Chain of thought describes a model working through a problem in visible intermediate steps — laying out the reasoning before committing to an answer, the way a maths teacher insists you show your work. It began as a prompting trick. Researchers found that appending a phrase as simple as let us think step by step to a question measurably improved accuracy on problems the same model had just failed. The instruction added no new information. It changed only the shape of the output, and the shape turned out to matter.
The reason it works is more mechanical than it sounds. A language model produces one token at a time, and each token it generates becomes part of the input it reads for the next one. A model that answers immediately has only the question to work from. A model that writes out its steps has given itself a scratchpad — every intermediate line is extra context it can condition on, and extra computation spent before the answer is due. Thinking out loud is not a metaphor here. The words on the page are, quite literally, where some of the work happens.
This is also where the trouble starts. A chain of thought looks like an explanation, and it is tempting to read it as one — as a window onto the model's actual reasoning. It is not, or at least not reliably. The steps are produced by the same process that produces everything else: plausible text, conditioned on what came before. Researchers have shown that a model nudged toward a wrong answer by a biased prompt will often generate a fluent, confident chain of reasoning that never mentions the nudge, arriving at the planted conclusion by an argument that reads perfectly well. The written rationale and the underlying computation can come apart, and whether they have come apart in any particular case is not something you can tell by reading.
The technique has since stopped being a prompt and become an architecture. The current generation of reasoning models is trained to deliberate at length before answering, sometimes for thousands of tokens the reader never sees, with the length of that deliberation treated as a dial that can be turned up for harder problems. The gains on mathematics, code, and multi-step logic are substantial and well documented. They also cost real money and real time per question, which is a genuine shift: for most of this field's history, the cost of an answer from a trained model was roughly fixed. Thinking for longer is now something you can buy.
The human translation will be familiar to anyone who has talked their way through a hard problem and found the answer somewhere in the talking. Reasoning aloud genuinely helps, and it helps for much the same reason it helps a model: it puts the intermediate steps somewhere outside your head, so they do not all have to be held at once. The uncomfortable half of the parallel is just as familiar. People also produce confident accounts of why they decided something, and those accounts are frequently reconstructions assembled after the fact rather than transcripts of the decision. We have known this about ourselves for decades. It should not be surprising to find it in a system trained on everything we wrote.
