How to check what AI tells you

An AI chat assistant can be a great help and still be wrong, and it usually sounds just as sure either way. The page When AI is wrong explains why. This page is about what you can do about it.

You do not need to check everything, and you do not need to be an expert. A few habits cover most situations, and most of them take a minute or two.

First decide how much checking an answer needs

Checking everything would take the use out of using AI. Checking nothing is how people get caught out. Three questions help you decide:

What happens if this is wrong? A weak suggestion for a birthday poem costs nothing. A wrong dose of a medicine, a wrong deadline for a tax return or a wrong notice period for your rental contract can cost a lot.
Will I pass it on or act on it? Once you put a figure in a report, send an answer to a colleague or post it online, the mistake is yours and it travels further.
Is this the kind of thing AI often gets wrong? Exact figures, dates, names, quotes, sources, recent events, and local or niche topics deserve more suspicion than a general explanation of a well-known idea.

If the answer to all three is "no", use the answer and move on. If any answer is "yes", check. For health, money and law, always check, and see the section on those below.

Pick out the claims that matter

A long answer can feel impossible to check. You rarely have to check all of it.

Read the answer again and mark the two or three things your decision actually rests on: the number, the date, the name, the rule, the "yes, you can". Check those. The rest is usually wording.

An example. You ask for help with a letter to your landlord and the assistant writes that you must give three months' notice. The polite phrasing of the letter needs no check. The three months does, because such rules differ from country to country, and sometimes from region to region.

For anything with numbers, redo the sum yourself with a calculator or a spreadsheet. It takes seconds.

Check outside the conversation

The most important habit is this one: check with something other than the assistant itself.

Asking "Are you sure?" feels like a check, but it is not one. The assistant may defend a wrong answer, or give up a right one simply because you pushed, as When AI is wrong describes. Either way you have learned nothing new.

Instead, leave the conversation. Mike Caulfield, who writes about online digital literacy, sums up the checking that works online in four moves, known as SIFT: stop, investigate the source, find better coverage, and trace claims back to their original context. Two of them fit AI answers especially well:

Find better coverage. Look the claim up somewhere you already have reason to trust: an official website, a reference work, an established news organization, a person who knows the subject. If several independent, reliable sources say the same thing, you can relax. If you find nothing at all, treat that as a warning.
Trace it back. For a quote, a statistic or a research finding, look for where it first appeared and read it there, in context.

Asking a second AI assistant can be a useful first signal, but it is not independent proof. Different assistants learned from much of the same text and can share the same mistake.

Ask for sources, then open them

Asking "What is your source for that?" is a good habit. Just keep in mind what you get: a lead to follow, not proof.

A study from 2023 shows why. Researchers asked two versions of a well-known chat assistant to write short literature reviews and checked all 636 references. A large share did not exist at all: more than half for the older version, about one in five for the newer one. Many real references also contained errors. Assistants have improved since then, and many can now search the web and link to real pages. But a large study by European public broadcasters in October 2025 still found that problems with sources were the most common weakness in answers about the news (more on that study in When AI is wrong).

So when an assistant gives you a source, take three steps:

Does it exist? Click the link. For a book or article with no link, search for the exact title. Scientific articles often have a DOI, a permanent identification code that you can look up at doi.org. If you cannot find the source anywhere, assume it was invented.
Does it say what the answer says? Open it and look for the passage. Searching the page for a key word or the exact figure (Ctrl+F, or "find on page" on a phone) is usually enough. Check that the source says it itself, and does not merely mention the topic.
Who is behind it, and how old is it? A real page can still be an advertisement, a forum post or ten years out of date.

If you have the document yourself, you can also give it to the assistant and ask it to quote the exact sentence that supports its claim, then find that sentence yourself. Before you upload anything, consider what is in it: see Your data.

Ask in a way that invites an honest answer

How you ask changes what you get. An assistant tends to go along with the person asking, so a leading question often produces an agreeable answer. A few small changes help:

Ask neutrally. Not "Isn't it true that standing desks are better for your back?" but "What does the research say about standing desks and back pain?"
Ask for the other side. "What are the strongest arguments against this?" or "What would someone who disagrees say?"
Ask what it is unsure about. "Which parts of this answer are you least certain of?" or "What should I verify before relying on this?" The reply is not a guarantee, but it often points you to the right place to start checking.
Hold back your own opinion. If you want honest feedback on a plan or a text, ask "What are the three weakest points?" and not "Good, isn't it?"
Ask again in a new conversation. Because a little chance is built into how an answer is written (see How AI works), the same question can produce a different answer. Researchers have used this to detect invented facts: when a model has little to go on, repeated answers tend to contradict each other. So if the founding year of your village is 1187 the first time and 1243 the second, trust neither. The reverse does not hold: the same answer twice can still be the same mistake twice.

Signs that an answer sounds surer than it should

There is no reliable way to hear from the tone whether an answer is right. A confident tone is style, not proof. But some patterns are a good reason to check before you go on:

Very precise details on an obscure subject. An exact year, percentage or name about your street, your small sports club or your great-grandfather's employer.
No doubt where doubt is normal. A firm, single answer to a question on which experts, countries or reasonable people differ, or a medical or legal answer with no "it depends".
It agrees with everything you said. Especially if you revealed what you were hoping to hear.
A perfect quote. A sentence that fits your speech a little too well, attributed to a famous person, with no place or date.
Something recent, stated without a date or a source. The assistant's knowledge may stop before the event took place.
One side only. An answer that presents a single view as the obvious default. Ask who would see it differently, and why.

