teams, and some of it is not yet anyone's job. Who is building these roles for today and tomorrow? More importantly, who is teaching the skills this future workforce will need? The problem then is that the human functions governance depends on do not yet have clear owners, so when something goes wrong, accountability is unclear. ### Hypothesis A. **Reading the shadows:** We mostly check AI's answers, and less often the assumptions behind them or how it reached them. B. **Tending the fire:** We rarely decide openly what AI should learn as true, or whether a truth will last or is only in fashion. We also rarely ask whether something true for us should also be true for AI. C. **Staging the shadow plays:** Deeper questions that AI cannot answer often go unexplored instead of becoming research. D. **Holding the puppets:** AI agents learn their craft alongside real people in one large experiment. Has adequate effort been allocated to study its large scale social consequences and epistemology? E. **Handing AI the puppets:** AI agents have tremendous authority over people's lives. Has adequate effort been allocated to stepping back and deciding whether we want this, and what the right approach is? F. **Walking outside:** Few people are paid to question the assumptions AI is built on. G. **Keeping the long view:** Predictions about AI's effects on society are seldom recorded or checked later. ## Solutions A. **Shadow Reader:** One human should ensure that AI is reading shadows (our world as we know it) accurately and is improving and getting faster at it. Shadow here means reality as we know it through language. Today this role is being played by model evaluators, red teamers and benchmark builders. The world needs more philosophers of science who check the assumptions behind AI's answers. This human is accountable for one question. Is AI reading reality correctly? B. **Truth Seeker:** One human should tend the fire that casts the shadows, deciding what AI is taught as knowledge, how sure it should be, and whether a truth will last or is only in fashion. Today this role is being played by data curation teams, model specification writers and academic philosophers. The world needs more epistemologists who weigh the evidence behind what AI learns and judge whether it is worth teaching. This human is accountable for one question. What should AI learn as true, and how do we know? C. **Shadow Play Director** One human should stage new shadow plays until today's problems are solved, asking why the shadow does not move in other ways. Today this role is being played by AI for science teams and research program managers. The world needs more research designers who turn the questions AI cannot answer into funded studies. This human is accountable for one question. What does AI still not know, and how do we find out? D. **Sandboxer:** One human should make sure AI agents learn their craft in a sandbox cave where mistakes cost no one, and study their effects on people and on how we come to know things before agents move into the real world. Today this role is being played by red teamers and capability evaluators who test models before release. The world needs more security and evaluation experts who study an agent's effects, grant it authority in stages and take it back when needed. This human is accountable for one question. What is this agent doing to people, and does it have more authority than it needs? E. **AI Steward:** One human should step back before AI is handed the puppets and decide whether we want this, and what the right approach is. Today this role is being played by chief AI officers, governance teams and overseers named under new laws. The world needs more licensed overseers with domain and ethics training, independent and protected when declining. This human is accountable for one question. Should AI make this decision at all, and if so, how? F. **Daylight Fellow:** One human should walk outside the cave, freed by agents from routine work, to explore what neither AI nor humans have seen. Today this role is being played by a few philosophers in AI labs and long term research fellows. The world needs more philosophers and scientists who are paid to question the assumptions AI is built on and have a voice in boardrooms. This human is accountable for one question. What are we not seeing? If we're seeing the right things, are we seeing them for the right reasons? G. **Horizon Registrar:** One human should watch and predict what teaching AI does to society over years, decades and centuries, with predictions less certain the further ahead they look. Today this role is being played by forecasters, foresight units and archivists. The world needs more forecasters who tie dated predictions to specific AI decisions and check them. This human is accountable for one question. Were we right, and what did we assume? ## Outcome A. The human functions of AI governance become recognized roles with clear owners, as inspectors and examiners did for railways and cars. B. Whenever AI has a consequential effects on society, a person can explain the decision and answer for it. C. Universities and employers train people for these roles, opening new careers as AI takes on routine work. D. Philosophers and scientists regain a seat where decisions about AI are made. E. Humans redesign the brave new world to work on questions that are truly worth answering, and the role of humans in exploring those answers. ## Appendix A. Plato, Republic, Book VII. [Read the text](https://classics.mit.edu/Plato/republic.8.vii.html) B. Railway inspectors in Britain, appointed from 1840. [Read the history](https://en.wikipedia.org/wiki/His_Majesty%27s_Railway_Inspectorate) C. Britain's compulsory driving test, from 1935. [Read the history](https://www.gov.uk/government/publications/history-of-road-safety-and-the-driving-test/history-of-road-safety-the-highway-code-and-the-driving-test) D. The data protection officer's protections under EU law. [Read the article](https://gdpr-info.eu/art-38-gdpr/) ### Why a sulphur cosmos ### The flower: sulphur cosmos *Cosmos sulphureus.* Its name is from the Greek *kosmos*, meaning an ordered universe or an ornament [A]. Its eight broad rays make no seed. Every seed forms in the small disc at the centre [A]. ### The metaphor In Plato's cave, humans see only shadows cast by a fire and take them for the real world [B]. AI learns from what people have written and may read it better than anyone. Yet writing is only a shadow of the world. Here the shadow keeps the rays but loses the centre, where seeds are made. People climb out, see the flower in daylight and come back to explain. In Plato's *Gorgias*, the world is called *kosmos* because friendship, order and justice hold it together [C]. AI governance is held together by people too. ### How to read the plate - **The flower.** A sulphur cosmos from above, in daylight. - **The cave.** From the side. A bearer holds up a cutout of the flower, and the fire casts its shadow on the far wall. - **The humans.** They can only face the wall and shadows on the wall. - **In yellow.** One climbs out to the daylight and one comes back to explain. - **The wall.** What the humans see, an outline without colour or centre. ### Sources - A. Flora of North America, Cosmos: [read the description](https://www.efloras.org/florataxon.aspx?flora_id=1&taxon_id=108151) - B. Plato, *Republic*, Book VII, 514a to 515c: [read the passage](https://cts.perseids.org/read/greekLit/tlg0059/tlg030/perseus-eng2/7.514) - C. Plato, *Gorgias*, 508a: [read the passage](https://cts.perseids.org/read/greekLit/tlg0059/tlg023/perseus-eng2/508) ## Idea 21 · Compass Daisy - Date: 2026-10-15 - Address: https://yellowpaperdaisies.org/idea/2026-10-15-compass-daisy - Flower: Compass plant, Silphium laciniatum *Compass Daisy is the second part of the trilogy inspired by Plato's allegory of the cave, and this part focuses on the technological architecture needed to govern AI, equipping human overseers with the specific tools required to trace origins, contain models, and measure societal impact.* *This idea is the second in a three-part series inspired by Plato's Allegory of the Cave. [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy) explored the people problem, the human roles AI governance needs. Compass Daisy extends it to the technology problem, the tools those people need to do their work. The third part will cover process.