How to check an AI answer in two minutes

Why a check is needed at all

A chatbot can be wrong and still sound sure. The wording is fluent whether the content is right or not, so tone tells you nothing. The useful question is not "does this sound right?" but "what happens if this one part is wrong?"

How much you check also depends on you. In a survey of 319 knowledge workers (Lee and others, Microsoft Research and Carnegie Mellon, CHI 2025, peer reviewed), people with more confidence in the AI reported less critical thinking about its output. People with more confidence in their own skill reported more. It is a survey of what people said they did, so it shows a link and not a cause.

The two-minute check

1. Guess first (15 seconds)

   Before you read the answer, write one line: what do I expect it to say? A guess gives you something to compare against. Without one, whatever the AI says becomes your starting point.

2. Find the load-bearing claim (15 seconds)

   Read the answer and pick the one fact, number or step your next action depends on. A date. A dose of fertiliser. A line of a contract. The function that does the real work.

3. Confirm it outside the AI (60 seconds)

   Look that one claim up in a source that is not a chatbot: the official page, the manual, the textbook, the documentation, or a person who knows. Asking a second AI is not a check. It may repeat the same mistake.

4. Open any source it cites (15 seconds)

   If the answer gives a link or a reference, open it. Check that it exists and that it says what the AI says it says.

5. Ask what would make it wrong (15 seconds)

   Type: "What is the weakest part of that answer? What would make it wrong?" You are not asking for reassurance. You are asking where to look.

Quick checks by type of answer

• Numbers: redo the sum yourself, roughly. If the AI says 15% of 240 is 46, you will catch it.
• Code: say what it will do before you run it, then run it on a case where you already know the result.
• Quotes and references: search for the exact quote or title. If you cannot find it, do not use it.
• Rules, law, tax, medicine: go to the official source or a professional. Treat the AI's answer as a list of things to ask about, not as the answer.
• Summaries: read the first and last paragraph of the original and check that the summary matches them.

When two minutes is not enough

Use the size of the consequence, not the length of the answer. If being wrong costs money, health, a grade or someone's trust, check more than one claim, and check with a person. If being wrong costs nothing (a film to watch, a synonym), skip the check and save the effort for where it counts.

Why the guess comes first

In a small classroom study of 42 university students (Islam and others, accepted at the AIED 2026 conference), students who were asked to think first and then consult the AI showed the largest drop in over-trusting what it told them. The study is small and it measured trust, not learning. It fits the survey above: people who bring their own view check more.

So step 1 is not a warm-up. It is the part that turns reading into checking.

Make the guess automatic

You can do all of this with nothing but a habit. If you want a prompt at the moment you ask, the think pad has a box for your attempt and sends it with your question, with the line "Check my thinking: tell me what I got right, then what I missed." The coach setup puts a standing instruction in your AI's own settings so it asks for your attempt before it answers. Both are free.

Questions