You asked an AI assistant a question, and the answer is clear, well written, convincing. Now what? You can't take it all at face value, but you're not going to spend an hour re-checking every word either.
You probably already know an assistant can make things up with a straight face. If you want to understand why, start with why ChatGPT makes things up. This article is the hands-on part: a simple five-step method to fact-check an answer from ChatGPT, Claude, Gemini or any other assistant, spending only as much time as the situation deserves.
Want to practice right away? Hallucination Hunt is a small free game: read AI answers, find the errors hidden inside.
Why "are you sure?" isn't enough
The most common reflex is to ask the assistant: "Are you sure?" Bad idea, or at least not nearly enough.
An assistant doesn't look anything up in a register of true facts. When you ask whether it's sure, it generates a new plausible answer to your new question. Sometimes it sticks to a mistake with the same confidence. Sometimes it apologizes and "corrects" something that was actually right, just because your question hints at doubt. Either way, you've learned nothing solid.
Asking for sources isn't enough on its own either: it can cite an article, a book or a page that doesn't exist, or that doesn't say what it claims. A source is only worth something if you open it.
So the checking happens outside the conversation. The assistant can help you prepare, but the final word belongs to a source you've looked at yourself.
Step 1: split the answer into claims
An AI answer usually mixes three things: general advice, explanations, and checkable claims. Those last ones are what you're after.
A checkable claim is a sentence that can be true or false: "the law dates from 2019," "this medication shouldn't be taken with alcohol," "the author of this novel was born in Boston," "you have 14 days to cancel."
Take the answer and list those claims, one per line. You can even get the assistant to help:
List as bullet points every factual claim in your previous answer (numbers, dates, names, rules, quotes), without commenting on them.
This has an immediate effect: a smooth paragraph that sounded solid turns into a list of points, each of which has to stand on its own.
Step 2: prioritize
You don't need to check everything. Sort the claims with two questions: how likely is this to be wrong? and what does it cost me if it is?
Given how assistants work, the riskiest ones are:
- numbers: amounts, percentages, statistics, quantities;
- dates: deadlines, years, time limits;
- names: people, places, organizations, book titles;
- quotes: a sentence attributed to someone;
- sources: references, links, studies it mentions;
- rules: laws, procedures, the terms of a contract or a benefit.
Then weigh the stakes. A wrong date in a dinner-party anecdote is harmless. A wrong deadline on an official form, a medication dose, or a figure you're about to present to your team is not. Start with what's both risky and important.
Step 3: go back to the original source
For each priority claim, look for the original source: where the information actually comes from, not a site repeating it.
- An administrative rule: the official website of the agency in charge.
- A law: the text of the law itself, on an official site.
- A figure: the body that produces it (statistics office, a company's annual report).
- A quote: the original book, speech or interview.
- Health information: your doctor, your pharmacist, or an official public health website.
If the assistant gave a source, open it and check two things: that it exists, and that it really says what's claimed. The classic trap is a real source that doesn't contain the quoted information.
Step 4: cross-check two independent sources
When there's no obvious official source, find two independent sources saying the same thing. Independent means they don't copy each other: ten sites rewriting the same article count as one.
Check the date of the source, too. Something true three years ago may be outdated today, especially prices, rules, offers and news. Assistants have a knowledge cutoff: what they learned stops somewhere.
Step 5: decide, claim by claim
For each claim you checked, pick one of three outcomes:
- Confirmed: the original source says the same thing. Keep it.
- Fix: the idea is right but a detail is wrong (it's 30 days, not 14). Correct it using the source.
- Delete: you can't find a source, or the source says the opposite. Remove it, even if the sentence was nicely written.
That last one is the hardest: we tend to keep what we failed to disprove. Simple rule: no source, no publishing.
The verification grid
Copy this table into a note and fill it in for any answer that matters:
| Claim | Type | Stakes | Source checked | Decision |
|---|---|---|---|---|
| "14-day deadline" | date / time limit | high | official site | fix: 30 days |
| "Founded in 1998" | date | low | two independent sites | confirmed |
| "According to a study by..." | source | medium | not found | delete |
With practice you won't need the table anymore: you'll split and sort in your head as you read.
What the assistant can do to help
An assistant isn't a reliable judge of its own answers, but it's still useful for organizing the work:
- listing the factual claims in a text (step 1);
- telling you what kind of source to check for each point;
- rephrasing a question for your search;
- flagging the parts it's least certain about, if you ask explicitly.
For each claim in your answer, tell me what kind of official source would let me check it, and flag the ones you're most likely to have wrong.
Keep in mind that this self-assessment is generated too: it points you in a direction, it doesn't replace the source.
Mini exercise: fact-check an answer in 15 minutes
- Ask a factual question (2 min) about something you know a bit: your town's history, a rule of your sport, something from your job. Ask for an answer with at least five specific facts.
- Split (3 min): write each checkable claim on its own line.
- Sort (2 min): highlight the three riskiest (numbers, dates, names, quotes).
- Check (6 min): for those three, find the original source or two independent sources.
- Decide (2 min): confirmed, fix or delete, and count them up.
What to notice: how many claims held up, which ones were wrong or approximate, and whether the tone of the answer let you guess which. Usually it doesn't, and that's the whole point of the method. To keep going, play a round of Hallucination Hunt.
Going further
The free game Hallucination Hunt trains exactly this reflex: spotting the number or name that's off in an answer that looks perfect. To understand where the problem comes from, read why ChatGPT makes things up, and if a term trips you up, the glossary is there.
At work, the checklist in our guide to using ChatGPT at work keeps you from sending out a mistake. And to lay the foundations (what a model generates, what it actually knows), play the free chapter "Light the Lantern" in your browser, no account needed: play the chapter.
Frequently asked questions
How can I tell if a ChatGPT answer is true?
Split the answer into checkable claims, check numbers, dates, names, quotes and sources first, then compare each one with the original source or with two independent sources. A confident tone says nothing about whether the answer is right.
Is asking the AI "are you sure?" enough?
No. The assistant generates a new plausible answer to your question: it may stick to a mistake, or change something that was right because you seem doubtful. Checking has to be done against a source you look at yourself.
Do I need to check everything in an AI answer?
No, that would be exhausting. Check what's both risky and important: numbers you'll repeat, deadlines, official rules, health, money or legal advice. For a recipe idea or a rough draft, a quick read is enough.
What if I can't find any source for a piece of information?
Remove it, or clearly label it as unverified. Information you can't find isn't necessarily false, but you can't rely on it. No source, no publishing.