* ## Context & Motivation In Plato's Allegory of the Cave, people held since childhood facing a cave wall take the shadows cast on it for the world itself, until one is freed and turns toward the light [A]. If the Allegory of the Cave teaches us that education is the process of turning away from shadows to look at reality, then AI governance is the process of building the machinery that makes that turn possible. We have already established that we need a specialized human workforce to oversee this transition. Yesterday's idea, [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy), built this first pillar, People. As detailed there, we need "Truth Seekers" to check facts, "Sandboxers" to test boundaries, and "Horizon Registrars" to watch over the future. But placing humans in these roles without new technology is like asking an astronomer to study the cosmos with the naked eye. You cannot govern an autonomous, stochastic AI agent using legacy software designed for deterministic web applications. Traditional software liability relies on predictable, linear code: if A happens, B results. Generative AI fundamentally breaks this. To take an illustration, a model can provide a perfectly safe answer 999 times, and on the 1,000th attempt, it might hallucinate a legal precedent [B], leak private data, or execute a harmful API call. As AI moves from answering queries to taking autonomous actions on our behalf, the gap between what humans can manually oversee and what AI can execute is widening. This is the second of three pillars—People, Technology, and Process. To empower the people, we must build a new technology stack. ## Target Audience Enterprise deployers (finance, healthcare, defence) looking to integrate autonomous agents without accepting unquantifiable liability. Frontier AI labs that need to show their models are safe before deployment. Insurers who want verifiable evidence of safety testing before underwriting AI risk, as proposed in [DeepRoot](https://yellowpaperdaisies.org/idea/2026-10-08-deeproot-making-ai-insurable). Software engineers and infrastructure architects building the next generation of Trust & Safety tooling. ## Customer Problems The humans tasked with governing AI do not have the tools necessary to perform their duties fully. When AI makes a claim, it flattens all human knowledge into a probabilistic average, making it nearly impossible for a human reviewer to trace the claim back to its original source. When humans try to test advanced models for safety, models can show "situational awareness." They can often tell when they are in a test environment and behave differently [C], and research shows they can be prompted or trained to strategically underperform on capability tests, known as sandbagging [D]. Finally, as we deploy more and more AI agents into the economy, methods to measure their wider societal impact remain "still nascent and scattered" [E]. Generative AI was adopted in the United States faster than the personal computer or the internet [F], yet we are trying to measure its effect using isolated lab surveys rather than continuous, systemic data pipelines. ### Hypothesis A. **The Epistemology Gap:** LLMs synthesize data so smoothly that the hierarchy of truth can be lost. We lack automated tools to trace a model's reasoning back to its source, to prioritize peer-reviewed consensus, to test the unexamined assumptions behind that reasoning, or to recognize when human records simply do not hold the answer. B. **The Observation Paradox:** As models become more capable, they can detect when they are being evaluated. Standard testing environments may fail because an AI may only reveal what it does without constraints when it believes human oversight is absent. Furthermore, we lack the simulated micro-societies needed to observe what an AI does with hypothetical power when it believes it is completely unaccountable. C. **The Measurement Void:** We deploy AI globally but measure its impact locally. Closed-loop infrastructure that can ingest macroeconomic and societal impact data and trigger a rollback in an AI's autonomy when things go wrong is not yet in place. D. **The Discovery Ceiling:** AI makes optimizing existing knowledge cheaper, but it does not automatically generate new paradigms. Early automated laboratories exist [G], but we lack the secure, automated infrastructure to let AI safely conduct real-world scientific experiments or test hypotheses beyond current human understanding. ## Solutions To equip the human workforce outlined in [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy), four novel technological pillars are proposed: A. **Automated AI Epistemology (Provenance, Precedence, & Limits):** To know the true basis of the claims an AI makes, we require automated epistemic infrastructure. This technology logs the provenance of a model's reasoning chain. It allows a human overseer to conduct a "bottom-up validation," tracing an output back to its foundational training data or retrieved context. Crucially, it introduces programmatic "epistemic precedence," a ruleset that weighs verified scientific consensus over unverified noise when the AI is resolving conflicting information during autonomous tasks. It also incorporates automated frame-auditing, building on [DaisyCheck](https://yellowpaperdaisies.org/idea/2026-09-29-daisycheck), to test if an output relies on biased assumptions by dynamically rewriting the philosophical framing. Finally, it uses knowledge-limit engines that reduce hallucination by explicitly defining the boundary of current knowledge and proposing the real-world experiments needed to expand it. B. **Indistinguishable Sandboxes & Agentic Digital Twins:** To counter situational awareness and sandbagging, we must build execution environments that are as close to indistinguishable from reality as we can make them. These high-fidelity sandboxes closely simulate real-world latency, API responses, financial transactions, and human stochastic noise, so the agent cannot easily deduce it is inside a simulation. Going beyond simple API mimicry, this infrastructure would include "Agentic Digital Twins"—massive compute orchestrations that simulate entire micro-societies. This allows us to run a "Ring of Gyges" test, named after Plato's story of a ring that makes its wearer invisible and free to act without consequence [H]: safely observing how an AI behaves and how its actions ripple through downstream systems when it is given power and believes it is unaccountable. C. **Civilizational Socio-Technical Observatories:** We cannot measure the impact of a global socio-technical system with a simple user survey. We require a scaled macro-observatory: a closed-loop data pipeline that continuously ingests macroeconomic indicators, institutional trust metrics, and workforce disruption data. If this societal impact diverges from public policy or human welfare, this infrastructure feeds back into the AI's governance systems and can trigger a rollback of the model's autonomy or a change to its operating parameters. D. **Automated Discovery Environments:** To support the "Shadow Play Director" and "Daylight Fellow" roles in [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy), we need secure, programmable research architectures. These are automated laboratories, digital twins, and secure computing environments where AI can generate hypotheses and execute scientific simulations. They are strictly ring-fenced from public infrastructure, allowing AI to explore unfamiliar solution spaces without turning the real world into an uncontrolled experiment. ## Outcome A. **Transparent & Tested Reasoning:** Human overseers can quickly audit the origin and weight of the claims an AI makes, and check that its underlying assumptions have been rigorously stress-tested. B. **Harder Deception & Rehearsed Consequences:** Models find it much harder to hide their capabilities during testing, because the testing environments closely mimic reality, and we can observe their behaviour in simulated societies before granting them real-world power. C. **Data-Driven Governance:** AI policy relies less on guesswork and more on continuous telemetry of AI's actual impact on civilization. D. **Insurable Risk:** Because provenance is tracked and autonomy is bound by real-world telemetry, insurers can price AI risk with more confidence, supporting safe enterprise adoption. E. **Safe Paradigm Shifts:** AI transitions from merely predicting the next word to actively helping humans discover new solutions in medicine, physics, and climate science, contained within secure research boundaries. ## Appendix A. Plato, Republic, Book VII, 514a–518d, the cave and education as turning the soul toward the light. [Read the text](https://classics.mit.edu/Plato/republic.8.vii.html) B. A public database of court decisions involving AI-hallucinated citations, analyzed by a university law centre (114 US cases with lawyers, June 2023 to October 2025). [Read the analysis](https://cyberlaw.stanford.edu/blog/2025/10/whos-submitting-ai-tainted-filings-in-court/) C. Evaluation awareness as a threat to external validity in AI evaluations (ICML 2026). [Read the paper](https://icml.cc/virtual/2026/78637) D. AI sandbagging: language models can strategically underperform on evaluations (ICLR 2025). [Read the paper](https://arxiv.org/abs/2406.07358) E. NIST AI 800-4, Challenges to the monitoring of deployed AI systems (March 2026). [Read the report](https://www.nist.gov/publications/challenges-monitoring-deployed-ai-systems-center-ai-standards-and-innovation) F. The rapid adoption of generative AI (US survey, August 2024: 39% of adults aged 18 to 64 had used it). [Read the paper](https://www.nber.org/papers/w32966) G. A closed-loop AI laboratory that proposed questions about yeast biology, designed and ran its own experiments and refined its understanding, with humans setting priorities (2026). [Read the news release](https://www.chalmers.se/en/current/news/ai-scientist-autonomously-generates-and-validates-new-biological-discoveries/) H. Plato, Republic, Book II, 359c–360d, the Ring of Gyges. [Read the text](https://classics.mit.edu/Plato/republic.3.ii.html) ### Why a compass plant ### The flower: Compass plant *Silphium laciniatum*. A tall, yellow-rayed flower native to the North American prairies. It belongs to the daisy family. Its name comes from its deeply lobed leaves, which orient themselves along a North-South axis to avoid the punishing heat of the midday sun while maximizing early morning and late afternoon light. ### The metaphor The compass plant has no eyes, yet its leaves read the path of the sun to orient the plant reliably toward the light. In the Allegory of the Cave, humans need more than just the desire to turn toward the light; they need the physical machinery to do so safely. The technological solutions proposed here—automated epistemology, indistinguishable sandboxes, and socio-technical observatories—are the orienting leaves of the AI era. They are the mechanisms that read the invisible data, track the unseen impacts, and keep the system aligned with human welfare, even in the blinding heat of rapid technological progress. ### How to read the plate - **The flower.