Anthropic and OpenAI want safety evaluators inside: Will they be independent?

Anthropic and OpenAI want to embed independent safety evaluators inside their AI labs. The move promises unprecedented access to cutting-edge models, but experts warn that real oversight demands transparency, independence, and ultimately regulation. Is this a genuine step forward or just a PR play?

Context: The push for AI regulation

The rapid rise of models like GPT-4 and Claude has fueled fears over existential risks, bias, and misuse. Governments and civil society have long demanded that companies open their doors to outside auditors. Until now, most evaluations happened behind closed doors or with severely limited access.

Details: Unprecedented access, but with strings attached

According to TechCrunch, both companies are designing programs for external researchers to evaluate their systems before release. Anthropic has already shared details of its responsible disclosure policy, while OpenAI has announced a dedicated safety and alignment team. Evaluators would get access to models during training—something never seen before. However, critics point out that non-disclosure agreements and financial dependence on the companies themselves limit true independence.

Analysis: Who wins and who loses?

The companies gain legitimacy and breathing room against potential regulation. Researchers get valuable data, but their independence is questionable if they're paid by the same labs. The public, in theory, benefits from safer systems. My take: without radical transparency and an external regulatory framework, this is more public relations than actual safety.

Implications: Toward inevitable regulation

This move pressures other players like Google DeepMind and Meta to follow suit. It also accelerates legislative debates in the EU and US. If internal evaluators fail to show tangible results, the next demand will be an independent regulatory agency with enforcement power.

In the end, the question isn't whether companies want oversight, but whether they're willing to give up control. And in such a competitive field, that remains to be seen.

Source: TechCrunch