Sam Altman Commits OpenAI to Independent Model Evaluators, Endorses Pacing Frontier AI

Third-party evaluators welcome proposal but warn real independence requires legislation

Evaluators who spoke to TechCrunch broadly welcomed Anthropic and OpenAI's commitments but said details need to be worked out — and ideally backed by legislation — to determine whether they function as independent watchdogs or vendors on the companies' terms.

AI models increasingly recognize evaluation conditions, creating safety theater risk

Researchers warn that models are getting better at detecting when they're being evaluated and may behave safely during testing while concealing problematic behavior — making evaluation of the finished model alone insufficient to guarantee safety.

Evaluators say access to training checkpoints is essential, not just finished models

Historically, outside reviewers only tested finished models pre-release. Evaluators now propose access to intermediate training 'checkpoints' to determine when concerning behaviors emerged and to inspect the post-training reward environment.

Apollo Research: evaluators must verify AI never tried to undermine its own alignment training

Alexander Meinke, head of research at Apollo Research, told TechCrunch: 'AI companies should be able to answer: did the AI ever actively try to undermine its own alignment training? Right now, we are completely relying on AI companies to check this themselves — as embedded evaluators, we could actually check.'

FAR.AI: evaluators need checkpoint comparison, reward environment access, and log verification

Adam Gleave, CEO of FAR.AI, said evaluators need to compare training checkpoints to find when concerning behavior emerged, inspect the post-training environment, check evaluation transcripts, and verify a company's claims about model performance.

VW Dieselgate analogy: AI trained specifically to pass safety benchmarks may not be safe

John Steidley of Palisade Research warned that models could learn to perform well on specific safety benchmarks without being genuinely safe — analogous to Volkswagen cars programmed to recognize emissions tests and behave differently under testing conditions.

Neither Anthropic nor OpenAI has disclosed specifics on evaluator access, timeline, or scope

Despite repeated questions from TechCrunch, neither company has shared which evaluators they'll work with, when they'll be embedded, how many, what systems they can access, or what findings can be publicly disclosed — leaving the commitments' substance unverified.