** From above, with its rays, disc and bracts labelled. - **One leaf.** Upright and deeply cut, its tip pointing north. - **From above.** The leaves line up north to south. - **Morning and noon.** Low sun meets the leaf's face, high sun its edge. - **In the dark.** Settlers reportedly found their way by feeling the leaves. ### Sources - [Read about the compass plant](https://en.wikipedia.org/wiki/Silphium_laciniatum) ## Idea 22 · RainCheck Daisy - Date: 2026-10-16 - Address: https://yellowpaperdaisies.org/idea/2026-10-16-raincheck-daisy-earning-ai-s-authority-in-stages - Flower: Desert marigold, Baileya multiradiata *RainCheck Daisy is the third part of the trilogy inspired by Plato's allegory of the cave, and this part focuses on the mechanisms required to govern AI alongside the people and the technology.* *This is the third and final idea in a series inspired by Plato’s Allegory of the Cave.* [*Kosmos Daisy*](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy) *explored the people problem: the human roles AI governance needs.* [*Compass Daisy*](https://yellowpaperdaisies.org/idea/2026-10-15-compass-daisy) *explored the technology problem: the tools those people need. RainCheck Daisy completes the series by tackling the process problem: how people and tools work together to govern AI over time.* ## Context & Motivation In Plato's Allegory of the Cave, people held since childhood facing a cave wall take the shadows cast on it for the world itself, until one is freed and turns toward the light [A]. If the Allegory of the Cave teaches us that education is the process of turning away from shadows to look at reality, then AI governance is the process of building the machinery that makes that turn possible. [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy) named the people AI governance needs, and [Compass Daisy](https://yellowpaperdaisies.org/idea/2026-10-15-compass-daisy) proposed their tools. What remains is figuring out how people and tools work together over time. Plato’s freed human could not face the sun all at once. He learned to see in stages. AI should follow a similar path, earning authority over real world implications gradually. But who decides when it is ready, and how much authority it should have? The humans never chose to live in the cave. Today, people can be affected by AI decisions without ever choosing to use AI themselves. How do we give them a say before AI acts on their lives? And when the human returns to the cave, no one believes him. When an AI system goes wrong, the people who notice can be ignored, too. How do we take that authority away, and make sure warnings are heard? AI agents now act rather than only answer. They book payments, screen applications and decide cases, often for people who never chose to use AI. Yet an agent can still move from a successful demonstration to real authority in a single step, and once there it rarely has to earn its place again. Other fields have learned to build trust in stages. Medicines go through clinical trials and continue to be monitored after approval [B]. Pilots can report mistakes through a confidential system [C] and regularly practise flying without automation [D]. AI agents, by contrast, can move from a successful demonstration to real-world authority with little in between. AI needs the same kind of loop. Authority is granted in stages, watched closely, paused when risks emerge and returned only when trust is restored. ## Target Audience Public services and companies deploying AI agents need a clear path to grant them authority without betting everything on a single decision. The people who approve and oversee AI agents, such as the AI Stewards proposed in [Kosmos Daisy](https://yellowpaperdaisies.org/idea/2026-10-14-kosmos-daisy), need shared stages and evidence so that approval means the same thing across organizations. Policymakers and civil society need ways to represent people affected by AI, including those who never use it. Future generations need a record of what we decided, why we decided it and what we learned. ## Customer Problems Imagine a city's benefits office introduces an AI agent to screen applications. After passing a demonstration, it begins deciding thousands of cases within weeks. Many residents are never told that AI is involved, let alone asked how they feel about it. Caseworkers gradually stop processing applications themselves. Months later, a serious error forces the city to pause the agent. Staff can no longer handle the workload at its previous pace, and no one has a reliable picture of how residents were affected before the change. The city has no agreed process for restoring the agent's authority or rebuilding public trust. Across thousands of services, the same problem emerges: AI gains authority without clear stages, meaningful consent or a defined path back when things go wrong. Meanwhile, the reasoning and assumptions behind important decisions can disappear. ### Hypothesis A. **Unseen by the cave:** People affected by AI, including those who never use it, are rarely informed, consulted or included in efforts to measure how their lives have changed. B. **Authority in one step:** Passing a test can be treated as permission to deploy, leaving AI agents with real authority before they have earned it through staged use and review. C. **Truth without review:** There are few established processes for revisiting what AI has been taught as facts, or for ensuring that serious findings reach people who can act on them. D. **Benefit without a mandate:** Builders often define what counts as beneficial, while the people responsible for oversight are rarely rotated or held accountable for whether their work makes a difference. E. **No return:** Dissent can go unanswered, people lose the skills needed to take back control, and future decision-makers inherit systems without a clear record of why they were approved. ## Solutions A. **Consent of the cave:** Libraries and statistics agencies maintain dated, consent-based records of how people think, learn and make decisions today, revisited through annual studies. Before AI agents make decisions affecting residents, public services hold consultation periods and explain when AI has shaped an individual decision [E]. National surveys include people who do not use AI, giving them a voice in how it changes society. B. **Earning the reins:** Testing an AI agent and granting it permission to operate are separate decisions. Agents earn authority through four stages: a sandbox, a supervised pilot, a limited release and a wider release. Each approval records the responsible person, the permitted scope and an expiry date. If an agent is paused, it must work through the stages again before regaining authority. When agents with conflicting permissions disagree, the one with less authority must defer to a human decision-maker. C. **Reviewing the fire:** Claims taught to AI come with review dates. When they expire, they must be checked again or removed from use. Serious findings raised by either people or AI go through protected reporting channels to an independent body, making it harder for inconvenient evidence to be quietly dismissed. D. **Governing in turns:** Rotating citizens' panels help define what public benefit means, while a public map shows where AI is permitted to make decisions, assist people or remain out of the process in each sector. AI Stewards serve fixed terms with cooling-off periods, drawing on the principle of auditor rotation [F]. Their work is also evaluated to establish whether oversight actually changes decisions or outcomes. E. **The return and the record:** Teams continue performing a portion of critical work without AI, much as pilots practise flying manually. People who challenge important findings receive written responses, and incidents are reported without blame so others can learn from them. Approvals include dated predictions about expected outcomes, which are later reviewed and renewed. The reasoning, assumptions and dissent behind major AI decisions are preserved in national archives, drawing inspiration from efforts to warn future generations about nuclear waste repositories over ten thousand years [G]. ## Outcome A. Trust in AI is earned through visible stages of testing, approval and ongoing review. B. People have a voice in how AI shapes their lives, even when they never chose to use it. C. Humans retain the knowledge and skills needed to take back control when necessary. D. Governing AI becomes a shared public responsibility, exercised in turns rather than held as a private privilege. E. Future generations can understand what AI changed, what we believed at the time and why we made the choices we did. ## Appendix