What is a Hallucination?
AI hallucination, told as a story about a parrot that continues confidently and plausibly even when it doesn't know.
A guest asks who wrote a paper, and the parrot gives a name, a year, even a title.Very confidently, very plausibly. But that paper doesn't exist.
In its books, a paper was always followed by a name and a year. So it continues in that shape.It has almost never picked the I-don't-know bean, so that one sits at the bottom. The parrot continues the shape, not the meaning.
A hallucination is the parrot continuing plausibly even when it doesn't know.The mountain-of-books parrot was born this way. Doubt names, numbers and dates most, and it errs far less when a real page the librarian found sits on its tray.
There are several ways to reduce it. The librarian's page works best; the dial only cuts randomness.A line in the note helps a little, and the short drill only fixes manners. Turning the dial down leaves confident errors in place — they aren't randomness.
It can't be removed completely — continuing is the bird's nature.So the answer is a draft and a person checks it. In front of risky jobs like money or medicine, put up a fence — the thief's parrot story lives on the security side.
Hallucination = the parrot continuing confidently and plausibly even when it doesn't know. Doubt names, numbers, dates; ask for sources; put the librarian's real page on the tray. It never fully goes away — the answer is a draft, a person checks.
A language model generates content that isn't true because it keeps producing the next token that looks plausible under its training distribution. Prompt instructions, RAG (injecting retrieved evidence), fine-tuning and low temperature reduce it but never remove it. Source citation, factuality evaluation, calibration (matching confidence to accuracy) and guardrails for high-risk tasks are used alongside.
When grown-ups say it
- Hallucination
- The plausible lie. Comes from continuing what it doesn't know. Not a mistake so much as a make-up.
- Grounding
- Speaking from real text. Put the page the librarian found on the tray, and it continues from that.
- Retrieval-augmented generation
- Librarian + parrot. For every question the librarian searches first, then the parrot answers from that page. The easiest way to add new knowledge.
- Source citation
- Saying where it saw it. An answer with a source can be checked by a person. No source means it's a draft.
- Temperature
- The randomness dial. Turn it down and it picks only the obvious bean. Less randomness, but confident errors stay.
- Factuality evaluation
- Checking answers against a key. An examiner measures how often it makes things up, using hundreds of questions with known answers.
- Calibration
- Is it right as often as it sounds sure? A well-calibrated parrot speaks softly when unsure. The plausible parrot is always loud.
- Prompt
- The trainer's note. Say so if you do not know — that one line helps a little. → the trainer's note