How to check an AI answer before you rely on it

AI models write fluently whether they are right or wrong. That is the problem: a made-up fact reads exactly like a real one. The fix is not to stop using AI. It is to check the parts that matter, quickly, before you act on them.

Know which answers need checking

Not every answer carries the same risk. Rewording an email you are about to read anyway needs no checking. These do:

• Facts, names, dates and numbers that you will repeat to someone else.
• Quotes and citations. Models are known to invent sources that look real.
• Anything legal, medical or financial. Check with a qualified person or a primary source.
• Code that will run on real data.
• Recent events. Every model has a knowledge cutoff, and may not know what changed since.

Make the model show its working

You can make checking much easier by asking for answers in a checkable form.

• Ask for quotes, not summaries. "Answer using direct quotes from the document, with the section each comes from." Quotes can be found and confirmed in seconds. The summarise with quotes prompt does this.
• Give it the source. Paste the document and say "answer only from the text below; if the answer is not there, say so". This is called grounding, and it cuts made-up answers sharply.
• Ask what it is unsure of. End your prompt with "List any claims you are not certain of." Models are surprisingly willing to flag their own weak spots when asked.

Check the claims, not the prose

Once you have an answer, pull out the specific claims: each number, name, date and cause. Then check each one against a source you trust. The fact-check claims prompt will list the claims in any text for you, with what kind of source would confirm each one.

If you check things like this often, the source checker skill teaches your assistant to do this step on its own.

Watch for agreement that is too easy

Models tend to agree with you. If you ask "isn't it true that...", you will often get "yes" even when the honest answer is "no". This is called sycophancy. To get a straight answer:

• Ask neutral questions: "Is this true or false, and why?" rather than "This is true, right?"
• Ask for the strongest case against your view.
• Start a fresh chat and ask the question cold.

Run the code before you trust it

For code, the check is simple: run it, on a copy of real data, and read the output. Ask the model to write tests alongside the code. The write unit tests prompt is a good start. If something fails, paste the exact error back. That works far better than describing it.

When to stop checking and ask a person

If an answer will be used to make a decision about someone's health, money, legal position or safety, an AI answer is a starting point for a conversation with a professional, not the end of one. Our terms say the same.

The 60-second routine

1. Is this low risk? If yes, use it.
2. Did I give the model the source? If not, do that and ask again.
3. Pull out every fact, number and quote.
4. Check each against a source I trust.
5. For code, run it.

It takes a minute, and it is the difference between AI that saves you time and AI that costs you trust.