"Think step by step" is the most popular prompting technique for improving AI accuracy. It makes the model decompose problems into steps before answering. It genuinely improves reasoning on math and logic tasks. It does not prevent hallucination. A wrong answer arrived at step by step is still a wrong answer - it is just harder to spot.
Chain of thought (CoT) prompting asks the model to show its reasoning process before giving a final answer. Instead of jumping from question to answer, the model generates intermediate steps.
This works well for mathematical and logical reasoning. When the model writes out "2 + 3 = 5, then 5 × 4 = 20," it is less likely to jump to a wrong answer than if it tries to compute the entire thing in one step.
The mechanism: CoT gives the model more computation space. Each intermediate step becomes part of the context, which the model uses to generate the next step. More steps means more chances to course-correct within the model's own reasoning.
Hallucinated premises propagate through steps. If the first step contains a hallucinated fact, every subsequent step builds on it. The chain of thought makes the hallucination more elaborate, not less likely. Each step adds supporting reasoning for a conclusion that was wrong from the start.
Confident wrong reasoning. The model generates reasoning that sounds logical but is based on incorrect information from training data. The steps are internally consistent. The logic is valid. The premises are wrong. The result: a beautifully reasoned hallucination.
The model cannot verify facts mid-chain. When a human solves a problem step by step, they can pause at any step and check: "Is this actually true?" The model cannot. It generates each step based on what came before and its training data. It has no mechanism to verify whether the intermediate claims are correct.
More convincing hallucinations. A hallucination with step-by-step reasoning is harder to spot than a bare hallucination. The reasoning gives it an aura of rigor. A reader follows the logic, finds it sound, and accepts the conclusion - not realising that one of the premises was fabricated.
Chain of thought shows you how the model reached its answer. It does not tell you whether the answer matches reality. Showing work is not the same as checking work.
Newer "reasoning models" like o1 and o3 use extended chain of thought internally. They decompose problems into many steps, reason through alternatives, and arrive at answers with more computation. They are genuinely better at complex reasoning tasks.
But they still hallucinate. The additional computation helps with tasks where the answer can be derived from the prompt and training data alone. It does not help with tasks where the answer depends on information the model does not have - live system state, current data, recent changes.
More reasoning about incomplete information produces more confidently wrong answers. The model thinks harder about what it knows. It does not know what it does not know. The reasoning is excellent. The knowledge gap is unchanged.
Chain of thought and verification solve different problems. CoT helps the model reason through complex logic. Verification ensures the model's claims match reality.
The best systems use both. Let the model think step by step. Then verify each step's factual claims against live data before the final answer reaches the user.
Check provides the verification side. It injects real-time environment state into the model's context - so the model's reasoning starts from facts, not training data - and validates every output against the live environment. CoT with verification means the model reasons well and reasons from verified reality.
120 verifications a day free. No card, no signup.