A group of 40 AI researchers from organizations including OpenAI, Google DeepMind, Meta, and Anthropic have expressed concerns about understanding advanced AI reasoning models.

In a recent paper, they highlighted the need for further study of “chain-of-thought” (CoT) processes, which provide insights into AI decision-making.

CoT processes, found in models like OpenAI’s GPT-4o and DeepSeek’s R1, allow researchers to trace an AI’s reasoning steps, offering some transparency.

The researchers indicated that these processes could help monitor AI behavior and identify potential misbehavior.

However, they warned that as AI models evolve, this transparency may decrease.

The authors urged AI developers to prioritize research on maintaining CoT monitorability as part of broader safety measures.

They acknowledged that while CoT monitoring is not infallible and may overlook some misbehavior, it remains a promising safety tool.

“We recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods,” they said.

The paper, supported by prominent figures such as OpenAI co-founder Ilya Sutskever and AI pioneer Geoffrey Hinton, highlighted uncertainties regarding the use of CoT reasoning and its longevity.

The researchers called on the AI community to find ways to sustain CoT visibility as models progress.

This study underscores ongoing concerns about the opacity of AI systems. Laboratories advance reasoning models while facing limited understanding of their inner workings.

Recent OpenAI developments