When a model answers a question, it isn't checking a database of truth — it's generating the most statistically plausible next words, one token at a time. That process can produce a perfectly true sentence or a completely fabricated one with exactly the same fluency, the same confident tone, and the same internal "this looks right" signal. The model has no fact-checker. It only has a sense of what sounds like it belongs. Below, try to spot the fabrications yourself — and watch the model's confidence stay high no matter which way the truth falls.
Each card shows a factual-sounding claim. Decide if you think it's true or false, then reveal the answer. Watch the "model confidence" meter — it's deliberately unhelpful: it stays high whether the claim is true or completely made up, because fluency and correctness are computed by different parts of your brain, not the model's.
A search engine looks something up. A language model fills a gap with the most likely-sounding continuation, the same way it would finish "Twinkle twinkle little ___." When the gap is a real fact it has seen often, that continuation is usually true. When the gap is obscure, the model still completes the pattern — it just makes something up that sounds like it belongs. Click each card to watch the process.
A query is matched against a verified, indexed source. If the fact isn't there, the honest answer is "not found" — there is no pressure to produce fluent text regardless.
The model predicts the next most likely token given everything before it — over and over. It always produces something fluent, whether or not that something is true, because "sounds right" is the only signal it optimizes for.
Four questions on why hallucination happens and how teams reduce it in practice. Pick an answer to reveal feedback.
Hallucination is one piece of a bigger picture — how these models actually generate text, and how good evaluation catches what confidence alone can't.
Knowing that confidence isn't correctness changes how teams should deploy AI for anything customer-facing or high-stakes. That's the kind of evaluation and guardrail work we help teams get right before they ship.
Talk to us about your AI roadmap