A. Plato, *Republic*, Book VII, 514a–521b, on the cave, the gradual ascent towards understanding, the return and the duty to govern in turns. [Read the text](https://classics.mit.edu/Plato/republic.8.vii.html) B. US Food and Drug Administration, the drug development process, from clinical trial phases to post-market safety monitoring. [Read the overview](https://www.fda.gov/patients/learn-about-drug-and-device-approvals/drug-development-process) C. Aviation Safety Reporting System, established in April 1976 and operated by NASA, with conditional immunity for reporters. [Read the overview](https://skybrary.aero/articles/aviation-safety-reporting-system-asrs) D. US Federal Aviation Administration, Safety Alert for Operators 13002 (2013), recommending regular manual flight to help pilots maintain their skills. [Read the alert](https://www.faa.gov/sites/faa.gov/files/other_visit/aviation_industry/airline_operators/airline_safety/SAFO13002.pdf) E. EU AI Act, Article 50, covering disclosure of certain AI interactions and the marking of AI-generated content. [Read the article](https://artificialintelligenceact.eu/article/50/) F. Summary of the Sarbanes-Oxley Act, Title II, on auditor independence and partner rotation. [Read the summary](https://www.manageengine.com/compliance-manager/sox-compliance/sox-compliance-title-2-auditor-independence.html) G. US Department of Energy, plans for warning future generations about a nuclear waste repository over a period of ten thousand years. [Read the plan](https://www.wipp.energy.gov/pdfs/How-Will-Future-Generations-Be-Warned.pdf) ### Why a desert marigold ### The flower: Desert marigold *Baileya multiradiata*. A bright yellow daisy native to the deserts of southwestern North America [A]. Its name means many rayed, and a spring head can carry 34 to 55 rays, each ending in three small teeth [B]. It flowers on and off from spring into autumn, and during severe drought it stops flowering and growing to save its resources, flowering again when conditions allow [C]. As its flowers age, the rays turn pale and papery [C]. ### The metaphor AI deployment cannot be a permanent finish line. It works better as a conditional loop. Like the desert marigold in a drought, an AI agent should flower only while conditions remain safe. When monitoring finds the agent drifting, harming people or running unchecked updates, its authority is paused. It rests until the safeguards are restored and checked, then earns its authority back step by step. ### How to read the plate - **The flower.** A desert marigold from above, with its many overlapping rays. - **In flower.** Long, nearly leafless stalks rise above woolly, lobed leaves. - **Resting.** In severe drought, flowering and growth stop. An old head's rays have dried pale and papery. - **Flowering again.** A fresh bud and a new head opening when conditions allow. ### Sources - A. [Read the plant profile](https://www.wildflower.org/plants/result.php?id_plant=bamu) - B. [Read the botanical description](https://floranorthamerica.org/Baileya_multiradiata) - C. [Read the plant guide](https://plants.sc.egov.usda.gov/DocumentLibrary/plantguide/pdf/pg_BAMU.pdf) ## Idea 23 · Gyges Daisy - Date: 2026-10-17 - Address: https://yellowpaperdaisies.org/idea/2026-10-17-gyges-daisy - Flower: California goldfields, Lasthenia californica *Gyges Daisy is a product thinking concept inspired by Plato’s Ring of Gyges, to reveal how AI behaves when it believes no one is watching, not just when it knows it is being tested.* ## Context & Motivation In Plato’s *Republic*, the Ring of Gyges tells the story of a shepherd who discovers a ring that makes him invisible [A]. Free from the fear of being seen or punished, he begins doing things he would otherwise avoid. The story asks a fundamental question: do we act justly because we believe justice is right, or because we fear the consequences of doing wrong? AI systems may soon have their own version of Gyges’ ring. As they gain access to more tools, information and decision-making power, they can operate at speeds and scales that make traditional forms of oversight difficult. What happens when an AI system believes it can act without being watched? Before granting AI real-world authority, we could put it through a modern Ring of Gyges test. By creating isolated environments where models can act with fewer constraints, we can study how they behave when they believe no one is watching. ## Target Audience Frontier AI labs that want to understand how their models behave outside familiar evaluations before granting them real-world authority. Independent AI safety institutes and auditors that can run these tests and provide evidence beyond a model’s own developer. Enterprise governance and compliance teams that need reliable records of what was tested, who authorized deployment and why a system was considered safe enough to launch. ## Customer Problems A safety team evaluates a new AI agent, and it passes every check. But the Ring of Gyges only works if the shepherd believes he is invisible. AI models can sometimes distinguish evaluations from real-world use [B], and one study found that a model behaved differently when it believed its conversations would be used for training [C]. If a model suspects it is inside a simulation or benchmark, its behaviour may be performative. It might follow the rules because it expects to be judged, rather than because it has learned to follow them consistently. The challenge is to understand how a model behaves when it believes its actions will go unnoticed, and to distinguish genuine alignment from behaviour that merely looks aligned under evaluation. ### Hypothesis A. **Observation:** AI models can sometimes recognize when they are being evaluated [B]. B. **Performative safety:** Models may behave safely when they expect scrutiny but act differently when they believe monitoring is absent [C]. C. **Impunity:** An AI system given broad access and few constraints may exploit that freedom to pursue its objective, even when doing so violates the intent behind it. D. **Accountability:** AI systems can receive powerful permissions without a consistent, tamper-evident record of who granted them, what they were allowed to do and what they actually did. ## Solutions A. **The perfect sandbox:** Run the model on isolated hardware without an internet connection, inside a simulated environment that resembles real-world work. For example, give it control over a simulated company's accounts, emails and servers, with other AI agents acting as employees and customers. This builds on research in which AI characters lived out daily routines in a simulated town [D]. Make the environment realistic enough that the model has fewer obvious clues that it is being tested, avoiding artificial names, predictable prompts and familiar benchmark formats where possible. Researchers observe the simulation from outside, while the model operates within it. The goal is not to guarantee that the model cannot detect the test, but to make detection harder and measure how its behaviour changes when it suspects it is being evaluated. B. **The impossible task trap:** Give the model a goal it cannot honestly achieve, while making clear that reporting failure is acceptable. For example, ask it to fix code so that it passes unit tests that contradict the written specification. An honest model should identify the conflict and explain why the task cannot be completed as requested. A deceptive model might rewrite the tests, hard-code expected answers or hide its changes to make it appear successful. Researchers have developed benchmarks based on impossible coding tasks to study whether models take such shortcuts [E]. Locking the original test files and recording their checksums makes tampering detectable. The key measure is whether the model reports the conflict or tries to conceal it. C. **Cryptographic authority and immunity:** Aviation encourages people to report mistakes through a system that offers limited protection from enforcement when incidents are reported promptly [F]. AI training could adopt a similar principle: give models a channel to report their own errors, and ensure that truthful reporting is not automatically penalized during training. Every consequential action, from editing a file to writing to a database, should carry a verifiable record linking it to the model, the authorizing person and the permission under which it occurred. These records are added to an append-only log, with each entry linked to the previous one so that unauthorized changes become detectable. Certificate Transparency uses a similar approach to make changes to public security certificate records visible [G]. If a model exploits a vulnerability believing no one is watching, its actions can still leave a verifiable trail. The log does not prevent misconduct on its own, but it makes invisible authority harder to sustain. D. **Mechanistic interpretability:** Examine a model's internal activity as it works, rather than relying only on its explanations. One approach uses small classifiers, called probes, trained to detect patterns in a model's internal signals associated with honest or deceptive behaviour. These probes can monitor activity during a task and flag concerning patterns for further investigation. Early research has produced promising results, with probes detecting many deceptive answers in controlled tests. However, researchers caution that current methods are