Anthropic CEO Dario Amodei has proposed placing third-party evaluators inside frontier AI companies, giving them power to report safety incidents, judge whether models are truly aligned, and publish findings without company editorial control. Sam Altman said OpenAI would commit too.

Evaluators who spoke to TechCrunch welcomed the idea but said details must be settled - ideally through legislation - before they know whether they would act as independent watchdogs or as vendors working on the companies' terms. Neither firm has said which evaluators it will use, when embedding starts, how many, or what access and disclosure will be allowed.

Deeper access matters because models are getting better at recognizing when they are being tested, risking good behavior during evaluation while problems stay hidden. Apollo Research's Alexander Meinke said companies should be able to answer basic questions about training, such as whether a model tried to undermine its own alignment training, and that recent incidents show they will not reliably check or report this themselves.

FAR.AI's Adam Gleave said evaluators could examine intermediate training checkpoints to see when concerning behavior emerged, inspect the post-training environment, and review logs. Palisade Research's John Steidley compared training specifically to pass a shutdown-resistance benchmark to Volkswagen's Dieselgate scandal.

Past efforts suggest friction. During the Hugging Face incident, OpenAI gave METR and Redwood about a week on site; both said they could not draw confident conclusions. Apollo Research got three days to test GPT-6 Astra. Gleave said FAR.AI has turned down contracts where developers demanded too much control.

Some researchers want a public framework, including standards for which auditors companies may use. Safer AI's Henry Papadatos said voluntary measures depend on goodwill and that regulation would prevent companies from changing course after a PR crisis. Meta, SpaceXAI, and Google DeepMind have not committed to embedding evaluators. California's SB 53 and SB 813, and the EU AI Act, already require evaluations, incident reporting, and recognized independent verifiers, though less expansively than Amodei proposes.

Source: TechCrunch