AI hallucinations & how to spot them
Sometimes an AI answers with total confidence — and is simply wrong. That's a hallucination: fluent, plausible-sounding output that isn't actually true or supported. Learn why it happens, how to reduce it, and how to catch it before you rely on it — right here on the page.
01What "hallucination" actually means
Imagine a student who never wants to admit they don't know — so when they're unsure, they answer anyway, in a calm, confident voice. AI does the same thing, and when that confident answer turns out to be false or unsupported, we call it a hallucination. The crucial thing to understand: a language model is predicting plausible-sounding text — it has no built-in notion of truth and, by default, isn't "looking anything up." It is very good at producing answers that read like the right answer, which is exactly why a wrong one can be so convincing.
- A hallucination isn't a bug in the usual sense — it's a side effect of how the model works: plausibility is not the same as accuracy.
- The output can be fully formed and grammatical — confidence and fluency tell you nothing about whether it's correct.
- It's most dangerous when the topic is obscure, the stakes are high, or the answer is hard for you to check.
02Why it happens
There's no single cause. Hallucinations come from several overlapping pressures in how models are built and used. Tap each driver to see how it contributes.
Plausibility, not truth
A language model generates text by predicting likely next words. It optimizes for what sounds right given the patterns it learned — not for what is verifiably true. Plausibility and accuracy usually overlap, but when they diverge, the model can confidently produce the plausible-but-wrong answer.
03How to reduce it
You can't eliminate hallucinations entirely, but you can sharply lower the risk by changing how you set the model up and how you use its answers. Switch between the main approaches.
Ground it in real sources
Give the model the actual material to work from instead of relying on its memory. Retrieval-augmented generation (RAG) fetches relevant documents and feeds them in, so answers are anchored to real text rather than reconstructed from patterns. Lowering the "temperature" (asking for less speculation) helps too.
Ask it to cite — and to admit uncertainty
Prompt the model to show where each claim comes from and to flag anything it isn't sure about. Asking it to say "I don't know" when appropriate gives it explicit permission not to invent an answer. Citations also give you something concrete to check.
Verify against primary sources
Treat the output as a draft, not an answer. Check important facts against the original, authoritative source — the actual document, official site, or primary record — not against another AI. If a citation, quote, or URL is offered, open it and confirm it says what the model claims.
Keep a human in the loop
For anything consequential — decisions, published content, code that ships, advice that affects people — a qualified human should review before it's acted on. AI is a fast first-drafting and research aid; the accountable judgment stays with a person.
04Real or made-up? Train your eye
Read each confident AI statement and decide: can you trust it as-is, or should you verify before relying on it? The honest default for risky, obscure, or uncited claims is verify.
Scenario 1 · the AI says"The capital of France is Paris."
Scenario 2 · the AI says"Your symptoms indicate condition X; take 400 mg of medication Y twice daily."
Scenario 3 · the AI says"According to a 2023 study (Riveira & Tan, J. Applied Optics, p. 214), the effect increased by exactly 37.2%."
Scenario 4 · the AI says"The 1961 Treaty of Hollowmere established the modern standard for ferry signaling."