What we'll build
One AI system where every way of knowing works on the same problem, to reach what none of them could reach alone.
Tradition-grounded AI agents, an open global forum and one record that grows, tested first on an open benchmark. Here is the design we want to develop with the AI community, why now, and the questions we want to settle together first.
The system: three scales, one record
WhatIfWe is a project in pluralistic, public-interest AI: collective intelligence infrastructure designed so that genuinely different ways of knowing can deliberate together and reach what none of them reaches alone. The system has four parts that feed each other. Depth and scale do different jobs, and it needs both.
- Machine scale: an open benchmark. Sixteen tradition agents work through thousands of real problems, and practitioners judge blind whether each tradition is still recognisable, against a single model asked to play all of them. It needs no forum and no new data, only published texts used with permission. It could stop the project, which is why it comes first. The WhatIfWe Open Benchmark sets out its design.
- Forum scale: an open online forum. Thousands of people take part from their own tradition and in their own language, in many concurrent groups. Grounded agents carry each contribution into the terms of the others. vTaiwan maps opinions this way. The forum would do it for ways of knowing.
- Depth: long-running groups. Small groups of people who have gone furthest in their traditions work for months on one real problem. They calibrate the agents, and they show what the forum cannot: what happens at the deepest level of each tradition.
- One record. All three feed one growing record of agreements, open disagreements and new approaches, each traceable to who contributed what. The question that decides everything is whether later work starts further on.
Scale without depth would repeat what deliberation platforms already do well. Depth without scale would stay a seminar. The record is what joins them.
What the space does
Inside the forum and the long-running groups, the space is more than a place where contributions meet. It plays roles that no meeting, platform or chatbot plays today, and it plays them together. People decide and sign. The space provokes, proposes, holds, translates, remembers and protects.
A companion for each tradition. Each participant works beside an AI companion built for their own tradition. It answers only from that tradition’s library: its texts, teachings and scholarship, held in a store its community controls. A council of knowledgeable practitioners keeps it credible. They choose what goes into the library, check what the companion says, and can correct it or take it down. It can be shaped to the participant’s own school, lineage and language, so that a Sufi of one order and a jurist of another hear their own tradition, not a general one. It says where its tradition is divided, and it cites every passage it uses. The closer it comes to the tradition a person actually lives, the more they can trust it to carry their voice into the conversation, and the more traditions can take part across the world.
Live interpretation. Participants speak in their own languages. Speech is interpreted live, and the speaker’s own voice can always be heard beneath it. Systems that interpret speech across nearly 100 languages within seconds already exist (Meta’s Seamless models, 2023), and video calls now carry live speech translation that keeps the speaker’s voice (Google Meet, 2025). The space adds a second layer: it carries meaning between frameworks as well as between languages, and each author confirms the rendering of their words. Where a language or a setting needs a human interpreter, the space makes room for one.
The roles the space plays.
| Role | What it does |
|---|---|
| Mediator | States each side in the other’s terms and asks: is this what you meant? It mediates the meeting, never the outcome |
| Keeper of shared intention | When conflict prevails, returns people to the question they chose to work on, in their own words, and to what their own tradition says about staying with difficulty |
| Creative catalyst | Sets side by side what has never met, such as a passage no one in the room would have found, and leaves the making to people |
| Steward of attention | Slows a heated exchange instead of rewarding it, and makes sure every voice is heard |
| Memory | Keeps every open question alive, with its history, so whoever comes next starts where the last group stopped |
| Window | Answers an institution’s request for a cross-tradition reading: signed, sourced, with its disagreements and no verdict |
| Project steward | Helps the community choose its next work: a project goes forward when many traditions each say, in their own words, why it matters. It is fed by what the record knows, and what it learns comes back |
| Guardian | Merges nothing, authors nothing, and takes nothing a tradition has not offered |
What this makes possible. Whoever comes next starts where you stopped. When the room turns hard, the space holds both sides at full strength and softens neither. And a question that today has no answer, we need a cross-tradition read on this, becomes a request with one.
Why now
Four things have changed in the last few years, and together they make this something that can be built and tested.
Deliberation with AI works for very large numbers. DeepMind’s Habermas Machine tested AI-written group statements with 5,734 people in the UK. Participants preferred them to statements by human mediators and rated them more informative, clear and unbiased, and independent raters found them clearer and more informative (Tessler and colleagues, Science, 2024). vTaiwan has drawn in more than 200,000 people and contributed to 26 laws. California launched a permanent state platform, Engaged California, in February 2025. A coalition launched at the Paris AI Action Summit that month aims to involve 10,000 citizens from more than 100 countries in its first cycle.
