Lesson 5 of 7
Watching it reason
8 min read
Read a reasoning model's working and you'll notice something odd: mid-thought it writes “wait — that's not right”, crosses out its own step, and tries again. Nobody programmed that hesitation. Where does it come from?
Thinking out loud
When a model practises on hard problems, the attempts that succeed tend to share a habit: they think out loud first. They restate the problem, lay out the steps, try a path, and check it. Because that habit keeps reaching the answer, reinforcement learning makes it stronger — so a well-trained reasoning model naturally produces a long chain of thought before it commits to an answer.
The step-by-step working isn't decoration. It's the strategy practice found most reliable — so the model learned to lay its reasoning out before answering.
“Wait — that's not right”
The striking part is what shows up inside that working: the model second-guesses itself. It'll follow a line of reasoning, notice it doesn't hold, write something like “wait, let me reconsider”, and backtrack. No human scripted that move — it emerged, because catching your own error mid-solution is a great way to end up correct, and reinforcement learning rewards ending up correct. Step through a reasoning trace below and watch it happen.
Self-correction wasn't taught — it was discovered. Because practice rewards getting the final answer right, the model learned to double back and fix itself, exactly as a careful person would.
Why doubling back pays off
It's worth being exact about why this shows up. During practice, an attempt that charges ahead on a wrong step usually ends up wrong — and gets thrown away. An attempt that pauses, notices the wrong step, and fixes it ends up right more often — and gets kept. Nobody ever rewards the word “wait”; practice only ever rewards the final answer being correct. Self-checking survives because it's a dependable route to that.
Self-correction isn't a personality the model was handed. It's a by-product of rewarding correct answers: the habit that catches mistakes is the habit that ends up right, so practice keeps it.
This self-checking behaviour appearing on its own was the headline of open reasoning models like DeepSeek-R1, whose write-ups nicknamed it the “aha moment”: partway through training, the model spontaneously began re-examining its own steps — a habit nobody put there by hand.
Recap
- —Practice pushes a model to produce a chain of thought — its reasoning, laid out step by step — before answering
- —Inside that reasoning it learns to self-correct: notice a wrong step, backtrack, and fix it
- —Nobody scripted the hesitation — it emerged because reaching the right answer is what practice rewards
A reasoning model, mid-answer, writes “wait, that doesn't work” and redoes a step. What best explains this behaviour?
Continue in the app
Take the whole How AI Learns to Think course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.