How to fact-check an AI answer in two minutes
A verification method you will actually use: find the sentence you would act on, check it by claim type, and open the citation instead of trusting it.
The standard advice about AI answers is to verify them against a reliable source. That advice leaves several practical questions unanswered. It does not tell you which claim in the answer to check, or which source counts as reliable for that particular kind of claim, or how long you should spend before you either use the answer or throw it away. Without a clear procedure, it is easy to skip the check, which is why what most people actually do is skim, decide the answer reads as though someone competent wrote it, and paste it into an email.
A two-minute budget is what turns verification into something you do on a Tuesday afternoon with a colleague waiting on your reply. It is short enough to survive contact with a real workload and long enough to catch the failure that costs you: a confident, well-formed sentence resting on something the model assembled rather than retrieved. Almost nothing in a typical answer needs checking, because one or two sentences carry the weight and the rest is scaffolding. What follows is a method sized to that budget, and then a straight account of the claims it cannot handle, since a method that pretends to cover everything gets abandoned the first time it quietly fails.
The one sentence that actually needs checking
Read the answer once and ask a single question of it: if this were false, what would go wrong? That question sorts an answer faster than any amount of suspicious re-reading, because it ignores the writing and goes straight to consequence. The sentence you would forward, quote, spend money on or repeat in a meeting is the one to check. Everything around it can be wrong at no cost to you.
You cannot triage by reading, because fluency is uniform across the whole answer. Sentences built from solid, heavily repeated material and sentences the model improvised arrive in the same register, with the same air of having been looked up, and nothing in the prose marks the join. The instinct to trust the passages that sound most authoritative is backwards on the mechanism involved. For why models produce confident invented material at all, we cover that in AI hallucinations explained; over the next two minutes, take it as given.
What you can use, weakly, is shape. Risk climbs with precision and with recency. A famous round number in a well-covered field is usually fine. An oddly specific figure about a small subject, a person's current job title, a price, a version number, a statutory deadline, a quotation attributed to someone: invention lives there, because those facts are too particular to be everywhere in the training material and too recent to have settled.
Then write the claim down as a proposition before you go looking. "It said the deadline is 31 January" gives you something to resolve; "it said something about the filing deadline" gives you a topic to read around, which is how a two-minute check turns into ten. Pinning it has a second effect worth having: some claims fall apart at this step, because you cannot state them precisely enough to check, and a claim too vague to check is too vague to use.
Different claims, different two-minute checks
Here is the real reason "verify with a reliable source" fails as instruction. The reliable source is a different thing for every kind of claim, the fastest route to it is different, and for several kinds there is no route at all inside two minutes. A date and a dosage do not verify the same way, and collapsing them into one category is what produces advice everybody ignores.
Claim type | Fastest check inside two minutes | Can a second AI settle it? |
|---|---|---|
A date or a matter of record: when something happened, a version number, an official figure | Go to the body that publishes it, not to a summary of it. Search that page for the figure | No. Two models can hold the same outdated fact with equal confidence |
A statistic with a study behind it, typically carrying a year and a percentage | Find the study itself. Check the year, the population and what was measured, all three | No. The number often survives the trip while the caveat does not |
A quotation | Search the exact words inside quotation marks. If only aggregators carry it, treat it as unattributed | No. Plausible paraphrase is precisely what a model produces |
A legal or regulatory claim: your rights, a deadline, a rule | Establish the jurisdiction first, then read the responsible body's own page. Two minutes gets you to the page, rarely to an answer | Never. Jurisdiction and effective dates are where this fails |
A medical or dosage claim | Check the general fact against a health authority. The application to you is not checkable this way | Never. The dangerous error is the applicability, not the fact |
A product capability: "app X can do Y" | Open the vendor's current documentation, or open the app and try it. Docs go stale, the app does not | No. Training data captures the product as it was |
A piece of code | Run it somewhere disposable. A function that does not exist announces itself immediately | Partly. A second model reviews; only running it verifies |
A recommendation or judgement | Nothing here is checkable. Check the assumptions it rests on instead | No, and the question is malformed for it |
The third column shows the problem with the check most people reach for first. Asking another model is fast, feels rigorous, and is close to useless for anything factual, because models built on heavily overlapping text fail in correlated ways. A second system restating a wrong date with total composure hands you the same error twice, and two copies of one mistake read exactly like agreement. Where a second model does earn its place is in surfacing that a question has more than one defensible answer, which is a separate service worth reading about on its own terms.
Open the link, then search inside it
Citations are where this method earns its keep, and where the standard advice does the most damage, because "check the source" gets read as "check that there is a source". There are three ways a citation fails and they are not equally visible.
