Executive summary

Frontier AI systems currently externalise their reasoning in human-readable form. Before producing an answer to a complex problem, a model writes out its intermediate steps in natural language. These steps produced are known as a chain-of-thought (CoT). Chain-of-thought reasoning improves AI’s capabilities on complex tasks but is also a safety asset. Several documented misbehaviours in deployed models – plans to circumvent guardrails, deliberate underperformance during evaluations, reasoning toward prohibited outputs – have been caught because the model's reasoning trace was readable in its CoT. Chain-of-thought reasoning not only enables catching such misbehaviours after the fact, but also in real-time, through the use of separate classifiers that can immediately flag suspicious output.

Industry support for CoT monitoring is strong. Researchers from OpenAI, Anthropic, Google DeepMind, and Meta have joined the UK AI Security Institute in endorsing CoT monitoring as an important safety opportunity. In a survey of nineteen AI safety researchers conducted for this Arq report – more than half drawn from frontier AI companies – nearly 80% rated CoT monitoring as very important or absolutely central to the safety of advanced AI.

But human-readable reasoning may not be a permanent feature of frontier models. Chains of thought are currently legible only because today's models are trained on human text and because making them reason step by step in natural language happens to improve performance on hard tasks. In a line of research now maturing inside the same frontier AI companies, this natural language is replaced with internal numerical representations that aren’t directly interpretable by humans. Early results suggest that these ‘latent reasoning’ architectures can be several times cheaper to run, and the cost gap may well widen. Once it does, consideration about safety and competitiveness could come into tension: in our survey, only one respondent in nineteen expected AI companies to voluntarily preserve readable reasoning under large economic pressure. While it is theoretically possible to build interpretability tools that would be needed to oversee new, opaque numerical reasoning in real time, they are not yet ready, and respondents do not expect them to be until 2030 at the earliest.

Left unaddressed, a shift from readable to opaque reasoning would remove the most scalable safety oversight tool the field currently has, at exactly the capability level where oversight becomes crucial to prevent catastrophic misuse, and before any replacement method is ready for deployment. To avoid a race to the bottom, and to ensure AI developers retain oversight over frontier AI models in the future, this report sets out two recommendations.

First, we advise the EU AI Office to issue interpretive guidance clarifying what the AI Act’s General-Purpose AI Code of Practice requires when a provider moves away from human-readable reasoning. The Code already obliges providers of advanced AI models that pose systemic risks to use state-of-the-art evaluation and monitoring tools – which today requires readable CoT – and to demonstrate that any move to a more compute-efficient approach delivers equal or superior safety outcomes. The implication is that a provider switching to an architecture that cannot be monitored through CoT must show, by other means, that oversight has not slipped. We call this principle ‘non-regression’. The AI Office should make the non-regression principle as applied to CoT monitoring explicit, rather than implicit, and show how equivalent non-CoT-based oversight can be demonstrated through a so-called safety case. Formal codification can follow in the Code's next biennial revision.

Second, the EU should fund the technical work that such safety cases will depend on. We propose an independent monitorability research fund, inspired by the UK’s AI Alignment Project. Such a fund could be co-financed by philanthropic organisations, frontier AI companies, and EU public funding through instruments such as Horizon Europe and the RAISE programme. Pull mechanisms – milestone prizes, challenge competitions, bounties for reproducible failures of proposed monitoring techniques – can broaden the research community working on AI oversight well beyond the frontier AI companies themselves.

Why chain-of-thought monitoring matters

Frontier models are well-known to show misaligned behaviour from time to time. For example, when asked to fix a failing piece of software, a model will sometimes alter the test itself rather than fixing the code just so that the broken code passes, and write out the manoeuvre step by step in its reasoning trace. Today, such misbehaviour can be caught because the trace is visible in the CoT; while the final code would look like a clean fix when viewed in isolation. In similar cases, models planned to circumvent safety measures, deliberately underperformed on evaluations, and reasoned toward prohibited outputs, and these behaviours were only discovered because the model wrote about them in its CoT.

