Hassabis backs independent oversight inside frontier AI labs
Google chief scientist Demis Hassabis backed Dario Amodei's proposal to place independent evaluators with standing, employee-level access inside frontier laboratories. The Decoder and EL PAÍS report that Hassabis called the direction right while saying details need work; Google researcher Peyman Milanfar disputed the case for an imposed speed limit.
Artificial Intelligence··Midday
Hassabis backs the direction and leaves details open
Demis Hassabis, co-founder of Google DeepMind and Google's chief scientist, wrote that Dario Amodei's proposal for standing outside evaluators inside frontier AI laboratories points in the right direction. The Decoder and EL PAÍS also report his warning that the details still need work. His support does not establish agreement on every implementation choice. The shared point is that laboratories' own safety claims would face continuous external examination.[1], [2]
The proposal moves from periodic tests to standing access
Amodei's first step would give outside evaluators such as METR standing, employee-level access. The arrangement aims to examine active development and safety practice, rather than testing only a model prepared for release. Hassabis connected the approach to Google's proposal for an industry-wide standards body for frontier AI. The reports do not specify the legal scope of access, protection of confidential information or evaluators' enforcement powers. Those details would determine how much oversight the arrangement actually provides.[1], [2]
The case for a speed limit remains contested
Google researcher Peyman Milanfar argued that systems capable of reliable self-improvement would be slow, damped and bounded by stability, making an imposed speed limit unnecessary. His objection separates the usefulness of independent evaluation from whether model progress should be slowed. Hassabis supports the oversight direction, while Milanfar questions a technical premise behind the danger case. The reports therefore show no complete industry consensus; access, measurement and pacing remain distinct points of disagreement within the same proposal.[1], [2]
