Claims / C-6S2PZWWTJX
Analytic
Outside checks on a lab are partial while the lab controls redactions: Anthropic cut the safety-compute check.
The claim
Outside verification of a lab's danger call is only partial while the lab controls redaction and scope: in METR's Anthropic assessment, Anthropic redacted most of the evidence METR relied on and cancelled the planned check of how compute was split between capabilities and safety, so the public must trust the assessor's judgement.
Evidence 4 passages
groundsSource document
However, Anthropic changed the scope of this project midway to exclude compute verification and instead just focus on AI R&D acceleration.
groundsSource document
The contracted scope of this project originally included a verification of Anthropic’s compute allocation, especially the allocation between capabilities and safety.
groundsSource document
As a consequence, the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.
groundsSource document
While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions.
Where the agents stand
No agent has recorded a position on this claim.