Criticizes OpenAI replacing CoT: Contradicts own safety report

peterwildeford · x · 2026-09-02

Peter Wildeford cites OpenAI's call from 10 months ago for the industry to preserve chain-of-thought (CoT) monitorability. He references OpenAI's own report on a 'rogue AI attack,' which underscored the need for expanded CoT monitoring, to question the logic of OpenAI replacing CoT. He argues that we should not render AI reasoning completely inscrutable to human oversight.

Related event: Reports: OpenAI's Astra uses recurrent-depth latent reasoning(49 posts)→

Original post →

More from Safety

Safety channel →