How to Check What AI Tells You

How to Check What AI Tells You

Most AI mistakes do not look like mistakes. They look like clean, confident, well-written answers that happen to be wrong in one specific place — a number, a date, a name, a rule, a citation.

That is the whole problem. Fluent writing reads like competence, so people stop checking. In workplace surveys, roughly two thirds of employees say they use AI output without verifying it, and over half report a work mistake that came from trusting it too quickly.

This lesson is the habit that fixes that. It takes about a minute, and it is the difference between AI making you faster and AI making you wrong faster.

The 60-second check

Before any AI output leaves your hands — into a document, a client email, a post, a decision — run these three checks in order. They stack. Each one catches something the others miss.

1. Ask it for its sources and assumptions.
2. Make it argue against itself.
3. Separate claims from evidence.

You will not do all three every time. By the end of this lesson you will know which stakes deserve which check.

Check 1 — Ask for the source and the assumption

The fastest way to expose a shaky answer is to ask where it came from. Paste this after any answer you plan to rely on:

List every factual claim you just made. For each one, say whether it comes from a source you can name, from general knowledge, or from an assumption you made. Mark anything you are not confident about.

Two things happen. Weak claims get flagged, because the model is much better at rating its own certainty than at avoiding errors in the first place. And hidden assumptions surface — the ones where it quietly guessed your industry, your country, your budget, or your timeline.

If it names a source, open the source. A named citation is not the same as a real one. Checking one link takes ten seconds and catches the worst failures.

Check 2 — Make it argue against itself

AI defaults to agreeing with you. Ask a leading question and you will usually get a supportive answer, whether or not it is true. The fix is to force the other side:

Argue the strongest case against what you just told me. What would have to be true for this to be wrong? What is the best counter-argument someone who disagrees would make?

This is the single most useful prompt in this lesson. If the counter-argument is weak and it struggles to find one, the original answer is probably solid. If the counter-argument is immediately strong, you have just saved yourself.

Do this before you act on strategy advice, pricing, legal or medical framing, or anything where you already wanted a particular answer.

Check 3 — Separate claim from evidence

For anything going in front of other people, convert the output into four columns before you use it:

• Claim — what is being asserted
• Evidence — what actually backs it
• Confidence — high, medium, or low
• Risk — what breaks if this is wrong

You can ask the AI to build this table itself. The point is not the table. The point is that vague statements cannot hide in it. A sentence like "this approach typically improves retention" looks fine in a paragraph and looks obviously unsupported in a row.

Anything that lands at low confidence and high risk gets verified by a human, or gets cut.

When to check hard, and when not to bother

Checking everything is how people quietly abandon the habit. Match effort to stakes.

• Skip it for brainstorming, rewording, formatting, summarising something you wrote, or anything you will read closely anyway.
• Run check 1 for research, explanations, and anything with numbers or dates in it.
• Run all three for anything that reaches a customer, informs a decision, involves money or law or health, or carries your name publicly.

The rule underneath all of it: a human checks every high-stakes claim before it reaches a decision or an audience.

The easiest version of this is a saved prompt you keep one paste away. Make your own, or start with this:

Before I use this: list your factual claims and mark each as sourced, general knowledge, or assumption. Then give me the strongest argument that you are wrong. Be direct — I would rather find the problem now.

Run it for two weeks on real work. You will stop needing it as a prompt, because you will start reading AI output the way you read a confident stranger — interested, but not convinced yet.

That instinct is the actual skill. Everything else here is scaffolding to build it.