Chain-of-Thought Reasoning — Explain Like I'm 5

Show Your Work

In school, you’re often told to show your work — not just write the answer, but write out the steps you took to get there. There are two reasons: first, you might catch mistakes in your own reasoning. Second, even if you get the answer wrong, the teacher can see where you went wrong and give partial credit.

Chain-of-thought prompting applies this same idea to AI.

Before chain-of-thought, AI would read a math problem and immediately jump to an answer: “12.” It might be right, but it was just pattern-matching, not real reasoning.

With chain-of-thought, the AI reads the problem, then writes: “Let me think through this. If there are 5 apples and we add 3 more, we now have 5 + 3 = 8 apples. Then we remove 2, so 8 - 2 = 6. The answer is 6.” Now the AI can catch its own errors as it goes, and the reasoning is much more reliable.

The Surprising Result

Jason Wei and colleagues at Google discovered in 2022 that adding just a few example reasoning chains to the AI’s prompt — showing it how to think step-by-step — dramatically improved performance on complex problems.

On grade-school math problems, accuracy jumped from 18% to 57%. On something called GSM8K (grade school math), GPT-3 went from basically useless to genuinely good, just by showing it examples of thinking out loud.

The trick works even with just one magic phrase: “Let’s think step by step.” Appending this to any question causes large AI models to spontaneously produce reasoning chains — a capability that was always there but needed the right prompt to activate.

What It Changed

Chain-of-thought prompting revealed that AI models contained far more reasoning capability than anyone realized — it just needed to be expressed externally rather than computed internally.

This discovery drove a wave of AI improvements: “reasoning models” like o1 and o3 are now trained to spend much more time on chain-of-thought reasoning, allowing them to solve genuinely hard problems in math, science, and coding that simpler prompting couldn’t touch.

One thing to remember: Chain-of-thought works because complex reasoning requires intermediate steps — when AI is forced to make those steps visible, it can follow them more carefully instead of jumping to a potentially wrong conclusion.