Why asking an AI to "think step by step" makes it smarter

Ask an AI a tricky math or logic question and it may blurt out a wrong answer instantly. But add five magic words — "let's think step by step" — and its accuracy can jump dramatically. This trick is called chain-of-thought prompting, and it's one of the most surprising discoveries in how large language models work.

The reason is mechanical, not magical. A model generates its answer one word at a time, and each word it produces becomes part of what it reads next. When it writes out intermediate reasoning, those steps act like scratch paper — the model can lean on its own partial work instead of trying to leap straight to a conclusion. Skipping the reasoning forces it to compress a multi-step problem into a single guess, which is exactly when it stumbles.

This is why newer systems increasingly do this reasoning before giving a final answer, sometimes hidden from you. The insight flips a common assumption: the words an AI writes aren't just its output — for hard problems, they're also part of its thinking process.