How do I check whether an AI answer is true?

Language models generate plausible text rather than retrieving verified facts. That produces genuinely useful output most of the time and confident, well-formatted errors some of the time, and the two look identical. Knowing where the failure modes cluster is most of the skill.

Step by step

1. Know where errors cluster.Specific numbers, dates, names, citations, quotations, legal and medical specifics, and anything very recent. Fluent general explanation is much more reliable than precise detail.
2. Ask it to show its sources, then check them.Do not stop at being given a citation. Fabricated references are a well-documented failure and they look entirely convincing. Open the source.
3. Check anything consequential against a primary source.Official guidance, the actual legislation, the manufacturer's documentation, a reference work. Match the effort to the stakes.
4. Ask the same question differently.Rephrase, or start a fresh conversation. Answers that shift substantially between attempts indicate the model is uncertain even though it does not sound it.
5. Watch for excessive confidence on obscure things.These tools rarely say "I don't know". A crisp, detailed answer about something very specialised deserves more scepticism, not less.
6. Ask it to argue the opposite.Requesting the counterargument or the weaknesses in its own answer surfaces caveats that the first response left out.
7. Never rely on it alone for medical, legal or financial decisions.Useful for understanding the shape of a question and what to ask a professional. Not a substitute for one.
8. Be most careful when it agrees with you.These tools tend towards agreement. A confirmation of what you already believed is the answer least likely to be independently checked.

Tips

• Good uses: explaining a concept, drafting, summarising text you provide, brainstorming, restructuring. Weak uses: precise facts, current events, exact citations, arithmetic.
• Giving it the source material and asking it to work from that is far more reliable than asking it to recall something.
• If you could not defend the answer to someone who knew the subject, do not act on it yet.

Common mistakes

• Trusting a citation without opening it — Fabricated references are a known failure mode and they look completely plausible — correct format, real-sounding journal, wrong or nonexistent paper.
• Treating fluency as accuracy — These systems are optimised to sound right. Confidence carries no information about correctness.

Questions people ask