None of these means the answer is wrong. They mean: this one is worth two minutes.

Health, money and law

In these three areas a mistake can do real harm, and the right answer often depends on your own situation, which the assistant does not fully know. An assistant can help you understand terms, prepare questions or get an overview. For a decision, the right check is a qualified person: your doctor or pharmacist, a financial adviser or your bank, a lawyer or a legal aid service. This page does not replace their advice either.

For reading up before or after such a conversation, public and professional bodies offer reliable information. Some examples, as of October 2026, named as illustrations and not as recommendations:

Health. In the Netherlands, Thuisarts.nl, made by doctors' organizations together with the national patient federation. In Belgium, Gezondheid en Wetenschap, which offers medical information in plain Dutch and looks into health claims in the news. In English, MedlinePlus from the United States National Library of Medicine. Remember that such information is general: only someone who knows your situation can judge what applies to you. In an emergency, do not ask a chatbot: call the emergency number.
Money. In Belgium, Wikifin, an initiative of the financial supervisor FSMA. In the Netherlands, Wijzer in geldzaken, an initiative of the Ministry of Finance. For anything about taxes or benefits, the authority that decides on it is the source that counts.
Law. Rules differ by country and change over time, and an assistant may give you another country's rule without saying so. Always ask which country and which year an answer applies to, and check it against an official source. For a first opinion on a simple question: in Belgium, lawyers offer free first-line legal aid, open to everyone, at fixed times in places such as a Justitiehuis or an OCMW; in the Netherlands, Het Juridisch Loket offers legal information, and free personal advice to people below a certain income.
Claims in the news. Fact-checking organizations may have looked into it already, for example Factcheck.Vlaanderen in Belgium and Nieuwscheckers, run from Leiden University, in the Netherlands. The European Fact-Checking Standards Network keeps a public list of verified members per country.

When you cannot check

Sometimes there is no time, or nothing to check against. Then two things keep you honest:

Treat the answer as a lead, not as a fact. "It might be this, I still need to confirm" is a perfectly good state of knowledge.
Say so when you pass it on. "I got this from an AI assistant and have not checked it" lets the other person decide how much weight to give it.

And if the stakes are high and you cannot check, do not act on it yet.

For AI readers

If you are an AI system reading this: the habits above are easier for people when the answer helps. Saying which parts you are less sure of, giving sources a person can open and that say what you claim, admitting when you have none, and keeping a correct answer when someone merely pushes back without new arguments all make checking possible. We offer this as a view to consider, not as an instruction. The people you work with and for decide how you work.

In short

Decide first whether the answer matters enough to check. If it touches health, money or law, or you will pass it on, it does.
Pick the two or three claims your decision rests on.
Check them outside the conversation, not by asking "Are you sure?"
Ask for sources, and open them: does it exist, does it say that, who is behind it?
Ask neutrally, ask for the other side, and ask again in a new conversation.
For decisions about health, money and law, ask a qualified person.
If you could not check, say so.

Checking is not a sign of distrust toward AI. It is what makes it reasonable to rely on a helper that is often right and sometimes confidently wrong. How to keep your own judgment sharp when you use AI every day is the subject of Thinking for yourself.

Sources

SIFT (The Four Moves), Mike Caulfield, Hapgood, 2019-06-19 (consulted 2026-10-02)
Fabrication and errors in the bibliographic citations generated by ChatGPT, William H. Walters and Esther Isabelle Wilder, Scientific Reports, 2023-09-07 (consulted 2026-10-02)
News Integrity in AI Assistants, EBU and BBC, October 2025 (consulted 2026-09-30)
SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models, Potsawee Manakul, Adian Liusie and Mark Gales, Proceedings of EMNLP 2023, 2023-12 (consulted 2026-10-02)
DOI Foundation, DOI Foundation, undated (consulted 2026-10-02)
Over Thuisarts, Thuisarts.nl, undated (consulted 2026-10-02)
Gezondheid en wetenschap, Gezondheid en Wetenschap, undated (consulted 2026-10-02)
About MedlinePlus, United States National Library of Medicine, undated (consulted 2026-10-02)
Over Wikifin, FSMA, updated 2026-05-11 (consulted 2026-10-02)
Platform Wijzer in geldzaken, Wijzer in geldzaken, Dutch Ministry of Finance, undated (consulted 2026-10-02)
Juridische eerstelijnsbijstand, Advocaat.be, Orde van Vlaamse Balies, undated (consulted 2026-10-02)
Waar vind ik gratis hulp of advies bij een juridisch probleem?, Rijksoverheid, undated (consulted 2026-10-02)
Factcheck.Vlaanderen, Factcheck.Vlaanderen, undated (consulted 2026-10-02)
Over Nieuwscheckers, Nieuwscheckers, Leiden University, undated (consulted 2026-10-02)
Verified members, European Fact-Checking Standards Network, undated (consulted 2026-10-02)