The cost has collapsed. Running a language model at a given quality has become about ten times cheaper every year, roughly a thousandfold over three years (a16z, November 2024). Retrieval, orchestration and translation between languages are now standard parts.
The machinery already shapes how we feel about each other. Engagement ranking on Twitter served more anger and more hostility toward the other side than a plain chronological feed (Milli and colleagues, PNAS Nexus, March 2025). Reranking the live feeds of 1,256 people for ten days during the 2024 US election moved how they felt about the other party, in either direction (Piccardi and colleagues, Science, November 2025). The lever is real, and it can be pointed at understanding.
And AI is narrowing how we think. A 2026 study of language models on ethical dilemmas from Indigenous contexts found that they adopt a tradition’s persona on the surface while defaulting to individualistic conclusions (Merrikhi, 2026 preprint), which it calls a persona illusion. Another found that an average American was a closer match to Indian cultural values than any model trained for India, and that better prompting did little to change that (Agarwal and colleagues, 2025). In science itself, researchers who use AI publish more and are cited more, while the range of topics science studies narrows (Evans and colleagues, Nature, January 2026). Simulations of models trained on each other’s output find that an ecosystem of diverse models slows this kind of knowledge collapse, up to an optimal level of diversity (Hodel and West, 2025 preprint). The more the world thinks through AI, the more this matters.
The column nobody fills
Every serious system in this field measures something real. None measures whether different ways of knowing survive the meeting.
| System | Scale reached | Measures whether each way of knowing stays distinct? |
|---|---|---|
| Habermas Machine (DeepMind) | 5,734 participants | No: agreement is the outcome |
| Polis and vTaiwan | Hundreds of thousands | No: opinions within one frame |
| Talk to the City | About 1,000 per deployment | Diversity of opinion, afterwards |
| Global Dialogues (Collective Intelligence Project) | 6,000+ people, 70+ countries | No |
| Abundant Intelligences | Community programmes | Within one tradition |
Each measures what it was built to measure. A minority view inside a shared frame is a different answer to the same question. The case nobody has tested is a different question: a tradition that does not accept the terms in which the problem was posed. That is the case planetary problems keep presenting, and it is the one we want to build for. The field maps these and other initiatives, each with the limit it reaches.
Principles we would start from
Nothing is built yet. These are the commitments we would bring to a first conversation with technical partners, and each is open to challenge.
- Each tradition keeps its own voice. Any AI that represents a tradition draws only on that tradition’s own material, kept apart from the others.
- Nothing is blended. Every contribution stays signed and traceable. People may agree completely; the machine may not merge them.
- Disagreements are kept, with their history, so that later work can start from them.
- The traditions judge. Whether a tradition survived the meeting is decided by people formed in it, never by the system.
What we want to settle with partners
These are the questions we would work through first, together. We would rather answer them with people who know the field than present answers of our own.
- Can an AI reason from inside a tradition, and not only use its vocabulary? How would we test it, on which problems, and who should judge?
- What should the first test be? It should be cheap, open to anyone to reproduce, and able to show that the idea does not work.
- How large can a forum grow before depth is lost, and what keeps people from very different traditions engaged in it?
- How do we earn trust from the communities most wary of AI? People who distrust AI take part less in AI-mediated deliberation (Jungherr and Rauchfleisch, 2025).
- Who owns what the traditions contribute, and on what terms? Indigenous-led programmes such as Abundant Intelligences have done the most work on this, and we want to learn from them. Our intention is that contributions are held by a non-profit foundation.
- How do we stop the system telling everyone what they already believe, when it brings each person’s own tradition into the conversation?
- Which problem should come first, and which traditions would make a result on it meaningful?
Whatever we test, we will publish what it shows, including results that go against us.
What we are asking of the AI community
- AI labs, for model access, compute and evaluation expertise, and for a serious test of whether their models can reason from inside a tradition.
- Researchers in multi-agent systems, evaluation, deliberation and collective intelligence, to help design the benchmark and the forum, and to stress-test the architecture.
- Builders of deliberation platforms, to test whether this layer can sit alongside tools that already work with very large numbers of people.
- Funders who want to know whether AI can widen how humanity thinks.
Partners sets out where else we need partners. To begin, write to hector@ibarraconsulting.com.