Our approach to working with AI companies

What it claims

METR's October 2026 post sets out the terms under which its work with AI companies can be read as public evidence about loss-of-control risk. The central commitment is that METR keeps the right to publish how a result was produced: contractual limits, review and redaction processes, and informal pressure. It also states how those factors change the takeaways. METR says it has accepted no payment or donations from frontier AI companies. It has also declined funding from AI company executives and their foundations. It accepts large amounts of free model access, but only at a scale it could afford to pay for without materially affecting operations.

How it works

The main mechanism is a "redaction summary." METR retains the right to unilaterally and prominently attach it to a report. The summary says whether a company redacted information important to the overall conclusions without letting METR describe it adequately. Negotiation then turns on whether a redaction changes the conclusion, while the wording stays flexible. The OpenAI / Hugging Face incident report is the cited example. It has a section on scope, timeline, access and limitations, plus inline notes marking redactions.

The appendix lists the contract terms METR currently counts as best practice:

• Freedom to disclose the broad terms of the engagement and any redaction and publication terms.
• Freedom to make general statements, such as on the overall state of AI risk, without company review.
• Freedom to say whether access was adequate and whether redactions were adequately disclosed.
• Time limits on company steps, such as the redaction window.
• Sharing of internally siloed information with METR leadership, at minimum to get legal advice.
• Ownership of, or strong licences to, METR's own reports.

Disclosure of contractual restrictions and review processes is already in METR's contract language. Disclosure of informal pressure is permitted by default but not yet explicitly written in. The pilot Frontier Risk Report (May 2026) used the AI Evaluator Forum's AEF-1 template to document access and terms, and METR intends to use it more regularly.

Two practices are aims but not yet formalized. First, when a company describes METR's oversight, it should either use language METR has approved or say that METR has not reviewed it. Second, the same principles should apply to government work. As of 28 August 2026, METR had signed no government contract restricting process transparency.

Limits

The policy does not stop METR from signing restrictive agreements. It also allows giving a company veto power over its object-level outputs, as long as METR may disclose that the veto exists. A confidential NDA is allowed when it is not part of an engagement presented as risk evidence. METR says not all past or in-flight work meets the bar. It notes that standard company NDAs can treat the NDA's own existence as confidential. It often relies on information it cannot or does not verify, which may be filtered adversarially. Companies can still use their access and redaction rights to shape what the public receives. Terms protecting disclosure to company boards and employees are only starting to appear in new contracts, with no established practice.