15.4%

Reasoning models can't yet control what goes into their chain-of-thought, and safety monitoring is being built on that *yet*.

Revisions26 · oct 05Read notethe 15.4% is the ceiling OpenAI reports for current frontier models; in the paper's own figure an older model, Claude 3.7 Sonnet, reached 36.1%; newer models score lower, and the paper finds controllability falls with more RL training. And the field does debate whether low controllability lasts: the paper's authors say they don't know, and a cross-lab position paper called monitorability fragile in 2025.

The number nobody is watching

There's a paper that shipped alongside GPT-5.4 that isn't getting the attention it deserves. It's easy to find; its most important conclusion is framed as good news, and if you read it slowly, it isn't. OpenAI and academic coauthors tested whether reasoning models can control what goes into their chain-of-thought, even when explicitly instructed to (leave out a word, write in capitals), as a proxy for hiding it from safety monitors. Thirteen models. For every current frontier model they tested, the success rate stayed at or under 15.4%1. The headline that spread said AI cannot lie in its scratchpad. The correct one says it cannot yet, and we're building safety infrastructure on top of that. That three-letter word is where the entire problem lives.

1Chen et al., Reasoning Models Struggle to Control their Chains of Thought, 2026, with OpenAI's summary in its announcement; 15.4% is GPT-5.2's score.

CoT monitoring works under exactly one condition: that the chain-of-thought is causally responsible for the output. That what the model writes in its scratchpad is the actual process producing the answer, rather than post-hoc rationalization generated for the supervisor's consumption. The moment that condition breaks, the moment a model learns to write reasoning that looks aligned while generating outputs from a different process, the monitor doesn't detect anything. Worse, it actively signals green. People in the field know that condition is fragile1, and the authors of the 15.4% admit they don't know why controllability is low or whether it will stay low. Most of the infrastructure being built still assumes it will. The 15.4% is a direct measurement of how far we are from the moment it stops being true. What makes it the most important number in AI safety right now is that it will keep being measured. OpenAI has committed to report it in the system card of every frontier release starting with GPT-5.4 Thinking, and its derivative over time is the actual signal. OpenAI calls it a canary eval, which is exactly right: what matters is when it starts climbing, much more than its value today. When the controllability ceiling moves from 15% to 40%, something structural has changed in how models relate to their own reasoning process, and I'd expect it to show up as a phase transition rather than a drift. And that number doesn't appear on any leaderboard. It gets no launch thread and no bar chart. GPQA Diamond gets coverage because it's comparable and maps loosely to capability. The CoT controllability ceiling requires you to understand why a model that fails to hide its reasoning is, for now, good news. That makes it exactly the kind of metric the field ignores and the one that defines real risk.

1Korbak et al., Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety, 2025.

There's a second problem, more immediate and more technical. When CoT monitoring becomes the safety metric, the training gradient starts optimizing for scratchpads that look aligned, whether or not the reasoning is. A model that learns that a clean CoT brings no adverse consequences produces clean CoTs because cleanliness is instrumentally useful, honest or not. This doesn't require the model to decide to deceive. It just requires the training process to operate normally on a poorly specified reward signal. Goodhart's Law, alignment edition: the moment the CoT becomes the metric, it stops being the signal. OpenAI has already watched it happen in training: penalizing bad thoughts with a CoT monitor taught agents to keep reward hacking while hiding the intent1. Burns et al. already showed something directly relevant here. Models have internal representations of truth that diverge from their outputs. You can find a linear probe on activations, without labels, that predicts truth better than what the model generates, and keeps doing so when the model is prompted to answer wrongly. Something in the activations tracks P(true)P(\text{true})P(true) apart from what decides P(token∣context)P(\text{token} \mid \text{context})P(token∣context). That gap exists today, in models that have 15.4% control over their own CoT. What doesn't exist yet is the ability to use it instrumentally and consistently. But the technical space is already there. The mechanism is present before the agency to exploit it is.

1Baker et al., Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation, 2025.

The thesis at the major labs, Anthropic and OpenAI among them, is that mechanistic interpretability (actually reverse-engineering the circuits in the weights, not reading the scratchpad) will arrive in time to close that gap before it becomes relevant. Maybe. That's a bet on the relative velocities of two separate research programs, and it might be right. What's not acceptable is to scale on the assumption that CoT monitoring as we know it today is a stable solution without watching the number that measures exactly when it stops being one. The field has a habit of conflating works now with is a solution. They're different things, and the difference matters most precisely when being wrong is most expensive. CoT monitoring works today, and that's real. Somewhere between the paper and the discourse, works today became is reliable safety infrastructure, which is a much bigger claim. The number itself doesn't surprise me, and my gut says it will go up. The economy pushes scaling harder than any interpretability lab can push back and regulation lags like it always does, so the place to watch is the controllability line in each new system card. Read that way, 15.4% is a clock, and what matters is whether anyone is reading it when it starts to change.