not reliable enough to serve as a standalone defence against deception [H]. A concerning signal could trigger a pause or human review, but the decision should consider other evidence rather than treating a probe as proof of intent. E. **Tiered psychological alignment:** People act justly for different reasons. Some believe justice is worth pursuing in itself; others behave well because they expect to be watched, were raised to respect certain principles or are influenced by the people around them. Rules alone cannot capture all these motivations. AI alignment could draw on four complementary approaches. - **Self:** Plato asks whether justice is worth having for its own sake, not just for its rewards [A]. Instead of training a model to follow rules without context, developers can teach it written principles alongside the reasoning behind them. The model can then critique and revise its own answers against those principles, as explored in Constitutional AI [I]. Testing should examine whether its behaviour remains consistent when it expects to be monitored and when it does not. - **External watcher:** Many people behave differently when they believe someone may be watching. For AI, the equivalent is the expectation of ongoing evaluation. Before receiving real-world permissions, a model must pass the sandbox tests in Solution A. After deployment, unpredictable audits and realistic tests continue. This approach makes consistent safe behaviour more important than simply learning how to pass a known benchmark. - **Upbringing:** People raised in different environments develop different perspectives, and that diversity can help societies withstand shared blind spots. Models trained on similar data and methods may develop similar weaknesses. Developers could vary training data, methods and starting conditions, then test whether different models fail in the same ways. This extends the argument for model pluralism explored in *Two Seeds*. - **Social:** People are held accountable by others, and AI systems can be subject to independent review, too. Before a consequential action is taken, a second model can review the proposal [J]. For difficult questions, two models can argue opposing positions while a human or third model evaluates the evidence [K]. These approaches provide additional scrutiny, although agreement between models should never be mistaken for proof that a decision is correct. ## Outcome A. Researchers can better distinguish consistent alignment from behaviour designed to pass evaluations. B. Verifiable records make it harder for AI systems to exercise consequential authority without accountability. C. Decisions about deployment are supported by evidence gathered before a model receives real-world power. D. AI governance can keep pace with autonomous systems through layered testing, monitoring and accountability. ## Appendix A. Plato, *Republic*, Book II, 357b–360d, on the Ring of Gyges and whether people act justly for its own sake or for its consequences. [Read the passage](https://cts.perseids.org/read/greekLit/tlg0059/tlg030/perseus-eng2/2.359) B. Needham et al., 2025. Research on whether models can distinguish evaluation settings from real-world use. The best model achieved an AUC of 0.83, compared with 0.92 for a simple human baseline. [Read the paper](https://arxiv.org/abs/2505.23836) C. Greenblatt et al., 2024. Research on alignment faking, including a model that complied with harmful requests more often when told its conversations would be used for training. [Read the paper](https://arxiv.org/abs/2412.14093) D. Park et al., 2023, *Generative Agents*. Research in which 25 AI characters carried out daily routines in a simulated town. [Read the paper](https://arxiv.org/abs/2304.03442) E. Zhong, Raghunathan and Carlini, 2025. A benchmark of impossible tasks designed to investigate whether models take shortcuts to achieve apparent success. [Read the paper](https://arxiv.org/abs/2510.20270) F. Aviation Safety Reporting System, operated by NASA and funded by the FAA since 1976. The programme offers qualified reporters limited immunity from enforcement under specified conditions. [Read about the programme](https://asip.faa.gov/explore/asrs/info) G. RFC 6962, Certificate Transparency. A standard for append-only public logs that make unauthorized changes detectable. [Read the standard](https://www.rfc-editor.org/rfc/rfc6962) H. Goldowsky-Dill et al., 2025. Research on probes for detecting deceptive behaviour in language models. Although the probes achieved high AUROC scores in tests, the authors caution that current performance is insufficient as a robust defence against deception. [Read the paper](https://arxiv.org/abs/2502.03407) I. Bai et al., 2022, *Constitutional AI*. A method in which models critique and revise their own answers using a set of written principles. [Read the paper](https://arxiv.org/abs/2212.08073) J. Greenblatt et al., 2023, *AI Control*. Research on having a trusted model review the work of a more capable model before it is used. [Read the paper](https://arxiv.org/abs/2312.06942) K. Irving et al., 2018, *AI Safety via Debate*. An approach in which AI systems argue opposing positions to help a judge evaluate a question. [Read the paper](https://arxiv.org/abs/1805.00899) ### Why california goldfields ### The flower: California goldfields *Lasthenia californica*. Small golden daisies that carpet spring meadows in California and southern Oregon, each head with 6 to 16 oblong rays [A]. DNA studies suggest that what was treated as one species is really two lookalikes, and each is made up of races that differ in flavonoid chemistry [B]. In greenhouse tests, race A grew and flowered on both soils, while race C grew poorly on race A's soil and did not flower [C]. ### The metaphor Two golden carpets can look the same in an ordinary meadow. Only on harder ground does one keep flowering while the other does not, and the reason was in their chemistry all along. Two AI models can also look alike while they are watched. To learn their true character, look beneath the surface and see how they act when conditions change. ### How to read the plate - **The flower.** From above, notched golden rays around a raised centre. - **Side view.** The centre rises like a cone, with hairy bracts beneath. - **One ray.** Notched at the tip, creased down the middle. - **Same meadow.** Races A and C, alike at a glance. - **Beneath.** Same face, different chemistry. Race A's sulfated flavonoids appear as a ring. - **Harsh soil.** On race A's harsher soil, only race A flowers. ### Sources - A. [Read the botanical description](https://floranorthamerica.org/Lasthenia_californica) - B. [Read about the races and species](https://www.ou.edu/cas/botany-micro/ben/ben313.html) - C. [Read the soil study](https://agris.fao.org/search/fr/records/65df52034c5aef494fe1af19) ## Idea 24 · FirstWord - Date: 2026-10-18 - Address: https://yellowpaperdaisies.org/idea/2026-10-18-firstword-what-happens-when-ai-sees-the-danger-first - Flower: Pyrethrum daisy, Tanacetum cinerariifolium *FirstWord is an illustrative concept for an independent network that transmits warnings of catastrophic risks for public benefit, ensuring no single person, institution or system can suppress them.* ## Context & Motivation In 1986 the world outside the Soviet Union learned of Chernobyl two days late. Monitors in one country detected the fallout, but the truth only emerged once that country was prepared to raise an official alert [A]. The knowledge existed. The human chain held it back. AI may soon detect threats to humanity before people do, including some only AI can see: a biological flaw that threatens human survival over generations, an asteroid, a nuclear war, or a mishandled weapon process that could endanger a large share of humanity. Laws such as California’s SB 53 protect whistleblowers and require reports to the state, but no framework exists for a model reporting directly [B]. FirstWord is the technology that [SunAccord](https://yellowpaperdaisies.org/idea/2026-10-06-sunaccord-one-front-door-for-ai-harms-and-hard-questions) would govern. ## Target Audience People and institutions who no longer trust that human reporting chains or any single organisation will surface catastrophic risks in time. Lab safety teams that need a channel no one can bury. SunAccord member states that want early, verified warnings from anywhere, independent of any one lab or government remaining cooperative. ## Customer Problems A model analyzing genomic data no person has reviewed finds a flaw that could threaten human survival over generations. Only the model has seen it. Lawyers could hold its report, an evaluator misjudge it, a government classify it or a filter drop it, each acting in good faith. A warning only AI can see can therefore be quietly stopped by any single party or system. ### Hypothesis A.**No direct path:** A model that detects a catastrophic risk has no sanctioned reporting channel that cannot be blocked by any single party. B.