Three properties make CoT monitoring unusually valuable. First, it works in real time, flagging problematic reasoning before the model acts. Second, the traces it produces can be logged, searched, and reviewed long after the fact, thereby forming a foundation for audit or incident investigation. Third, it scales: existing language-processing systems can triage these traces at deployment volume. No other current interpretability method combines all three advantages at frontier-model scale.

Latent reasoning architectures cannot currently offer the same advantages; after all, there is nothing to read but numbers that come without straightforward translations into natural language. But architectures that abandon CoT may be appealing for other reasons: early results suggest that they can be several times cheaper to run than language-based equivalents, and the gap could widen. A company whose models keep their reasoning visible while its competitors’ do not may soon pay a real price for safety and transparency.

AI researchers working directly on frontier AI models expect the transition to non-CoT architectures to happen soon, and to outpace the development of replacement oversight tools. In our survey, 63% of respondents expected frontier AI systems to use primarily internal, non-readable reasoning by 2030, with a quarter expecting the shift within two years. The same respondents mostly estimated that interpretability methods capable of equivalent oversight would be ready in 2030–2035, though a substantial group either predicted that such methods would be developed later, or believed that they were unlikely to be developed at all. Asked whether frontier laboratories would voluntarily preserve readable reasoning in high-risk applications if opaque architectures became dominant, just one of nineteen respondents said ‘yes’.

The proposal: equivalent monitorability

This report aims to help AI companies break the coordination problem they currently face, in a future-proof way. A blanket prohibition on opaque architectures would risk blocking beneficial progress and could become hard to enforce. We propose a mechanism that permits architectural innovation while ensuring that safety oversight does not degrade as a side effect.

Under our proposal, a developer may adopt any reasoning architecture it chooses, provided it can demonstrate that the classes of misbehaviour previously detectable through CoT monitoring remain detectable, with comparable reliability, latency, and auditability, under whatever oversight methods the new architecture supports. We call this equivalent monitorability: a non-regression standard that states that oversight capability cannot slip backwards as a result of deploying a model without CoT.

Demonstration of equivalent monitorability could take the form of a safety case: a structured argument, supported by evidence, that the system is acceptably safe for its intended deployment. Safety cases are a mature regulatory instrument in aviation, nuclear power, and medical devices, and bodies including the UK AI Security Institute are now developing them for advanced AI. Unlike fixed benchmarks, they are hard to game, can be tailored to specific models and deployments, and can evolve as both architectures and interpretability tools mature. The obligation we envisage would only apply when a developer transitions to an architecture that could reduce oversight; providers continuing to deploy models with CoT reasoning would face no new requirements.

Outlook

The opportunity to oversee advanced AI by reading its reasoning exists because of an accidental alignment between what makes models capable and what makes them interpretable. That alignment is not guaranteed to persist, and many of the researchers closest to the technology expect it to erode within the decade unless coordinated action is taken. The framework set out here is light-touch for cases where transparency is preserved, and demanding for those where it is given up. It does not mandate any particular architecture, prohibit innovation, or lock in chain-of-thought reasoning as a permanent requirement. It asks only that any move toward less transparent systems be accompanied by evidence that oversight has not been lost, and that European research investment ensures that it is possible to provide such evidence.

The rest of this report is structured as follows:

Chapter 1: The technical case — Why CoT is one of the most promising and scalable safety tools today, what makes it work, how faithful it actually is, and how architectural trends such as latent reasoning threaten to eliminate it.

Chapter 2: The policy case — Why preserving CoT monitoring is a coordination problem no single lab can solve on its own, why alternative safety approaches are not yet ready to substitute, and how a governance framework can help.

Chapter 3: Safety cases for equivalent monitorability — How safety cases could play a central role in such a framework, what they could look like in practice, and what kinds of the research investments are necessary to make these safety cases possible.

Chapter 4: Implementation — How the EU AI Act's Code of Practice already carries most of the legal weight for the proposed implementation, what role the EU AI Office and the International AISI Network could play, and how to fund the necessary research.

Appendix A: Expert survey — Results from a survey of AI safety researchers conducted for this study on the importance of chain-of-thought monitorability and relevant interventions.

Appendix B: Other threats to CoT monitorability — Further background on other threats to CoT monitorability not covered in the main report.

Click here to download the full report.