The first is a URL that does not resolve. Plausible domain, plausible path, plausible title, and a 404 at the end of it. This is the least dangerous failure in the set, because it announces itself the moment you click.
The second is worse, and it is the one to build a habit around. The link resolves. Real page, real organisation, roughly the right subject, and nowhere on it the claim it was cited for. A reader who clicks, watches an official-looking page load, and comes back satisfied has verified exactly one thing: that a page exists. The citation survived the click, which was the only test it was ever going to face, and the claim is now wearing a source it never had.
The third is that failure wearing better clothes. The page carries something adjacent: the right topic in a different year, the right rule in a different country, the right statistic for a different population. That is how a check produces false confidence instead of no confidence. If your claim has a number in it and the page has a number in it, they need to be the same number.
The fix costs about fifteen seconds and it is mechanical. Open the page, then use your browser's find function (Ctrl-F, or Cmd-F on a Mac) and search for the most distinctive token in the claim: the figure, the surname, the year, the section number. Search for the token, never the whole sentence, because sources paraphrase and your search will fail on wording while the fact sits right there. If find comes back empty, the citation does not support the claim, no matter how authoritative the page looks. Then glance at two more things: the date on the page, and whether you are looking at the source or at a report of the source. A news article about a study is not the study, and the abstract usually carries the population and the qualification that the news write-up dropped.
One sequencing point matters more than it sounds. Turn search on before you ask, instead of asking a model to source a claim it has already made. The second request asks it to defend a position it is committed to, and what comes back tends to be pages that look like support.
What our app hands you, and what it leaves to you
MultiChats can run the search for you, and it is worth knowing exactly what that gets you, since what arrives is raw material to check and no judgement about it. Turn the search toggle on and the answer is built from pages fetched during the request. On the web that toggle arrives once you switch Intent Detection off in Settings, under Tool Preferences; leave it on, as every account does at first, and the app decides per message whether to search. Underneath sits a row saying how many websites were fetched; open it and each source is listed with its title and domain, and each one opens in a new tab. Where the model cites inline, the marker links to the page it came from. Four search back ends sit behind that toggle (OpenAI Web Search, Google Search, Anthropic Web Search and Perplexity Sonar), and the line printed under the answer names the one that ran, alongside the model. That line is also how you answer the first question of any citation check, which is whether a search happened at all. No sources row means no search, and any URL in the text came from memory.
On free accounts, web search runs two searches per 30 days on the web app, and on mobile the toggle sits behind Pro and raises an upgrade prompt on a free tap. Two searches is a budget for the claim that matters, which is another argument for finding the load-bearing sentence before you spend anything.
Now the part product copy leaves out. Nothing in the app checks a citation for you. There is no badge for a dead link and no flag on a claim the fetched page fails to support. Opening the link and searching it is a human step and there is no version of this where it stops being one. What the app removes is the friction around that step, which is a smaller claim than "verified" and a true one.
Where two minutes runs out
Some claims cannot be checked in two minutes, and a method that will not say so is how careless advice gets laundered into confident advice. Four kinds resist the budget. Anything legal, where the answer turns on jurisdiction and on which version of a rule was in force on a given date. Anything medical, where the general fact can be perfectly accurate and the application to a particular person is the whole question. Contested empirical territory, where the authorities disagree and no amount of clicking resolves it. And private matters with no public record, where you can search all afternoon and find only other people guessing.
When a claim fails the check, you have three moves available, and using it anyway is not among them. Give it a longer look with a real search instead of a fast one. Take it to somebody whose profession it is, which for money, health and law is the right move often enough that it should arrive well before the last resort. Or drop the claim. That last one is badly underrated: most sentences survive the removal of their most impressive number, and a paragraph that says "filing deadlines vary, check with the tax office" is worth more than one that names 31 January and is wrong.
Be precise about what an unverifiable claim is. It may well be true. What it cannot be is used, at least not in your name and not in writing somebody else will act on. That distinction, between false and unusable, keeps the method honest, because it lets you stop looking without pretending you settled anything.
Two minutes, before you send
The method is short enough to hold in your head. Find the sentence you would act on. Say what it claims, precisely. Work out what kind of claim it is, since that decides where the check lives. Go to the thing itself and search the page for the specific token, instead of admiring the fact that a page loaded. If it does not check out inside the budget, escalate it or cut it.
Run it before you send. Running it after somebody has questioned you is damage control, and by then the claim is already out. What you end up with is a smaller set of sentences you are prepared to defend, plus a clear view of how much of the answer you were about to pass on under somebody else's authority without ever checking whose.