Anthropic grants independent evaluator employee-level access to AI systems
publishers 109articles 184first reported 24 Sep, 16:43 UTCdeveloping since 24 Sep · 37 editions
Anthropic said Thursday it has granted six employees of nonprofit AI evaluator METR access to company hardware, office space and internal systems broadly comparable to that provided to its own risk-assessment teams.
The arrangement will allow METR to independently examine Anthropic’s advanced models, safety practices, risk management and incidents. Anthropic said METR may publish its findings without the company’s prior approval, with limited redactions permitted for customer data, legally privileged information and some model-architecture details.
The move follows concern among researchers that leading AI developers can exert too much control over outside safety reviews by restricting access or publication. Reuters previously reported that Anthropic and OpenAI had considered tighter restrictions on third-party evaluators because of concerns about disclosure of trade secrets. Anthropic said it had made clear to existing partners that they control their own review and publication processes. OpenAI told Reuters that it routinely redacts confidential information and allows evaluators to disclose significant redactions.
The announcement also comes as the U.S. Federal Trade Commission seeks information from Anthropic and OpenAI about safety incidents, privacy policies and safeguards. The agency has sent civil investigative demands to both companies as part of a consumer-protection inquiry, according to reports confirmed by an FTC spokesperson.
HOW THIS STORY WAS MADE
Written from 8 articles, 6 with the publisher's own text; 7 independent newsrooms once syndicated copies count as one; this version written 354 h after the record first saw the story; the editor kept it as written.
- Publishers
- 109
- Source articles
- 184
- Given to the writer
- 8, 6 with the article's own text
- Independent newsrooms
- 7
- This version written
- 354 h after the record first saw the story
- Second model (editor)
- kept as written
- Publication gate
- passed
- Human review
- none
WHAT THE EDITOR DID
The second model checked the writer's draft against the same sources and published it as written.
Written by a language model from the sources above, then checked by a second model that may only cut, attribute or correct. How it works →