AI Tip

Make AI Show Its Work Before You Trust the Result

AI is very good at handing you a conclusion before you have seen the reasoning that supports it.

That is convenient when the task is low stakes. It is a problem when the answer will become a customer email, a project decision, a published claim, or the next step in an automated workflow.

The usual advice is to fact-check the output afterward. That still matters, but you can make the checking easier by changing the request. Do not ask only for the answer. Ask for a small evidence packet alongside it.

A useful version looks like this:

Give me the recommendation first. Then list the evidence you relied on, the assumptions you made, the parts you are uncertain about, and the checks I should perform before acting on it. If information is missing, say so instead of filling the gap.

This does not make the model truthful by magic. It makes the result easier to inspect.

Ask for four things

1. Evidence

Ask the model to identify the source, input, or observation behind each important conclusion. If the answer came from material you supplied, it should point back to that material. If it did not have a source, you need to know that before treating the statement as a fact.

For research work, ask for links or source titles. For an internal document, ask for the exact section or paragraph. For a data task, ask which fields were used and which were ignored.

An answer that cannot show where a claim came from may still contain a good idea. It has not earned the status of verified information.

2. Assumptions

Most useful answers contain assumptions. The dangerous ones are the assumptions nobody can see.

Ask: “What had to be true for this recommendation to make sense?” You may discover that the model assumed a deadline, a budget, a customer type, or a technical constraint that was never in the brief.

This is often the fastest way to find the real problem. You are not correcting the answer yet. You are checking whether it answered your question or a cleaner question it invented.

3. Uncertainty

Do not rely on a confidence score. Ask the model to name the uncertain parts in plain language.

“What would you want to verify first?” is a better prompt than “How confident are you?” It forces the review toward actions rather than a made-up percentage.

You can also ask it to separate known facts, reasonable inferences, and guesses. Those are different kinds of material and should not be allowed to blend together in a final email or plan.

4. A review checklist

Finish by asking for the smallest set of checks that would change your decision. Keep it short. A ten-point list that nobody uses is worse than three checks that happen every time.

For example:

• Confirm the date and owner of the source data.
• Verify the two numbers that drive the recommendation.
• Test the proposed action on one small example before applying it broadly.

That final item matters. A review should lead to an observable check, not just a feeling that the answer looks sensible.

Use a two-pass workflow

For anything important, separate generation from approval.

In the first pass, ask AI to produce the answer and its evidence packet. In the second pass, ask it to act as a reviewer: find unsupported claims, hidden assumptions, missing inputs, and instructions that are more confident than the evidence allows.

Then do the human check on the few items that can actually hurt you.

This is not about forcing AI to expose some perfect internal chain of thought. You do not need a theatrical explanation of every token. You need an auditable summary: what it used, what it assumed, what it does not know, and what you should verify.

The goal is simple. Make unsupported certainty inconvenient before it reaches the next person.

Practical prompt:

Answer the task using only the information provided. Return: (1) recommendation, (2) supporting evidence with source locations, (3) assumptions, (4) unknowns and risks, and (5) three checks before use. Label guesses clearly. Do not invent missing facts.

Use that prompt as a starting point, then adapt the checks to the cost of being wrong. A shopping list does not need a review board. A public claim, financial decision, or automated action does.

AI does not become trustworthy because it sounds careful. It becomes more useful when the workflow makes its uncertainty visible and gives you a cheap way to test the parts that matter.