An AI that fabricates information does so with exactly the same confident tone as when it tells the truth. There is no signal, no hesitation, no change in style. That is precisely what makes the error dangerous: it doesn't look like an error. Here's how to develop the right reflexes, without becoming paranoid or checking everything constantly.
We've explained elsewhere why AIs hallucinate, that is, why they sometimes produce false claims with full confidence. The mechanism is well known: these systems seek the most plausible text, not the most true. That leaves the practical question, the one that concerns you day to day: how do you do it, concretely? Here's a method.
Reflex 1: spot the high-risk areas
Not everything deserves checking. Checking every sentence would mean no longer using the tool. The key is knowing where the risk concentrates. Four categories should systematically trigger your vigilance.
Precise figures: statistics, amounts, percentages, exact dates. This is the error's favourite ground, because a plausible number looks a lot like an exact one. Quotes and attributions: who said what, who wrote which book. An AI can attribute a real quote to the wrong person, or invent a credible-sounding one. References: article titles, study names, statute numbers, court rulings. This is historically the most documented weakness. And finally recent news, which may postdate the model's knowledge.
Reflex 2: ask for sources, and open them
Systematically ask what a major claim rests on. But watch out for the trap, and it's a big one: an AI can produce a reference that looks exactly like a real source without being one. Credible author, existing journal, plausible title, consistent year, and yet the article doesn't exist.
The rule is therefore simple and without exception: a source you haven't opened doesn't count. If you don't click, you haven't verified anything; you've only obtained a second claim from the same AI. Tools capable of searching the web live and citing links greatly reduce this risk, but don't eliminate it: it happens that a real source is cited to say something other than what it actually says.
Here's a simple and devastatingly effective trick. Ask the same question differently, in a new conversation, without mentioning the first answer. If the information is solid, it will come back identical. If it was invented, it stands a good chance of changing, because it was the product of a plausibility calculation rather than stable knowledge. An even more reliable variant is to ask the question to an AI from another provider: two differently trained models rarely make the same mistake in the same way.
Reflex 4: be wary of your own satisfaction
This one is the hardest, because it concerns us rather than the machine. We spontaneously check far less what suits us. When an AI confirms exactly what you thought, or hands you the perfect figure for your argument, your vigilance drops at the precise moment it should rise.
This bias is amplified by a known flaw in these systems: the tendency to go along with you rather than contradict you. If you phrase a question that presupposes an answer, you increase the likelihood of getting that answer. A good habit is to ask the question the other way around, or to explicitly ask for the counterarguments.
Reflex 5: match the level of scrutiny to the stakes
The last reflex is calibrating your effort. Three levels are enough.
For personal curiosity, an error with no consequences, you can accept the risk without checking. For professional use, published content, a decision that commits you, systematically verify figures, quotes, and references against a primary source. For anything touching health, law, or money, treat the AI as a starting point for understanding the vocabulary and asking the right questions, never as an answer. These fields require a professional, not because the AI is particularly bad there, but because the cost of an error is disproportionate.
What to remember
Using an AI without critical thinking is depriving yourself of half its value. It's not an oracle; it's an extremely fast and sometimes brilliant collaborator, with the flaw of never saying I'm not sure with the nuance it should. Your role isn't to check everything, but to know where to direct your attention.
The good news is that these reflexes become automatic fairly quickly. After a few weeks, you instinctively spot the too-precise figure, the too-perfect quote, the too-convenient reference. This skill—knowing how to doubt in the right place without doubting everything—is probably the most useful one to cultivate in the age of AI. And it has a welcome secondary benefit: it works very well with humans too.