Chain of thought means getting a language model to write out intermediate steps before its final answer. It started as a prompting technique: asking it to "think step by step" or showing it worked examples. On multi-step problems it tends to reduce mistakes. Many reasoning models generate internal steps without that instruction.
A question from the app. Pick an answer.
A model shows a step-by-step explanation. What can you conclude?
They are not a recording of what happens inside the network. It can show tidy steps that the answer does not really follow from.
The steps are generated text. They can help you check, but they may contain errors or leave out how it got to the answer.
A convincing explanation can hide a mistake. Check the math and the claims.
A notebook and a pen cost 1.10 USD in total. The notebook costs 1 USD more than the pen. How much is the pen? The quick answer is 0.10. It is wrong: the notebook would cost 1.10 and the total 1.20.
With steps: the pen is x and the notebook is x + 1. So 2x + 1 = 1.10, and x = 0.05. A language model that answers directly can fall for the same 0.10. If it writes the math first, each step is out in the open and becomes part of the text it uses to predict the next one.
As a technique, you just put it in the prompt: "think step by step", "show your work", or a couple of solved examples with their steps so it copies the format.
Many reasoning models generate internal steps without being asked: those are the reasoning tokens. What you see may be a summary. In some APIs they are billed as output and the effort setting adjusts how much it reasons. With these models, asking for "step by step" is often unnecessary, and some providers advise against it. Check the guide for the one you use.
The steps you see are generated text, just like the answer. They can help it get there, but they are not a recording of what happens inside the network. A model can show tidy steps that the answer does not actually follow from, or get a step wrong and still land on the right answer.
They are useful for checking, not as proof. Convincing steps do not make the answer true (they can also contain a hallucination): if it matters, redo the math yourself.
For math, logic, code or decisions with several conditions. For a simple fact or rewording a text, asking for steps may add little and cost extra tokens.