How to fact check an AI answer before you rely on it
You don't need to check everything an AI assistant tells you. Sort the answer by what it would cost you to be wrong, check only the claims that carry that cost, and check each one against the original source, never by asking the same assistant again. A name in a recipe can wait. A dose, a deadline or a quote you're about to repeat can't.
By Meagan Victoria Angelucci, Eunosa’s founderLast updated About 7 minutes to read
On this page
Which parts of an AI answer are worth checking?
The parts you're about to act on, repeat or pay for. An AI answer is a mix of claims, and they don't all carry the same risk. A wrong detail in a trivia answer costs you nothing. A wrong date on a tax form, a wrong drug interaction or a quote you put in front of your boss costs you a lot.
So before you check anything, ask one question of each claim: what happens if this is wrong and I rely on it? If the honest answer is "nothing much", move on. If it's "I lose money, time, health or trust", that claim gets checked.
This matters because AI assistants make things up in a fluent, confident voice. Researchers call this hallucination, and a survey of hallucination in language generation (arXiv:2202.03629) describes it as text that reads well but isn't supported by any source. The answer gives you no signal about which sentence is the made up one. You have to decide where to look.
The triage table
Find the kind of claim in the left column, decide the stakes, and run the check in that cell.
| Kind of claim | Low stakescuriosity, small talk | Medium stakeswork, school, a purchase | High stakeshealth, money, law, safety |
|---|---|---|---|
| Nameperson, product, place, law | Skip | Search the exact name and confirm it exists | Find it on the official source: the organization's own site or a government register |
| Numberprice, amount, statistic | Skip | Find the number on a page you opened yourself | Find it at the original source, read the date beside it, and confirm the units |
| Date or deadline | Skip | Confirm on one reliable page | Confirm on the issuing body's own page, for the current year |
| Quote | Don't repeat it word for word | Find the exact words in the original speech, book or article | Don't use it until you've read the original in context |
| Citation or link | Skip | Open the link and find the sentence that backs the claim | Open it, read the passage, and confirm the paper or page says what the answer says |
| Advicewhat you should do | Treat it as one opinion | Compare it with a second source written by a person who knows the field | Ask a qualified professional before you act |
The table is the whole method. The rest of this page explains how to run each cell.
How do you split an answer into claims you can check?
Go through the answer one sentence at a time and pull out every statement that could be true or false. A paragraph that says "The deadline is 30 April, and late filers pay a penalty based on what they owe" holds two claims: a date, and a rule about penalties. Check them separately, because one can be right while the other is wrong.
Leave out opinions, summaries and advice phrased as "you might consider". Those aren't facts to check. They're suggestions to weigh.
Then mark each claim with its stakes from the table. You'll often find that only one or two claims in a long answer actually need a check.
Where do you check a name, a number, a quote or a date?
At the place the fact comes from, not at a page that repeats it.
Names. Search the exact spelling. For a person, look for their own page or their organization's staff page. For a law or a program, go to the government site that runs it. If the name only appears in the AI answer and nowhere else, treat it as invented.
Numbers. Find the original report, survey or official table. Read the date it was published and what exactly it measured. A figure can be real and still wrong for your question, because it's from another year, another country or another group of people.
Quotes. Find the original speech, interview, book or article and look for the exact words. AI tools often blend a real person with words they never said, or tidy a real quote into something smoother. If you can't find the words, don't put them in quotation marks.
Dates and deadlines. Go to the body that sets them: the tax agency, the school, the court, the company. Deadlines change from year to year, and an assistant may be repeating last year's.
Citations. Open every link and every reference. A real looking reference is not proof. In one test of literature reviews written by ChatGPT, 55% of the references from the older model and 18% from the newer one did not exist (Walters and Wilder, Scientific Reports, DOI 10.1038/s41598-023-41032-5). And a link that does exist can still point at a page that never makes the claim. The test is simple: can you point to a sentence on a page you opened yourself that says what the answer says? If you can't, the claim isn't checked yet.
Why doesn't asking the AI "are you sure?" count as a check?
Because you're asking the same source to grade its own work. When you type "are you sure?", the assistant has no new information. It looks at its own words again and either repeats them more firmly or changes them to please you. Neither tells you which version is true.
Researchers have tested this. A study presented at ICLR 2024 found that language models struggle to correct their own reasoning without outside feedback, and sometimes get worse after trying (arXiv:2310.01798). Asking a second assistant has the same weakness: two tools trained on similar text can share the same mistake, and agreement between them feels like confirmation when it isn't.
A check needs something outside the conversation: an original document, an official page, a person who knows. Asking the AI again is fine as a way to find out where to look. It isn't the look itself.
What should you do when you can't confirm a claim?
Treat it as unconfirmed and act accordingly. You have three honest options:
- Drop it. If the claim isn't essential, leave it out of whatever you're writing or deciding.
- Say it's unconfirmed. If you need to pass it on, say plainly that you couldn't verify it and where it came from.
- Ask a person. For anything in the high stakes column, a pharmacist, an accountant, a lawyer or the organization itself can answer in minutes what an assistant can only guess at.
What you shouldn't do is keep rephrasing the question until the assistant gives you an answer you like. That's how a guess turns into something that feels like a fact.
Frequently asked questions
Can I trust an AI answer if it gives me a link?
Not on the link alone. Open it and find the sentence that supports the claim. Assistants can attach a real page to a statement that page never makes, and they can invent references outright. A link is where to start checking, not proof that the checking is done.
Is it enough to check with a second AI assistant?
No. A second assistant is useful for spotting disagreement, which tells you where to look harder. But two tools can repeat the same error, and neither one has read the original source on your behalf. The check is the original source.
How long should checking an AI answer take?
As long as the stakes justify, and often less than you'd think. A low stakes answer takes no time, because you skip it. A medium stakes claim usually takes one search and one page. High stakes claims deserve the time it takes to reach the original source or a qualified person, because being wrong there is expensive.
Which topics should I always check?
Health and medication, money and taxes, legal rights and deadlines, safety instructions, and anything you're about to publish or say under your own name. Also check anything about recent events, because an assistant may not know what has changed.
Sources
- survey of hallucination in language generation (arXiv:2202.03629) Checked 7 November 2026.
- Walters and Wilder, Scientific Reports, DOI 10.1038/s41598-023-41032-5 Checked 7 November 2026.
- struggle to correct their own reasoning without outside feedback, and sometimes get worse after trying (arXiv:2310.01798) Checked 7 November 2026.
An assistant that shows its work
Checking is easier when the assistant tells you where each answer came from in the first place. We're building Eunosa to do that: it's designed to show the source behind what it says and to tell you when it isn't sure, so you know which claims to check. You can read how we built and tried the honesty check and our honesty check research.
Eunosa isn't out yet. Join the waitlist to hear when it is.