METR explains how it works with AI companies on risk assessments

In short

METR, a research nonprofit, has published a post explaining how it works with AI companies on projects the public might use to understand risk. METR says it will publicly describe limits on its access, redactions by companies, and pressures it faced. It also says that relying on companies' cooperation limits how independent it can be.

What happened

METR published the post on October 14, 2026. It says its aim is to gather and publish evidence on whether humans could lose control of AI agents on a large scale. It works with AI companies on evaluations, assessments and investigations.

METR says people increasingly expect these voluntary projects to give "safety assurance." The post is meant to explain its view of that role. METR says it collects evidence for the public and does not intend to decide how companies build or release their AI agents.

For work that bears directly on public understanding of risk, METR says it takes these steps:

• It will describe the scope and length of its work.
• It will describe any limits on access, and whether they affected its conclusions.
• It will say whether a company redacted (removed) information that mattered to its conclusions.
• It will describe formal restrictions and informal pressures that may have affected its conclusions.
• It will disclose staff conflicts of interest.

METR says it has not accepted payment or donations from frontier AI companies. It says it has also turned down funding where a company would have influence over it. It accepts a large amount of free access to AI models, but says it can afford to pay for the free usage without materially affecting its operations.

METR also describes an arrangement it has found useful. It keeps the right to attach a "redaction summary" to its reports. This says whether a company removed information that was important to the overall conclusions.

The post says this approach came from several years of work. It points to a pilot Frontier Risk Report from May, which grew out of work with Anthropic, Google, Meta and OpenAI.

What was done about it

This was a policy explanation, not a response to an incident. METR says it is still working out how to apply the policy. It says not all past or ongoing work meets its standard. It also says it expects to update the post over time.

METR says it aims to use a standard disclosure template, called AEF-1, more regularly in future reports.

Why it matters

METR says readers need to know the processes, limits and incentives behind its findings to interpret them correctly. It admits that its independence is limited. It says it depends on companies to take part, give access and allow publication. It also says it often relies on information it cannot or does not verify, which might be filtered in a way that misleads. Even with transparency, it says, a company could use its right to withhold access or redact to shape what the public sees.

Read the full post at the original report.