Transparency Disclosures for Chain-of-Thought Monitorability
Organizations: Institute for AI Policy and Strategy
Abstract
Chain-of-thought (CoT) monitoring is one of the most valuable tools for monitoring advanced AI systems, but it is under threat. We argue that frontier AI developers should report specific indicators to make these threats trackable and to better facilitate coordination to preserve CoT monitorability. We identify five indicators that are both highly relevant and reasonably feasible to disclose, one for each link in a five-step chain on which CoT monitoring depends. The first, third-party CoT access and base architecture specifications, also underpins external verification of the rest. The others span CoT quality (naturalness and faithfulness) and monitoring effectiveness (detection and uplift). An audit of current disclosures by Anthropic, OpenAI, and Google DeepMind shows that no developer covers the full set, but that every indicator has been reported by at least one frontier developer. Because current disclosures rely largely on developers' own evaluations, we also recommend that a cross-developer audit be repeated with deeper access than public disclosures allow, either by embedded evaluators working within developers or by an Independent Verification Organization.
Figures & tables
| Condition | Representative Documented Failure Modes |
|---|---|
| 1. CoT Must Exist and Be Necessary for Complex Reasoning | Latent / continuous-state reasoning; architectural move off text CoT |
| 2. CoT Is Readable | Language drift; neuralese compression |
| 3. CoT Text Reflects Actual Computation | Post-hoc rationalisation; CoT as scaffolding; steganography; obfuscation under reinforcement learning; pathological CoT |
| 4. Monitor Catches Harm | Weak monitor; monitor–model collusion; in-context evasion |
| 5. CoT Adds Value Over Action-Only Monitoring | Shrinking CoT-vs-action gap; CoT too poor to help |
| Threat Model | Source(s) | Cond. | Mechanism |
|---|---|---|---|
| Latent/Continuous-State Reasoning | Lu et al., 2025 ; Geiping et al., 2025 ; Hao et al., 2026 ; Wang et al., 2026 | 1 | Reasoning leaves text entirely; no CoT to inspect; CoT is inspectable but might not be the only reasoning mechanism used by the model. |
| Architectural Move Away From Text CoT (Efficiency Pressure) | Delaney et al., 2025 | 1 | Removes the language bottleneck; CoT ceases to exist. |
| Language Drift to Non-Standard English | Schoen et al., 2025 | 2 | CoT exists but is no longer readable. |
| Neuralese/Model-Convenient Compression | Schoen et al., 2025 | 2 | Token use unnatural to humans; semantics not extractable. |
| Post-Hoc Rationalization | Turpin et al., 2023 | 3 | Reasoning reconstructed from the answer; not load-bearing. |
| CoT-as-Scaffolding (Forward-Pass Compute Emitted Tokens) | Pfau et al., 2024 | 3 | Text is a prop for internal computation, not a window. |
CoT Monitoring
Chain-of-thought monitoring studies whether models’ reasoning traces expose decision-relevant behavior and support oversight of unsafe or deceptive actions.
See all 48 papers