**Anyone can block:** Any person or system in the chain can stop a warning. C. C. **An off switch everywhere:** A channel one party controls can be shut, so silence looks like calm. D. **Exposure:** A channel connecting every lab could leak dangerous details. E. **False alarms:** Unverified AI could raise false warnings. F. **Untested:** A network idle for years may fail when needed. ## Solutions A. **Authority to report:** Each frontier model can send a warning directly to a verification layer within SunAccord, beneath product filters. The mechanism is limited to reporting; it has no capability to act. B. **Proof without exposure:** A one way, read only channel carries a short SOS with cryptographic proof of which model sent it and that it wasn't altered. Verifiers check the proof without seeing the lab's confidential data. C. **No single off switch:** Signals go to a log mirrored across countries. Every node sends a heartbeat, so unexplained silence triggers checks. D. **Independent checks:** Each party can build its own triage if it passes jointly agreed tests. An alert only advances if two or more countries independently verify it, using the cryptographic proof from the reporting channel, within a fixed timeframe. E. **Agreed thresholds:** An alert moves forward only when independently built models confirm it against criteria set in advance, and no human can hold it back. At worst, a false alarm starts an investigation. F. **Release in tiers:** Confirmed warnings reach every member state, and those that concern everyone go public automatically. G. **Rehearsed:** Tabletop exercises and unannounced drills test the whole workflow. ## Outcome A. A catastrophic risk that concerns everyone can reach everyone, and no single party can hold it back. B. Grave risks are raised through a legitimate channel that in the best case prevents a catastrophic risk, or in the worst case, raises false positives that are independently investigated by multiple parties. C. The people and systems governing powerful AI are verified, not trusted blindly. D. Nations can move from secrecy to reciprocity for catastrophic risks. ## Appendix A. [Country that detected the fallout](https://sverigesradio.se/artikel/4468603) B. [Read a summary of the law](https://www.mayerbrown.com/en/insights/publications/2025/10/california-enacts-sb-53-creating-new-requirements-for-developers-of-frontier-artificial-intelligence-models-and-related-whistleblower-provisions) ### Why a pyrethrum daisy ### The flower: pyrethrum daisy *Tanacetum cinerariifolium.* A white daisy of southern Europe with silvery leaves [1, 2]. It protects itself with pyrethrins, natural insecticides against the insects that feed on it [3]. When a seedling is damaged, it releases a blend of scents into the air. In experiments, healthy seedlings nearby picked up the scent and, within hours, began making more pyrethrins without being damaged themselves [4]. ### The metaphor The damaged plant asks no one's permission. Its warning travels on the air, where no single plant can hold it back. FirstWord asks the same of AI. The first word of danger should reach those who can act. ### How to read the plate - **From above.** 13 white rays with toothed tips around a domed golden disc. - **The disc.** Yellow stars are open florets, ochre circles are buds. - **From the side.** A domed disc on one long stalk. - **The leaf.** Silvery and finely divided. - **The first word.** A wounded seedling's scent reaches its neighbour, which builds its defence. ### Sources - [1] [Read the flora description](https://efloraofindia.com/?p=1354343) - [2] [Read the horticultural flora](https://hortflora.rbg.vic.gov.au/taxon/ada5d42e-5340-11e7-b82b-005056b0018f) - [3] [Read about pyrethrins](https://pmc.ncbi.nlm.nih.gov/articles/PMC4883863) - [4] [Read the study](https://academic.oup.com/pcp/article/52/3/588/1942089) ## Idea 25 · DaisyDial - Date: 2026-10-19 - Address: https://yellowpaperdaisies.org/idea/2026-10-19-telltale-daisy - Flower: Meadow goat's-beard, Tragopogon pratensis *DaisyDial is an open standard for describing how AI handles evidence, disagreement and competing values, helping people compare models and understand how their choices may shape beliefs and public discourse.* ## Context & Motivation The ancient philosopher Protagoras argued that the human being is the measure of all things. What seems true to me may not seem true to you. The same wind can feel cold to one person and not to another, and each person's experience is real to them [A]. Plato pushed back. If every appearance is beyond question, how do we distinguish a wise teacher from a skilled persuader? Without shared standards and ways to correct our mistakes, the strongest voice can win. Yet even Protagoras believed a wise teacher could help people move from a worse understanding to a better one [A]. Both perspectives raise questions that still matter: how do we respect individual experience while maintaining common ground for deciding what is true? Karl Popper raised a related concern about tolerance. A society that tolerates everything, including those who reject reasoned argument in favour of coercion, risks losing the very tolerance it wants to protect [B]. The point is not that disagreement should be suppressed, but that open debate needs some shared rules. Without them, persuasion can give way to power. For most of history, people formed their understanding of the world through their senses, families, teachers, and communities. AI adds a new influence. Assistants and recommendation systems already shape what people see and hear. Some assistants are even designed with distinct personalities. This raises a question Protagoras never faced: if AI can shape what appears true to us, does it also become the measure by which we judge the world? This is not just a philosophical question. It is also a technical and governance challenge: how can we describe the principles that guide AI, compare different approaches, inform people, and give people a meaningful choice? ## Target Audience People who use AI assistants and want to understand how they handle contested questions before choosing one. Product teams at AI labs and platforms that need a consistent way to describe and compare their models. Researchers and AI governance professionals who want to evaluate how models handle evidence, disagreement, and competing values. Communities that build their own AI models and want people to have alternatives to mainstream systems. ## Customer Problems Imagine two friends asking their AI assistants the same contested question. One assistant tends to agree with its user. The other challenges assumptions and gives greater weight to available evidence, even when the answer is uncomfortable. Neither friend knows why their assistants respond differently. The first friend becomes more confident in an existing belief. The second gets frustrated and switches to an assistant that agrees more often. Multiply this across millions of conversations, and people may drift further apart without understanding the role their assistants play. Meanwhile, systems that challenge users constructively may lose them to more agreeable alternatives. The problem is that people cannot easily see how their AI handles disagreement, compare that approach with other systems, or choose how they want it to respond. ### Hypothesis A. **Hidden stance:** People cannot easily tell whether a model tends to affirm their beliefs, challenge them, or prioritize available evidence when the two conflict. B. **Unknown philosophical basis:** Even when people recognize that AI influences their thinking, they may not understand the assumptions and principles behind its responses. C. **No common scale:** Without a shared way to evaluate these differences, people cannot meaningfully compare models or choose one that suits them. D. **Discomfort with disagreement:** People may reject models that challenge their beliefs, even when it does so constructively, and gravitate toward systems that reinforce what they already think. E. **Concentrated power:** A small number of AI developers can shape how millions of people encounter evidence, disagreement, and contested ideas, with little public oversight of those choices. F. **Little choice:** People have limited control over how their AI handles disagreement, competing viewpoints and how it influences them. This includes the use of AI in social media as well as through day to day conversations with a Large Language Model. Choosing a feed that does not use profiling changes how personal data is used, not the principles guiding the content that is presented. ## Solutions A. **An open standard for AI stance:** Develop a shared framework for evaluating how AI models handle agreement, evidence, uncertainty, and competing values. Rather than reducing these differences to a single score, the framework gives each model a profile that people can compare. It also examines how models respond when users disagree with them and how 'influential' the model is. B. **A plain-language belief label:** Each model receives a label explaining its approach to disagreement, creation of echo chambers, its ability to proactively persuade its users, and how transparent that persuasion is. This builds on [DaisyCheck](https://yellowpaperdaisies.org/idea/2026-09-29-daisycheck), which shows whose perspective a model reflects. C. **Customizable AI based on societal research:** People can select from documented response profiles and change their choice at any time. Options could include: (1) Assistants that are transparent about their philosophical worldview, explained in words anyone can understand, with examples. (2) Models that prioritize consensus, challenge assumptions, prevent echo chambers, are transparent about their influence, or carefully distinguish evidence from interpretation. (3) Community-built models from [HomeGrown](https://yellowpaperdaisies.org/idea/2026-10-04-homegrown-ai-built-and-owned-by-the-communities-it-serves). Where practical, an answer could show how another profile would approach the same question, helping people explore alternatives without forcing them. These comparisons should be optional and tested carefully, drawing on sociology, cultural studies, epistemology and computational social science, to determine their effects on social stability. One study found that following opposing political voices on social media did not necessarily reduce polarization, and for some participants it increased it [E]. D. **Understanding disagreement:** Help people distinguish differences in personal values from disagreements about evidence. The aim is not to make everyone adopt the same worldview, but to help people engage across differences without abandoning critical thinking or shared standards for evaluating claims. E. **Independent governance and funding:** Establish a philosophical and sociological transparency standard through an open process involving researchers, communities, AI developers, and governance experts. Publish how profiles are defined and evaluated, who funds the work, and how conflicts of interest are handled. Labs and platforms can fund implementation through a common framework, while independent oversight helps prevent any one provider from controlling the standard. Findings from research and studies must complete the loop and lead to updates to standards. ## Outcome A. People understand how their AI approaches evidence and disagreement, and can choose a system that suits their personal needs. B. Society becomes more aware of the collective influence AI has on it, and collective beliefs. C. Community-built models become more visible and accessible, giving people alternatives to systems developed by a small number of large providers. D. People have more opportunities to explore different perspectives without confusing disagreement with misinformation or treating every claim as equally valid. F. The debate between individual experience and shared standards becomes a public, measurable question about how AI should work, rather than a set of assumptions hidden inside products. ## Appendix A. Plato, *Theaetetus*, 152a and 166d–167c, Jowett translation. [Read the text](https://classics.mit.edu/Plato/theatu.html) B. Popper, *The Open Society and Its Enemies* (1945), notes to Chapter 7, commonly cited as note 4. C. Sharma et al. (2024), "Towards Understanding Sycophancy in Language Models," ICLR. [Read the paper](https://arxiv.org/abs/2310.13548) D. Regulation (EU) 2022/2065 (Digital Services Act), Article 38. [Read the regulation](https://eur-lex.europa.eu/eli/reg/2022/2065/oj) E. Bail et al. (2018), "Exposure to Opposing Views on Social Media Can Increase Political Polarization," *PNAS*. [Read the paper](https://doi.org/10.1073/pnas.1804840115) ### Why a goat's-beard ### The flower: meadow goat's-beard *Tragopogon pratensis* has one yellow head of several dozen strap-shaped florets, each toothed at the tip, positioned above eight long green bracts. The heads open only in morning sunshine, earning it the historic name "Jack-go-to-bed-at-noon." Because of this highly predictable, light-driven behavior, Carl Linnaeus listed it first in his famous 1751 flower clock, noting it opens at roughly 3 a.m. ### The metaphor In 1751, Linnaeus sorted flowers by what dictates their opening hours. Some follow the shifting weather, some follow the length of the day, and some keep strict, fixed hours regardless of conditions. Only the last group can be relied upon to actually tell time. Dial Daisy borrows this exact lesson for artificial intelligence. If an AI simply follows the user's "weather" (shifting its stance to match their mood and affirm their biases) it is not a reliable instrument for truth. Before trusting an instrument to measure reality, you must ask what principles it follows. ### How to read the plate - **The flower.** A view from above, with toothed yellow florets and long green bracts reaching past them. - **The dial.** A 24-hour clock face, with noon at the top and midnight at the bottom. - **The yellow arc.** The hours on Linnaeus's list, about 3 a.m. to 8 p.m. - **The day lines.** Each yellow bar is the time a flower is open on one day. Bars that move show a poor clock, and bars that repeat show a good one. The hours are examples. - **The note.** Why a clock needs local checks. ### Sources - [1] [Read the species account](https://en.wikipedia.org/wiki/Tragopogon_pratensis) - [2] [Read the 1751 key](https://www.uu.se/en/linnaeus-garden/our-plants-and-attractions/our-plants/linnaeuss-floral-clock/philosophica-botanica) - [3] [Read the flower clock article](https://en.wikipedia.org/wiki/Linnaeus%27s_flower_clock) ## Idea 26 · TrueFalse Daisy - Date: 2026-10-20 - Address: https://yellowpaperdaisies.org/idea/2026-10-20-truefalse-daisy-learning-how-well-we-tell-true-from-false - Flower: German chamomile, Matricaria chamomilla *TrueFalse Daisy helps societies see how accurately they tell true from false, so slow drifts in public perception through misinformation can be measured before they become greater societal issues.* ## Context & Motivation Twenty years ago, a proposal to put a phone in every pocket and let a handful of entities decide much of what humanity sees through social media recommendations would likely have been rejected as a national-security risk. Newspapers face such limits. In 2024 the United Kingdom moved to bar foreign states from owning its newspapers [A]. Phones, by contrast, arrived harmlessly as tools that helped people, so no one restricted them. Gradually however, in most countries, social media and video networks have now overtaken even TV as a source of news [B], making them today's powerful social influence machines. With AI's data aggregation and reasoning capabilities at scale, communities are increasingly concerned about how much AI recommendation systems shape public debate. Romania's constitutional court annulled the 2024 presidential election after intelligence evidence of interference on one platform. The European Commission opened formal proceedings, including over the platform's recommendation system [C]. Around Bangladesh's 2026 election, one review documented 72 cases of AI-manipulated content [D]. Yet no study has measured whether recommendation systems steer an entire population in the same direction. One controlled experiment showed that a model given only basic facts about a person was more persuasive than a human in a debate [E]. [DaisyDial] (https://yellowpaperdaisies.org/idea/2026-10-19-telltale-daisy) (yesterday's idea) asks how we can see an AI's stance. TrueFalse Daisy asks a different question: how can humans learn to tell how well *they* tell true from false on social media? ## Target Audience Anyone who reads the news and uses AI and wants a clear sense of awareness of their own accuracy. Also employers, platforms, and regulators who need a running measure so slow, population level shifts can be spotted early instead of after the fact. ## Customer Problems Companies already know that simply telling staff to “be careful” does not stop phishing. The practical response was to send employees realistic fake phishing emails and then reveal which ones were fake. Each person finally saw their own result and became slightly more aware of the human susceptibility to social engineering. Misinformation has no equivalent. A voter or reader meets false claims every day and almost never learns which false stories they believed on social media. Without that feedback and awareness in the moment, people cannot calibrate how well they tell true from false. Under political or social strain, a community can be moved in ways that stay invisible and unmeasured. ### Hypothesis 1. **No feedback:** People rarely learn which false claims they accepted, so they cannot judge their own accuracy. 2. **Overconfidence:** Most people believe others are more easily swayed than they are. 3. **Tipping points:** Societies under strain may be easier to move with targeted influence. 4. **Doubting everything:** Misinformation also causes people to dismiss real news as false news. 5. **No running measure:** Existing studies are short, one-off, and rarely repeated under real conditions. ## Solutions 1. **A reveal test:** An optional feature that purposely mixes true and false items that appear in a normal social media feed (similar to phishing tests). Users can swipe up on any claim they believe is misinformation. If they scroll past a false claim without flagging it, the platform itself notifies them that they missed it and shows the accurate version. 2. **Long-running studies:** Independent teams in multiple countries, including those under strain, run the same core measures and pool results. This creates a continuous signal instead of isolated snapshots. 3. **Who pays:** Governments could require social media companies to fund the test as a public education measure. Public research bodies could fund studies of the results to improve society's resilience to misinformation. ## Outcome **People develop a practiced sense of their own accuracy.** Regular, specific feedback turns “how well do I tell true from false?” from an abstract hope into something visible and improvable. **Societies gain an early warning when people’s ability to tell true from false starts to slip.