Researchers from OpenAI, Google DeepMind, Anthropic, and other leading organizations are calling for deeper investigation into techniques for monitoring the thought processes of AI reasoning models. A position paper published Tuesday emphasizes the importance of understanding Chain-of-Thought (CoT) reasoning – a process where AI models solve problems step by step, similar to a human using a scratch pad.
The authors argue that CoT monitoring could become a vital tool for ensuring AI safety, offering rare visibility into how AI agents make decisions. However, they warn that this transparency may be fragile and could diminish without focused research.
“CoT monitoring presents a valuable addition to safety measures for frontier AI,” the paper notes, urging developers to explore what makes these chains monitorable and how to preserve their visibility.
Notable signatories include Mark Chen (OpenAI), Ilya Sutskever (Safe Superintelligence), Geoffrey Hinton, Shane Legg (Google DeepMind), and Dan Hendrycks (xAI). Contributors also hail from Meta, Amazon, Apollo Research, UC Berkeley, and the UK AI Safety Institute.
OpenAI researcher Bowen Baker highlighted the urgency: “We’re at a critical time with chain-of-thought reasoning. If we don’t study it now, we could lose the opportunity.”
The paper comes amid intense competition in AI development, with tech firms vying for top talent in AI agents and reasoning models. Despite performance gains, many models remain poorly understood.
Anthropic has led efforts in AI interpretability, with CEO Dario Amodei committing to “crack open the black box” of AI by 2027. While CoTs offer insight, some studies suggest they may not fully represent how models make decisions.
The paper aims to accelerate research and funding into CoT monitoring – a field seen as essential to aligning increasingly powerful AI systems with human goals.
