On 10 November 2026, researchers Jacob Davies and Theo Bearman submitted a paper proposing disclosures to track whether advanced AI systems’ chain-of-thought reasoning remains monitorable. Chain-of-thought (CoT) monitoring examines a model’s reasoning as a way to monitor its behavior. The authors argue that developers should report indicators that make threats to this monitoring trackable and support coordination to preserve it.[1]
The paper’s abstract identifies five indicators across a five-step chain on which CoT monitoring depends. The first covers third-party access to CoT and base architecture specifications; the authors say it also supports external verification of the remaining indicators. The others concern CoT quality, through naturalness and faithfulness, and monitoring effectiveness, through detection and uplift. The abstract does not describe how each indicator is measured.[1]
The authors report an audit of public disclosures by Anthropic, OpenAI and Google DeepMind. According to the abstract, no developer disclosed the full set, although each indicator had been reported by at least one of them. The authors say current disclosures rely largely on developers’ own evaluations and recommend repeating a cross-developer audit with deeper access, using embedded evaluators or an Independent Verification Organization.[1]
Lead photo: Jacob Davies, Theo Bearman · CC BY 4.0.
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
f4cd81b3329fdfb0225268533d06e2e94fd71fff2ab76b8f60378c295ab4d20c