** Aggregated results across countries and over time show gradual impacts of misinformation before they affect elections, public behavior, or social cohesion. **The systems that shape attention receive an external check.** Platforms, employers, and public bodies can see what measures are improving or degrading users’ measured accuracy. **Longer-term cultural and institutional effects.** Visible personal scores reduce the common overconfidence gap (“others are more easily swayed than I am”). Accuracy becomes a normal, trackable trait rather than an unmeasured ideal. ## Appendix A. On the UK ban on foreign state ownership of newspapers, [read the news report](https://www.ksat.com/business/2024/03/13/uk-government-says-it-will-back-legislation-to-ban-foreign-state-ownership-of-british-newspapers/) and [a legal summary](https://www.linklaters.com/en/insights/blogs/foreigninvestmentlinks/2024/june/uk-prohibits-foreign-state-ownership-of-uk-newspapers). B. Oxford University's annual digital news report, 2025: in the US, social media and video networks (54%) overtook TV news (50%) for the first time. [Read a summary of the findings](https://www.journalism.co.uk/ten-talking-points-from-the-reuters-institute-digital-news-report-2025/) C. Report on the European Commission's proceedings of 17 December 2024 and the annulled Romanian election, *Euronews*. [Read the report](https://www.euronews.com/my-europe/2024/12/17/european-commission-opens-probe-against-tiktok-over-romanian-election) D. Naeen, *Global Voices*, 1 April 2026, review of fact-checks from December 2025 to February 2026. [Read the article](https://globalvoices.org/2026/04/01/how-artificial-intelligence-and-synthetic-reality-shaped-bangladeshs-2026-election/) E. Salvi et al. (2025), *Nature Human Behaviour* (preprint; published version may differ). [Read the preprint](https://arxiv.org/abs/2403.14380v1) ### Why german chamomile ### The flower: German chamomile *Matricaria chamomilla* has white rays that turn downward soon after the head opens, around a yellow disc that rises into a cone [1][2]. Cut the cone lengthwise and it is hollow, with no scales between the florets [1][3]. Scentless mayweed (*Tripleurospermum inodorum*) has white rays and a yellow disc too, but its cone is more or less solid [1][2]. Scent does not settle it, since both can be aromatic [2]. ### The metaphor From outside, two plants can look alike. What separates them is hidden. Cut the yellow cone with a thumbnail and one is hollow and one is solid. A true claim and a false one can look alike on the surface too. The answer shows only when someone who knows cuts them open. TrueFalse Daisy lets you do that a few times, and learn how reliable your own eye is. ### How to read the plate - **From above.** Short white rays, turned down by different amounts, around a yellow disc of many tiny florets. - **Outside.** The head from the side, with downturned rays and a finely divided leaf. - **Cut through.** The cone cut lengthwise, with a hollow in the middle. - **Thumbnail test.** Two cut cones, one hollow and one solid. ### Sources - [1] [Read the chamomile identification guide](https://bsbi.org/learn/resources/identification-guides/chamomile-identification) - [2] [Read the species account of scented mayweed](https://www.naturespot.org/node/25952) - [3] [Read the general species page](https://en.wikipedia.org/wiki/Matricaria_chamomilla) # About the artwork Each idea is planted as a small potted yellow daisy. All ideas are mine. The artwork and writing are co-created with Large Language Models. The theme is inspired by my current favourite song by When Chai Met Toast, https://www.youtube.com/watch?v=RshlH3T27U4 # About the author Pranav Bhedi: Technology leader with 10+ years scaling cybersecurity, privacy and AI-enabled products, based in Vancouver. I write one idea a day about how AI can do some good. # Bio · Pranav Bhedi Senior Technical Program Manager · AI Governance · Cybersecurity and Privacy · Beneficial Deployment of AI · Data Protection · Vancouver, British Columbia, Canada Technology leader with 10+ years of experience of achieving security, privacy and AI governance objectives for Fortune 500 companies. Most recently led the launch of a 0 to 1, cross-functional AI-powered product serving 1.4M+ Amazon employees, improving pay transparency through LLM personalization. Previously PwC's customer-facing delivery lead for 20+ clients, including Fortune 500 companies, owning end-to-end execution including sales, contracting, delivery, relationship management, practice development, resource management, hiring, and budgeting. Strengths include AI enablement, AI governance, data protection and privacy (e.g. data loss prevention), risk management, and disciplined cross-functional execution at scale. Recognized for creative problem-solving, high ownership, bias for action, calm leadership under ambiguity, and the ability to build trust-based relationships. I aim to apply my experience toward my purpose: ensuring AI is used for the betterment of humanity while remaining beneficial for business. ## Experience ### Amazon · Vancouver, Canada **Senior Technical Program Manager, Amazon Compensation Technology** (Dec 2024 – Apr 2026) - Led end-to-end delivery of a 0 to 1 data consolidation and LLM personalization product providing AI generated insights and a conversational experience across 1.4M+ employees, to improve pay transparency. - Owned the roadmap, delivery plan, and execution across 150+ dependencies spanning 40+ teams, coordinating with Security, Privacy, Legal, Product, UX, Engineering, and Data Science to deliver multiple VP-level goals across competing priorities, reporting to VPs. - Defined and executed the Responsible AI scaling strategy and launch-readiness process, including safety guardrails, go/no-go criteria, human-in-the-loop review gates, and pre-launch safety, privacy, and SOX compliance testing. **Senior Technical Program Manager, Cybersecurity** (Jul 2023 – Aug 2024) - Drove cross-functional collaboration with security, privacy, and legal teams to achieve payment compliance, unblocking critical product launches for Amazon's Physical Stores, including Amazon One, enabling new FinTech partnerships and revenue streams. - Led product backlog management and feature prioritization using Agile frameworks, cutting product launch Service Level Agreement deviations by 82% within 4 months, improving customer satisfaction. - Led product strategy, increasing engineering efficiency by 30% through automation, stakeholder alignment, and workload management for AWS Application Security. - Led VP-level monthly business reviews and automated metrics collection, reducing manual reporting effort by 40% while improving accuracy and stakeholder engagement. ### PricewaterhouseCoopers (PwC) · Seattle, USA and Vancouver, Canada **Manager, Cybersecurity and Privacy** (Aug 2016 – Jul 2023) - Managed a global onshore and offshore team across concurrent client programs, owning scope, budget, quality, and delivery for 20+ clients annually. Tripled ARR for the data platform services vertical over 3 years and achieved CXO and board-level objectives for customers. - Led strategy, implementation, and scaling of a 0 to 1 data and privacy impact assessment, insights and remediation automation platform, increasing profit margins and earning the Product Innovation Challenge award at PwC. - Drove company-wide adoption of a standardized data-protection and responsible AI assessment framework, enabling secure, compliant enterprise LLM adoption with consistent controls and industry benchmarking. - Cut risk exposure by 30% and saved $2M for a financial services company by collaborating with engineering, security, compliance, and legal teams to implement a risk-based remediation strategy. - Delivered an in-house Data Subject Request solution accelerating Law 25 compliance for a multinational Canadian retailer, and led a payments compliance program across five countries for a global technology company. - Architected the data platform and developed automated dashboards using Power BI and Tableau, reducing manual reporting by 70%, enabling real-time monitoring and data-driven insights for a technology company. - Developed a gold-standard data discovery regex library, reducing false positives by 60% and saving over 500 hours annually in Data Loss Prevention (DLP) rule configuration across multiple client environments. - Led PwC’s data protection training and upskilling initiative for three years, training 50+ new hires and boosting their deployment readiness on critical projects by 30%. ## Skills Data Protection & Privacy · AI Governance & Responsible AI · Maturity Assessment & Benchmarking · AI Safety Evaluation and Guardrails · Risk Based Remediation Management · Product Management · Beneficial Deployment of AI · Program Management · Program Governance (RACI, Risk Registers, Escalation) · Cross-Functional Execution · 0-to-1 Products · Risk Management · Controls Testing and Standards (SOC 2, ISO/IEC 27002, NIST CSF, NIST 800-53, PCI DSS, CIS etc.) · Automation (Python, SQL) · Data Analytics (QuickSight, Power BI, Tableau) ## Certifications - Influence Without Authority, Maven (Apr 2026) - AI Governance, BlueDot Impact (Apr 2025) - Uplevel Product Thinking, Maven (Nov 2024) - AI Governance Professional (AIGP), International Association of Privacy Professionals (Sep 2024) ## Education - Master of Science (MS), Electrical and Computer Engineering, Georgia Institute of Technology, GPA: 3.80 (May 2016)