Chain-of-thought started out as a prompt technique: adding "think step by step" made the model write out the intermediate steps, and on complex tasks that made it measurably more accurate. The mechanism is simple — each step it writes down lands in the model's own context, and the next step builds on what is already there.
Today the technique is largely built in. Models with a thinking mode work through the reasoning on their own before they answer, so asking for it separately is usually pointless. In some places it actively hurts: on a simple, one-step task, forced reasoning is longer, more expensive and occasionally leads the answer astray.
The most common misconception is that the written reasoning is the model's real train of thought. It is not — it is an after-the-fact text explanation, which can be wrong next to a correct answer and right next to a wrong one. That makes it only partly usable for checking.
What is left of it in 2026, and when asking for the steps still pays, is the subject of Chain-of-Thought: When Asking for the Steps Still Pays, a post of its own. Directly related terms: reasoning effort, prompt and few-shot.