How do I get an AI to show its work?

To get an AI to show its work, ask it to do three things for every important claim: name a source you can actually open, label how confident it is (documented fact, reasonable inference, or guess), and state what evidence would prove the claim wrong. This forces the reasoning and the sourcing into the open instead of a smooth paragraph. Then verify the sources yourself, because models sometimes invent citations that look real. A claim that can name its own undoing is trustworthy. One that just insists is not.

By Jason Santiago, author of Let the AI Be Smart. As of 2026-07-06.

Why the default answer hides its work

Left alone, a model gives you a fluent conclusion with the reasoning compressed out of sight. That reads well and tells you nothing about whether it is right. You have to explicitly pull the work into view.

The three asks that force it open

Name a source. For each load-bearing claim, ask: where does this come from, and can I open it? Not "studies show." An actual, checkable source.
Label the confidence. Ask the model to mark each claim as a documented fact, a supported inference, or a hypothesis. This alone stops guesses from wearing a fact's clothes.
Name what would break it. Ask: what evidence would make this claim wrong? A real claim can answer. A hallucination cannot, and that tells you which is which.

The catch: verify the sources

Making a model show its work does not make the work true. Models fabricate citations that look completely real. So the final step is yours: open the sources it named and confirm they say what it claims. If you cannot open it, it does not count. See how to fact-check AI without being an expert.

This "show your work" discipline is the Claim Ledger, one of the free tools from the book, and it is a core part of the Council Method. The full method is in Let the AI Be Smart.