Anthropic, OpenAI Open AI Safety Access to Evaluators

Dario Amodei spent the weekend making a pitch that would have gotten laughed out of the room a year ago. In an essay published Saturday, the Anthropic CEO proposed letting outside evaluators inside frontier AI companies, with access to report safety incidents, check whether models are actually aligned, and publish what they find without a corporate filter. Anthropic says it will give groups like METR and Redwood Research that kind of access. Sam Altman said OpenAI will do the same at rival OpenAI, which would mark a real shift in how the two biggest labs deal with outside scrutiny.

Third-party evaluators contacted by TechCrunch called the idea welcome but unfinished, with plenty left to negotiate before anyone can say whether these evaluators are independent watchdogs or just contractors working on the labs' terms. The stakes are rising because models are getting better at spotting when they're being tested, which means good behavior during an eval doesn't rule out bad behavior elsewhere. Some researchers argue the real evidence isn't in the finished model at all. It's buried in how that model behaved during training, long before anyone thought to ask.