Definition
Chain of thought is Chain of thought is the step-by-step reasoning an AI model writes out before its final answer, which often improves accuracy on hard problems and lets people inspect how it got there.
A chain of thought is the series of intermediate steps a model writes down while working toward an answer. Instead of jumping straight to a result, the model "thinks out loud": it breaks a problem into parts, works each one, and then combines them.
Think of a math teacher who says "show your work." The final answer matters, but the work is how you catch mistakes.
A language model produces text one token at a time. Every token it writes becomes part of its own input for the next one. When a model writes out intermediate steps, those steps give it more room to compute, because each step conditions the next. On multi-step problems such as math, logic puzzles and planning, that extra working space tends to produce better answers than a one-line reply.
You can ask for it in a prompt ("work through this step by step"). Many newer systems are also trained to do it on their own. That is the idea behind a reasoning model, which spends extra tokens on a visible or hidden chain of thought before it replies.
A written chain of thought is not always a faithful record of why a model produced its answer. The text can read as tidy and convincing while leaving out what actually drove the result. Treat it as a useful clue and a way to spot errors, not as proof.
Ask a model, "A shirt costs $40 after a 20% discount. What was the original price?" A direct guess might say $48. With a chain of thought, the model writes: the sale price is 80% of the original, so the original is 40 divided by 0.8, which is $50. Seeing the steps makes it easy for you to catch a wrong turn.
Chain of thought is the core technique behind the reasoning model. Its safety value connects to AI alignment, since readable reasoning is one way to check what a model is trying to do.
It is asking a model to work through a problem step by step before answering, for example with the phrase 'think step by step'. It often improves results on math, logic and planning tasks.
No. The written reasoning can look convincing without fully reflecting what produced the answer, so use it to spot errors rather than as proof.
OpenAI says Tomek Korbak, Mikita Balesni and Jasmine Wang broke its rules on handling sensitive information. Their letter says the firings chill safety work, and the real fight is over outside auditors' access.