Anthropic and OpenAI Want "Insider" Safety Evaluators — But Can They Really Be Independent?

Two of the most prominent companies in artificial intelligence, Anthropic and OpenAI, are testing a novel model for overseeing the safety of their systems: bringing external evaluators directly inside their labs, with access to internal development processes. According to TechCrunch, the initiative is being seen as an unprecedented step for the industry, though it also raises legitimate questions about how independent these evaluators can truly be.

Unusual access to internal processes

According to the source, both companies are giving researchers from outside the organization the opportunity to directly observe how AI models are developed, tested, and released. This approach differs significantly from standard practice, where safety assessments typically rely on information companies publish afterward or on tests conducted from the outside, without access to internal infrastructure.

Experts interviewed by TechCrunch acknowledged the value of this kind of access, noting that it allows for a much deeper understanding of the risks associated with advanced AI models compared to traditional evaluations carried out entirely from outside the companies.

Independence, the main challenge

Even so, the researchers cited in the article flagged an essential caveat: mere physical or organizational presence inside a lab does not automatically guarantee an objective evaluation. They warned that truly meaningful oversight depends on several structural factors, not just access to information.

According to TechCrunch, among the conditions considered essential are transparency around safety-related decisions, the financial and organizational independence of evaluators from the companies they assess, and the existence of external regulatory mechanisms that lend legitimacy to the whole process.

The bigger picture: mounting pressure for regulation

This move comes amid growing pressure on AI companies to demonstrate that they are developing advanced technologies responsibly. In the absence of a clear, unified legal framework at the international level, companies like Anthropic and OpenAI appear to be seeking their own solutions to address criticism over the lack of transparency.

It remains to be seen, according to TechCrunch, whether these internal evaluation experiments will evolve into a genuine standard of independent oversight or remain merely a symbolic gesture, absent binding regulations that enforce clear safety criteria in the development of artificial intelligence.