Frontier AI models now make decisions with real legal and financial consequences. But evaluating whether they'll actually fail in practice is tricky—labs have access to real user conversations that reveal problems, but external researchers don't. OpenAI's new work on 'Deployment Simulation' asks whether public chat datasets can predict rare model failures, potentially giving independent safety researchers a way to audit models without proprietary access.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton