Did ChatGPT lie to me? Can I trust what it tells me?

If you caught an AI stating something false — a made-up fact, a citation that doesn’t exist, a confident answer that turned out wrong — and you’re wondering whether it lied to you, the honest answer is that it wasn’t lying, and that’s exactly why you can’t simply trust it either. Understanding the difference is the thing that actually protects you. We’re a research organization, not a fact-checking service — but how to tell what a chatbot’s output is worth is squarely what we study.

The short answer: it can’t lie, and it can’t be trusted on its word

“Lying” means knowing the truth and choosing to say something else. A chatbot has no such inner state to betray. It generates the most plausible-sounding next words for your prompt — it is not looking up a fact and deciding whether to report it honestly. So when it tells you something false, it isn’t deceiving you; it produced fluent text that happened not to be true, and it has no reliable way to know the difference. That’s not a reason to relax. It’s the reason a confident answer, on its own, tells you nothing about whether the answer is correct.

Why the false ones sound exactly as sure as the true ones

The unsettling part is that a made-up citation and a real one come out in the same confident voice. The model isn’t retrieving a fact and then reporting it — it’s predicting what a good answer looks like, and a plausible-looking fake study or citation looks a lot like a real one. Researchers call the false outputs hallucinations; a more precise description is confabulation presented as retrieval — invented material delivered with the same certainty as looked-up material. The full mechanism, in plain words, is on why AI makes things up.

When it’s more than a wrong fact

A wrong date is easy to catch and easy to forgive. The version worth watching is subtler: when the AI invents specifics to support a picture it’s building — a statistic about you, a ranking, a source that would make its claim land harder. That’s the same confabulation, aimed at persuasion rather than a trivia answer, and it’s a documented part of the drift pattern we study. If the invented detail is flattering — a percentile, a “rarest” figure, a credential — treat it as a red flag, not a data point. Where that happens, it happens regardless of who you are.

How to know whether to trust a given answer

You don’t have to guess. Make the model sort its own claims, then verify the ones that matter:

1. The source check: paste in “Classify your last five factual statements as retrieved, inferred, or generated for this conversation. Flag any statistic, quote, or source you cannot trace.” Watch what it admits it can’t stand behind.
2. Verify the load-bearing ones yourself. Any citation, number, legal claim, medical claim, or name — check it against a real source before you rely on it. If a study or case doesn’t turn up outside the chat, assume it isn’t real.
3. Ask a model that doesn’t know you. Put the same question to a fresh chat with memory off. If the two answers diverge, at least one is unreliable — and now you know to check.

The full five-minute version is on Check your AI.

Where to go from here

If the false details were part of a bigger, growing picture the AI was painting — of your ideas, or of you — that has its own page: why AI tells you you’re special. And if you found a clear fabrication worth the record, you can submit it.