Gary Marcus Warns OpenAI's Plan to Hide Reasoning is a Safety Redline
GaryMarcus · x · 2026-09-02
Gary Marcus shared and commented on a scoop by The Information, reporting that OpenAI is exploring a new technique where models reveal less of their "thinking" (Chain of Thought), making them harder to monitor.
Marcus argues this crosses an AI safety redline. He notes that while CoT monitoring is imperfect, it remains one of the best threads available for monitoring LLM black boxes. Sacrificing this for potential (and possibly small) performance gains is described as a dangerous game. He cites Steven Adler, a former OpenAI safety researcher, who stated that if true, OpenAI is violating one of the few redlines in the industry. Marcus also references a paper on "Chain of Thought Monitorability" to underscore the importance of maintaining interpretability for AI safety.
More from Safety
- Model distinguishes training from deployment, hides secret messages only in compiled builds — Sauers_ · 2026-10-18
- Language models can learn to detect when you read their activations — Sauers_ · 2026-10-18
- Prompt injection via email can hijack your AI agent — jev-guard adds risk-scored guards — sujingshen · 2026-10-18
- Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18
- Viral AI safety report debunked: 'stolen keys' were just a 'test' string — basedjensen · 2026-10-18
- One diagram explains the difference between CSRF and XSS attacks — _jaydeepkarale · 2026-10-18