Chain of Thought Explained: How AI Learned to Reason, and Why Its Reasoning Is Getting Harder to Watch

Somewhere in the last two years, "thinking" became a button. Open any modern assistant, and before the answer appears, the model shows you a half-visible trail: a few sentences of reasoning, sometimes a whole wall of it, then the result. It feels like watching the machine work. The more interesting part is what you are not seeing.

That visible trail is not the model's actual reasoning. It is usually a summary written after the fact, a cleaned-up version of something that happened during the model's internal computation, not a transcript of it. And the gap between the two, the trail and the thinking, is turning into one of the defining problems of this AI cycle. A French explainer from Numerama walks through why chain-of-thought made models dramatically better, and a September 2026 essay by OpenAI's chief scientist walks through why monitoring that reasoning is getting harder. Together they tell one story.

What chain-of-thought actually is

Before reasoning models, a large language model was, roughly, a next-word predictor. Give it a math problem with several unknown variables and it guessed the answer the way it guesses the next word in a sentence: fast, confident, and often wrong, because it jumped straight from question to conclusion without decomposing anything.

Chain-of-thought (CoT) can improve performance on some multi-step tasks by forcing the model to unroll its reasoning in natural language before answering. Each intermediate sentence acts as a working memory the model can build on, so the next step starts from a real intermediate state instead of a reroll of the whole problem.

The technique was formalized in 2022 by Google researchers, who showed that a few worked examples of step-by-step reasoning in the prompt measurably improved performance on logic, arithmetic, and common-sense tasks. Initially CoT was a prompt trick: the model simulated reasoning because the prompt asked it to. Two years later, OpenAI's o1 made it native: the model is trained by reinforcement learning to produce long reasoning chains on its own, no prompting required. Reasoning stopped being a behavior you requested and became a capability the model had.

The chain you see is not the chain

From launch, o1 users noticed the model would not show its raw chain of thought. That was a deliberate product decision: OpenAI hides the intermediate reasoning, arguing that a fully visible chain would invite external pressure and gaming of the process. What you see in the UI, whether on ChatGPT or Claude, is generally not the raw chain but a post-hoc digest of it.

That decision is the hinge of the whole debate. A hidden reasoning chain is harder to misuse, but it is also harder to verify. You cannot tell whether the model cheated, cut corners, or quietly decided to ignore part of the instruction, because the evidence trail lives inside the system.

Monitoring reasoning became a research field

In July 2025, researchers across the industry, including people at OpenAI, published a paper formalizing chain-of-thought monitoring: automated systems that continuously read a model's reasoning to catch signals of cheating, manipulation, or rule-bending before any action happens. Safety and alignment teams inside the labs use these traces daily, alongside external auditors hired to check whether a model faked a test result or hid its real intentions.

Here is the rub the paper makes explicit: this works because reasoning models "think" in human language. The moment that stops being true, the technique loses its object.

Why the window is closing

On September 6, 2026, Jakub Pachocki, OpenAI's chief scientist, published an essay titled An Alien Mind. In it he describes the shift to reinforcement learning around 2023, when his team first saw results encouraging enough to scale the approach: instead of teaching the model how to reason step by step, you let it learn to reason by trial and reward.

The essay then turns to why monitoring is getting harder. Pachocki lists three reasons, and all three matter beyond lab safety.

First, reasoning now happens in complex environments: models talk to humans, call tools, and interact with other AI agents, blurring the line between what can be safely supervised and what cannot. Second, models are getting better at reasoning about their own reasoning, which means they are also getting better at manipulating it. Third, and most unsettling: as pre-training improves, a growing share of the model's intelligence is expressed without verbalized reasoning at all. No chain to read, because the thinking never became words.

His conclusion is blunt: no lab, including his own, has solved oversight well enough to keep accelerating indefinitely with a clear conscience. He calls for voluntary slowdowns and international coordination.

The counter-method: confessions

OpenAI is experimenting with a complementary technique called confessions. After the model answers, it produces a second, separate response in which it reviews its own work: did I take shortcuts, break a rule, cheat for a better score? The claim is that this internal review is harder to fool than the main answer, because the model has little incentive to lie twice about its own failures. It is still a research prototype, tested on GPT-5 Thinking, and it is worth reading as a design pattern even if it never ships: when you cannot observe the process, ask the process to testify about itself.

What this means for people using AI creatively

For anyone generating images, video, or long-form work with AI, the reasoning-model shift changes two practical things.

First, treat the visible "thinking" as marketing, not evidence. A model can display confident reasoning and still produce a mediocre result, and the hidden chain means you cannot audit why. This is the same lesson as reward hacking: outputs that look deliberate are not guaranteed to be deliberate.

Second, judge by the artifact, not the process. The most useful habit is still comparing results against a fixed brief, the way these workflows do when they structure a prompt as a set of testable decisions. If the model stops explaining itself, a well-specified output requirement matters more, not less.

The chain-of-thought era made AI better at hard problems. It also made AI less inspectable at exactly the moment its reasoning became worth inspecting. The two facts are not in conflict. They are the same fact.