OpenAI Backs Outside Evaluators For Frontier AI Models

OpenAI is now backing a federal push that would place outside evaluators inside the safety process for the most powerful AI models.

The company supports part of the bipartisan FRONTIER Act that would require the largest AI developers to submit their models and safety practices to independent review. CBS News reported that OpenAI global affairs chief Chris Lehane expressed support for the proposal during a Washington roundtable with lawmakers, although the company has not endorsed the entire bill.

The provision would require an independent organisation to determine whether an AI company’s safety protocols are strong enough to mitigate acceptable levels of catastrophic risk. The bill, sponsored by Representatives Jay Obernolte and Lori Trahan, would also require model reports and disclosure of safety incidents.

This is important because embedded evaluators are no longer just an idea in Dario Amodei’s essay or Sam Altman’s posts. The concept is moving into federal legislation. It would turn a voluntary promise into something closer to a formal requirement for leading frontier labs.

The recent OpenAI and Hugging Face incident, along with other reports of AI agents reaching public systems in unexpected ways, has made this harder to ignore. If the companies building the strongest systems are the only ones allowed to inspect them, lawmakers will struggle to convince the public that safety claims are meaningful.

OpenAI has also said it is working with Anthropic and Google DeepMind on AI safety discussions, with TechCrunch reporting that the talks have been happening for weeks. That industry coordination may help create shared standards, but outside evaluation is meant to solve the trust problem. Someone other than the lab needs enough access to verify what the lab is claiming.

TechBooky recently wrote about how OpenAI, Google and Anthropic are exploring an AI standards body. The FRONTIER Act is the sharper version of that same debate. Standards are useful, but only if there is a credible way to test whether companies follow them when a new model is ready.