How to Fact-Check AI Answers

AI chat answers are fluent, well organised and often right. That is exactly what makes the wrong ones risky: an invented statistic looks just as convincing as a real one. You do not need to check everything, but you do need a fast, reliable routine for the things that matter. This guide gives you one.

Why AI answers can be wrong

• It predicts plausible text. A language model generates words that fit; it does not look facts up unless a search tool is involved. Our guide on how AI chatbots work explains why.
• Its knowledge has a cutoff. Prices, laws, office holders, product versions and timetables change after the training data was collected.
• It can invent sources. Titles, authors, page numbers, links and court cases can be fabricated in a perfectly realistic format.
• It tends to agree with you. A leading question such as "Isn't it true that...?" can push it towards confirming what you already believe.
• It mixes things up. Two people with similar names, two similar laws, or two versions of the same software.

Decide what needs checking

Match the effort to the stakes. As a rough guide:

Specific claims deserve the most attention: numbers, dates, names, quotes, laws, doses, prices, technical specifications and references.

A five-step fact-checking routine

1. Break the answer into claims

Ask the AI to do the first part for you:

List every factual claim in your previous answer as a numbered list.
For each one, say how confident you are (high, medium or low) and what
kind of source would confirm it.

The confidence labels are not reliable on their own, because models can be confidently wrong, but the list shows you exactly what to check.

2. Look for the original source

For each important claim, find where it comes from. Prefer primary sources: the official statistics office for numbers, the text of a law on the government's own legislation site, the study itself in the journal, the manufacturer's manual, the organisation's own website. A news article or blog that quotes a number is one step removed from it.

3. Read laterally

Professional fact-checkers do not judge a source by how professional it looks. They open new tabs and check what other reliable sources say about the claim and about the source itself. If several independent, reputable sources agree, you are on solid ground.

4. Check that citations exist and say what is claimed

If the AI gives you a reference, search for the exact title in a search engine, a library catalogue or a scholarly database. Then open it and confirm it actually supports the claim. AI sometimes cites a real paper for something the paper does not say.

5. Check the date

Is the information current? Look at when the source was published or last updated, and whether anything has changed since then: new rules, new versions, new prices.

Prompts that make answers easier to check

How you ask can reduce errors before they happen:

Weak:
What are the side effects of [medication]?

Stronger:
What are the commonly listed side effects of [medication]? Separate
well-established facts from things you are less sure about, and say
clearly if you do not know. Tell me which official document, such as
the patient information leaflet, I should check to confirm.

Other useful lines to add:

• "If you are not sure, say so instead of guessing."
• "Answer only from the text I pasted below. If the text does not say, tell me."
• "What would someone who disagrees with this answer say?"
• "Which parts of this answer might be out of date?"

Answering from text you supply is especially dependable. Paste the document and ask questions about it, or use the summarizer, then spot-check the result against the original. You can also open a claim-checking prompt in the chat to plan your checks.

Red flags in an AI answer

• Very precise numbers with no source, such as a percentage to one decimal place.
• Quotes attributed to famous people that you cannot find anywhere else.
• References with odd details: a journal that does not cover the subject, an author who does not work in the field, a link that does not open.
• An answer that changes completely when you regenerate it or ask "Are you sure?"
• Legal or medical statements with no mention of country, dose or circumstances.
• Information about recent events presented without a date.

Asking "Are you sure?" can be revealing, but do not treat a changed answer as the correct one. Models sometimes back down from right answers under pressure as well. Check outside the chat.

Does asking another AI help?

Comparing answers from different model styles in the chat menu, or from another AI service, can reveal disagreements worth looking into. But AI systems are trained on overlapping data and can share the same mistakes, so agreement between two AIs is not proof. Use them to find what to check, not as the check itself.

Students and researchers will find more ideas in AI for students, ChatG for researchers and the study prompts.

Where to check different kinds of claims

FAQ

How often is AI wrong?

It varies a lot by model, topic and type of question, so there is no single honest number. Well-known topics are generally handled better than niche subjects, recent events, exact figures and citations.

Is it safe to use AI for medical or legal questions?

Use it to understand terms, prepare questions and get general background. Do not use it to make decisions about treatment, medication, legal action or money without a qualified professional. In an emergency, contact your local emergency services.

Should I ask the AI for sources?

Yes, because it tells you what to look for. But always open each source yourself, because AI can produce references that do not exist or do not say what it claims.

Why do links in AI answers sometimes not work?

Unless the AI actually looked a page up, a link in its answer is generated text like everything else. It may point to a page that has moved, or to an address that never existed but looks right. Search for the title or the organisation instead of trusting the address.

What is the fastest check for a single fact?

Search for the specific claim and look for a primary or highly reputable source among the first results. If you cannot find one within a few minutes, do not rely on the claim.