Validation year, 2026 · v1.2 · updated September 2026
Public principles for how frontier AI behaves
The companies building frontier AI write down how their models should behave. OpenAI calls the document a model spec; Anthropic calls it a constitution. Polaris Collective is a non-profit working to make these documents trackable, contestable and independently tested. First topic: concentration of power.
The problem
A model spec is the public document in which an AI developer states how it intends its models to behave. It is the one document that makes those decisions legible and testable from outside, and it is starting to carry weight in regulation. Yet few frontier AI companies publish one, those that exist say little about the highest-stakes behaviours, and nobody outside the companies regularly tests whether models follow them.
The stakes are highest where model behaviour meets democracy: whether a model will help a person or small group seize or entrench power. That is where we start.
What we do
Track
We keep a public record of published model specs: what each one says, how it changes over time, and how regulation treats them.
Recommend
We research and publish what a model spec should contain and how it should be governed, for frontier AI companies and for policymakers.
Evaluate
We run independent, recurring evaluations of whether models follow their spec, with a focus on hard cases and on models working as autonomous agents.
Where we are
This is early. 2026 is our validation year: first publications arrive in autumn 2026, followed by feedback from companies, regulators and adjacent organisations, and a decision on scaling in early 2027. Our tracking tool, the AI Constitutions Index, is in development, with a public launch at our Westminster event on 20 November. Polaris Collective is a non-profit in formation, with a team based in London and Brussels. The work is anchored at the Oxford Institute for Ethics in AI through its Accelerator Fellowship Programme, and supported by the Talos Network incubation track and the LISA Founders Residency.
Launching at our Westminster event · Friday 20 November 2026
AI Constitutions Index
Read every published constitution in full in the doc reader, passage by passage, with the parts that bear on each behaviour marked.
- What the constitutions say
- How constitutions are governed
- How constitutions are regulated In preparation
- Adherence of models to constitutions In preparation
What the constitutions say looks at the document: how clearly it settles conflicts between its own rules, and how deeply it covers the behaviours a model should show, a list that is not exhaustive and that we are still adding to. How constitutions are governed looks at the company: what it publishes about its constitution, and how far it holds its models to it.
Keep me posted on the launch.
Convenings
Model specs sit inside a wider question: who has standing to shape how frontier AI is built and governed, inside the companies and outside them. Alongside our core work, we help bring researchers, policymakers and practitioners together on that question, with academic partners at the University of Oxford and Harvard Law School. A first closed, invitation-only convening takes place in London in late 2026, and we are exploring further convenings in 2027. We also contribute to academic research on how companies govern AI in practice.
Friday 20 November 2026 · Palace of Westminster, London
AI: Worker Voice as Critical Internal Condition
A one-day parliamentary conference on governance and worker voice inside frontier AI labs.
Where this goes
2026
First publications, feedback from companies, regulators and adjacent organisations, and a decision on scaling in early 2027.
2027
A public model spec tracker, regular publications on the state of model specs, and a first full evaluation run across frontier models.
Beyond
Further topics beyond concentration of power, and a reporting standard that regulators can reference.
We work in the open and hold ourselves to explicit milestones; we will update this page as results land.
Team
Executive Director. Governance & Operations.
Spent five years co-founding and running Wequity, a B2B AI company. Oxford PPE; MSc Economics UCLouvain, with a thesis on firm democratisation.
Executive Director. Research & Tooling.
CTO of Wequity for four years, and previously an AI researcher at Hexo Labs. MSc Aeronautical Engineering, UCLouvain and Keio.
Academic direction.
Research Director FNRS-UCLouvain, Senior research associate of the Center for Labor and a Just Economy at Harvard Law School and Distinguished research fellow of the Institute for Ethics in AI at the University of Oxford.
In 2025-2026, she chaired the International High-Level Expert Committee on Democracy at Work, set up by the Spanish Government. She was nominated an Accelerator Fellow of the Accelerator Fellowship Programme 2026-27 of the Institute for Ethics in AI at the University of Oxford.
We are recruiting a senior technical AI safety evals researcher, potentially at co-founder level for the right person. Write to us.
What we're looking for
We want feedback from the people closest to the problem: people working on model specs and safety inside frontier AI companies, evals researchers, and policy staff. Interest and scepticism are equally useful, and anything you share is treated in confidence.
We are also looking for senior advisors from AI safety, evaluations and policy, and for introductions to funders working on concentration of power or AI governance capacity. More detail is available on request.
Stay updated
Occasional updates on what we publish and where the work goes next. No more than one email a month.
We use your address only to send these updates. Unsubscribe any time by replying to any email.