Public principles for how frontier AI behaves

A model spec states how a company intends its models to behave: the closest thing a model has to a public constitution. It is the one document that makes those decisions legible and testable from outside, and it is starting to carry weight in regulation. Yet as of September 2026, only OpenAI, Anthropic and Alibaba publish one, and Alibaba's does not yet cover its most capable systems. Published specs cover consumer and API products, changes are only partly logged, and adherence rests on the companies' own claims. The gap between how much these documents matter and how little independent scrutiny they receive is the problem we work on.

The stakes are highest where model behaviour meets democracy: whether a model will refuse to help a person or group seize or entrench power, through coups, election manipulation or the erosion of institutional checks. That is where we start.

This is early, and already producing. Our first publications arrive at the end of September 2026: the first state of model specs memo, the first governance recommendations, and first evaluation results on whether a model's refusal holds when a plan arrives one innocuous step at a time. We are taking over the AI Character Index, an open tool that reads a model spec and pulls out every passage covering a given behaviour, and will keep it current across every published spec. In November we co-host a working day at Westminster on frontier AI governance and concentration of power. The work is anchored at the Oxford Institute for Ethics in AI through its Accelerator Fellowship Programme, and supported by BlueDot, the Talos Network incubation track and the LISA Founders Residency.

1. To the end of 2026

   First publications in September, feedback rounds with companies, regulators and adjacent organisations, the Westminster convening in November, and a go/no-go decision on scaling in January 2027.

2. 2027

   The public spec tracker live with change alerts, quarterly memos, content principles on power reviewed by external experts, and the first full evaluation run across frontier models, published as a paper with an open repo.

3. Beyond

   Evaluations that test models working as autonomous agents over long horizons, a reporting standard regulators can cite, and further spec topics beyond concentration of power.

We work in the open and hold ourselves to explicit milestones; we will update this page as results land.

• Gabriel Levie, co-founder.

  Governance pillar: tracking, recommendations, fundraising and external work.

  Spent five years co-founding and running Wequity, a B2B AI company. Oxford PPE; MSc Economics, with a thesis on firm democratisation.

• Samuel Verboomen, co-founder.

  Research pillar: evaluations and tooling.

  CTO of Wequity for four years, and previously an AI researcher at Hexo Labs. MSc Aeronautical Engineering, UCLouvain and Keio.

• Isabelle Ferreras, senior advisor.

  FNRS-UCLouvain, Harvard Law School and the Oxford Institute for Ethics in AI. She chairs the International High-Level Expert Committee on Democracy at Work, set up by the Spanish Government.

Our academic and advocacy work is anchored at the Oxford Institute for Ethics in AI, through its Accelerator Fellowship Programme.

We want feedback from the people closest to the problem: spec owners and safety researchers inside frontier AI companies, evals researchers, and the policy staff who would use a reporting standard. Interest and scepticism are equally useful, and anything you share is treated in confidence.

We are also looking for introductions to funders working on power concentration or AI governance capacity, and to senior advisors from AI safety, evaluations and policy. One-pager and full memo available on request.

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.