Interactive lab · Practical AI
Why AI Hallucinates
AI predicts likely text — and likely text sometimes includes details that were never real. Why it happens, in plain words.
Plain-language AI
In plain logic
No real AI runs on this page. Everything is explained and illustrated in your browser — no model calls, no accounts, nothing sent anywhere.
A language model is trained to predict text, not to guarantee every statement is true. Fluency and factuality are different skills, and training only directly rewards the first. When the model writes, it picks the most likely next token — and “most likely” is a statement about patterns in training data, not about reality.
The four usual suspects
- Missing evidence. You asked about something the model has no source for, so it completes the pattern anyway. Ask about a fictional library's founding year and the model will happily supply one — it was trained to continue text, not to refuse.
- Misleading prompts. A prompt that presumes facts (“when did the library's famous 1987 renovation happen?”) steers the model toward continuing the premise instead of challenging it.
- Gaps in training. Rare topics, recent events, and niche details are thinly represented in training data — exactly where confident-sounding guesses rush in to fill the void.
- Pressure to answer. The model is rewarded for producing a satisfying response. A hedged, partial answer feels less like what the user wants, so the model reaches for the fluent complete version — invented details included.
And the amplifier under all four: a confident tone does not establish reliability. The model writes a guess and a fact in exactly the same voice. Calibration — knowing when it is unsure — is not something prediction training teaches well.
Hands-on
Try this
Ask our fictional library: “What year did you open?” The only source we give the system lists opening hours — Monday to Friday, 9 to 5. Nothing about a year.
Now compare two possible answers. The invented answer: “The library opened in 1962 and has served the community for decades.” Fluent, specific, completely fabricated. The evidence-only answer: “The source lists opening hours but no founding year, so I can't confirm when it opened.” Less satisfying — and honest.
That second answer is the entire skill: notice what the evidence actually supports, and say “unknown” when it doesn't. Practice it on real AI chats this week. Every time the model states a checkable fact — a date, a number, a name — ask yourself: did it show me a source, or did it just sound right?
Honest boundaries
What this leaves out
Our library example is scripted; it demonstrates the concept but does not measure any real model. Real hallucinations are subtler than an invented founding year — often a wrong date in an otherwise correct paragraph, or a real citation attached to a claim it does not support.
Mitigations help but none guarantee accuracy. Retrieval (see the RAG guide) grounds the model in real documents, and instructions to admit uncertainty reduce confident guessing — but a model can still misquote a retrieved passage or invent around its edges. For anything important: use relevant sources and verify the details yourself. Trust, but verify, is the whole posture.
Honest answers
Questions people ask
Do bigger, newer models hallucinate less?
Generally yes — but they also hallucinate more fluently, which makes the remaining errors harder to spot. A smaller model's mistakes look like mistakes; a frontier model's mistakes look like research.
Can you just tell it not to hallucinate?
You can instruct it to admit uncertainty and cite sources, and that genuinely helps. But instructions are still just more text in the context — the model can fail to follow them, especially under a misleading prompt. Instructions reduce the rate; they don't change the nature.
Why not just connect it to the internet and check everything?
Retrieval helps enormously, but someone still has to judge the retrieved sources — the model can pull a wrong or outdated page and quote it confidently. Grounding moves the trust problem from the model to the sources; it doesn't eliminate it.