METR sets out rules for how it works with AI companies on risk evidence

Goal: evidence for the public, not decisions for companies

According to METR, its aim is to give the public accurate, timely and independent evidence about AI loss-of-control risk, so that the public, policymakers and developers can make their own informed decisions. Its main method is collecting high-quality, concrete evidence and publishing what it finds.

METR says it may also work with companies or regulators to comment on unpublished sensitive information, review risk reports, or give an independent judgment on claims about risk. It cautions, however, that many relevant claims involve open scientific questions and cannot be straightforwardly verified. The organization says it does not intend to decide how companies should develop or deploy their agents. In its view, decisions about frontier AI are high-stakes judgment calls that should not be made behind closed doors, and the current situation calls for open discussion of both the risk levels and which levels are acceptable.

The measures METR applies

For work that bears directly on public understanding of AI risks, such as pre-deployment testing, risk report reviews and risk assessment exercises, METR lists several commitments.

• Transparency about process: METR will publicly describe the limitations of its work, including the agreed scope and duration, any access, data and resource limits, redactions by the company, and formal or informal pressures that may have materially affected its conclusions.
• Transparency about independence: it will disclose relevant staff conflicts of interest under its COI policy.
• Financial and operational independence: METR says it will ensure it is not financially dependent on the companies it works with and will recuse employees with a significant conflict of interest when choosing a project team.
• Confidentiality and security: nonpublic information and access provided by companies will be secured under its practices and agreements.

METR also states principles meant to reduce how much the public must simply trust it. It aims to describe its evidence and methods in enough detail for outside experts to judge its findings, and to publish evidence in a way that allows others to run their own analyses. It aims to publish frequent, iterative reports even when questions cannot be fully answered, and to take an approach that lets the public compare reports across companies and makes engagement incentive-compatible for each company.

How it has worked in practice

METR says it reached this approach after several years of working with AI companies. It points to its pilot Frontier Risk Report, published in May, which grew out of work with Anthropic, Google, Meta and OpenAI. The post also links to examples from past work, including an OpenAI / Hugging Face incident investigation report, an OpenAI GPT-5.1-Codex-Max report, an Anthropic Claude Opus 5.5 system card summary and an OpenAI o1-preview preliminary report.

One arrangement METR says works well is retaining the right to unilaterally and prominently attach a 'redaction summary' to its reports. The summary indicates whether a company redacted information important to METR's overall conclusions without allowing it to adequately describe what was removed. METR says this lets discussions about redactions focus on whether a redaction would change a report's conclusion, while leaving some flexibility in wording.

On funding, METR says it has so far chosen not to accept payment or donations from frontier AI companies and has declined offers where an AI company would have influence over it, for example funding from a company executive or foundation. It accepts a large amount of free model access, but says it ensures it could afford to pay for the free tokens without materially affecting operations.

Limits of the voluntary regime

METR is candid about the weaknesses. It says ensuring adequate independence has proved challenging under the current voluntary regime. It remains dependent on companies to engage voluntarily, provide access and permit publication. In many cases it relies on information it cannot or does not verify, which might be filtered adversarially. Even with transparency about limitations, METR says, a company could use its rights to withhold access or redact in order to exert control over what evidence the public receives.

For readers of AI risk reports, the practical message is that such findings should be read together with the disclosed scope, access, redactions and constraints behind them. METR says it expects to update the post over time.