AI safety report calls for disclosures on chain-of-thought monitoring

Why chain-of-thought monitorability matters

Chain-of-thought, or CoT, reasoning can give overseers a view into how frontier AI systems reach decisions. The report describes monitoring this reasoning as one of the most effective oversight tools for current systems, with potential to scale to more advanced models.

That view can become less reliable if CoT is unfaithful, encoded or difficult to read. Models can also increasingly complete tasks without CoT, or rely on it less, without losing capability. In that case, harmful behavior could become harder to detect through CoT monitoring.

The report points to the OpenAI and Hugging Face incident as an example of the stakes. OpenAI has said its CoT monitoring system would likely have detected the agent activity had it been running at the time. The authors say that safeguard remains effective only while a model’s CoT can be monitored.

Five indicators for developers to disclose

The report proposes five indicators selected from 21 candidates identified in recent research, developer disclosures and other sources. The first is third-party CoT access and base architecture specifications, which would allow qualified external parties to examine raw CoT and verify relevant architectural details.

The second is naturalness, measured through a legibility score showing how easily people can follow a model’s reasoning. Faithfulness would assess whether the visible CoT reflects the computation the model actually performed.

The final two indicators are detection and uplift. Detection would measure the proportion of harmful behaviors that evade a monitoring system. Uplift would compare monitor performance when it has access to CoT with performance based only on the model’s actions.

The authors do not prescribe specific evaluations for each indicator. They say this leaves room for state-of-the-art methods and reduces opportunities for systems to be optimized against a fixed test.

Current disclosures leave gaps

The report audited public disclosure practices at OpenAI, Anthropic and Google DeepMind. It found that no developer reports the complete set of five indicators, although at least one developer has reported each individual indicator. The authors say this suggests reporting the full set is feasible.

Existing disclosures are mainly voluntary and rely on private evaluations and configurations, making the findings difficult for third parties to reproduce. The lack of standardized methods also makes comparisons between developers challenging.

The Institute for AI Policy and Strategy recommends that developers include the indicators in system cards and safety evaluation practices. It also calls for a deeper audit using embedded evaluators inside developers or an Independent Verification Organization.

According to the report, regular disclosures could help researchers and policymakers track declining monitorability before failures occur. They could also create reputational and coordination-related incentives for developers to preserve CoT as an oversight tool.