How to check if an AI answer is actually true

Why a confident answer is not the same as a correct one

You asked a question, received a clear answer and it sounded right. That is the moment worth slowing down for.

A generated answer can be fluent without having been verified. Wrong answers often arrive with the same steady tone and tidy structure as correct ones. OpenAI, Google and Anthropic all publish warnings that their assistants can provide inaccurate or misleading information.

In a 2025 study coordinated by the European Broadcasting Union, journalists from 22 public-service media organisations in 18 countries assessed more than 3,000 AI answers about news. Almost half had at least one significant issue, a third had serious sourcing problems and a fifth contained a major accuracy issue.

An answer and evidence are two different things

An answer is what the assistant produced. Evidence is something you can inspect independently, such as an official rule, product page, research paper or current timetable.

When an assistant tells you a hospital’s opening hours, that is an answer. When you open the hospital’s own website and read the hours, that is evidence. The two often match. When they do not, the original evidence wins.

Ask for sources, dates and uncertainty

Do not accept the first reply as the end of the conversation. Ask the assistant to separate its factual claims and admit what it could not verify.

You may receive useful links, but a link is still only a lead. An assistant can invent a source or attach a real source to the wrong claim. Tools such as Perplexitymake sources more visible, but visible does not mean verified.

Open the source

Click the link and find the part that supports the claim. Check that the page exists, belongs to the organisation it appears to belong to, and actually says what the AI claimed.

Prefer the organisation that owns the fact. Use a government page for a law, the provider’s pricing page for a subscription, the venue’s own site for opening hours and a health service or peer-reviewed paper for a medical claim.

Check the date

A correct answer from 2023 can be wrong today. Prices, software features, tax rules, visa procedures, eligibility, opening hours and public office-holders can all change.

Look for a publication date and an updated date. Then ask whether the type of information is likely to have changed since then. If a page has no date, treat that as another reason to compare it with something current.

Compare with a second independent source

For anything important, look for a second reliable organisation that says the same thing. Two news sites copying one press release are not really two sources; they are one source wearing two coats.

Independence matters because separate sources can catch different mistakes. Agreement does not guarantee truth, but disagreement tells you immediately that more checking is needed.

Match the check to the risk

Not every answer needs a research project. A restaurant suggestion might need one quick opening-hours check. A decision about a mortgage, medicine or legal deadline needs stronger evidence and a qualified person.

Ask one final question: if this turned out to be false, what would it cost? The answer tells you how far the checking needs to go.

How to recognise an invented citation

An assistant can produce a reference that looks completely normal: a plausible author, a professional title and a sensible year. None of that proves the source exists.

• The link fails, opens an error or lands on an unrelated page.
• Searching the exact title returns no original publication.
• The page exists but does not contain the quoted words or claim.
• The author is real but has not written the named publication.
• The title fits your question suspiciously perfectly.

This has caused real consequences. Courts have sanctioned lawyers and firms after submissions contained nonexistent cases, invented quotations or false descriptions of real decisions. The simple rule is non-negotiable: if you plan to cite it, open it first.

Health, law, money and safety need a different standard

AI can still help: ask it to explain terminology, organise your questions or summarise a document before an appointment. That can make a conversation with a professional more productive.

Do not use it as the decision-maker. A diagnosis, dosage, legal position, investment or safety judgement needs current evidence and a person who is qualified and accountable. In an emergency, contact the appropriate local emergency service rather than a chat window.

What to do when two good sources disagree

Check the dates first. Then ask whether the sources are answering the same question: a national average and a local figure can both be correct. Finally, consider who is closest to the fact. A regulator publishing a rule usually outranks an article describing that rule.

If a genuine disagreement remains, report it. “Reliable sources disagree” is a more useful answer than picking the version you prefer. Honest uncertainty is better than false confidence.