What are AI hallucinations? Explained with a confident quiz guess

In one sentence: an AI hallucination is when an AI tool gives an answer that sounds confident and correct but is false or made up, because it was built to produce a likely-sounding answer rather than to check facts.

The everyday picture

Picture a quiz night. A team gets a question they do not really know: "In what year did this small museum open?" Leaving the answer blank scores nothing. A guess might score a point. So the team writes down a year that sounds about right, in neat handwriting, with no question mark.

The answer sheet looks just as sure as when they truly knew. The quizmaster cannot tell the difference from the handwriting. Only checking the real answer shows the guess was wrong.

An AI model is that quiz team. When it does not know, it often writes down a smooth, likely-sounding answer anyway. That confident wrong answer is a hallucination.

How it works, step by step

1. You ask a question (the quizmaster reads it out). You type a prompt, such as "Which court case set this rule?" or "What is this supplier's phone number?"
2. The model draws on patterns it learned (the team searches its memory). A large language model learned by predicting the next word across huge amounts of text. OpenAI explains that this text carries no "true" or "false" labels. The model learns what sounds right, not what is right.
3. It hits a gap (nobody on the team knows). Common facts appear many times in that text, so the model usually gets them right. Rare, one-off facts, like a particular person's birthday, cannot be worked out from patterns. Those are where errors pile up.
4. It guesses with confidence (the team writes a year anyway). OpenAI's researchers found that most tests used to score AI models reward a right answer and give zero for "I don't know". Like a quiz with no penalty for wrong answers, that teaches models to guess.
5. Someone checks the answer (the quizmaster marks the sheet). The guess reads just as fluently as a real fact. Only checking against a trusted source, or asking the model to point to one, catches it.

Where the picture stops

• The model does not know it is guessing. A quiz team usually knows when it is bluffing. An AI model has no reliable inner sense of which answers are facts and which are guesses, although newer models are getting better at saying when they are unsure.
• It is not lying. The team chooses to bluff to win. The model is not trying to fool anyone. It produces the most likely-sounding words, and sometimes those are wrong.
• It can invent whole things, not just one answer. A model can make up a quote, a source, a link or a step in its reasoning, and wrap it in convincing detail.

Why it matters for your business

• Risk to customers and reputation. A chatbot that invents a refund rule or a product feature can mislead a customer, and the customer will hold your business responsible.
• It will not go to zero. OpenAI says accuracy will never reach 100%, because some questions cannot be answered from what a model knows. Anthropic's guidance says its techniques reduce hallucinations but do not eliminate them, and that critical information should always be checked.
• "I don't know" is worth more than it looks. In OpenAI's own test, an older model that almost never declined was wrong on 75% of questions. A newer model that declined about half the time was wrong on 26%, with similar accuracy. Fewer bluffs means less cleaning up.
• Simple habits cut the risk. Anthropic recommends telling the model it may say "I don't know", asking it to answer only from documents you provide, and asking it to quote or cite its source for each claim.

An example

Imagine a 20-person accounting firm that uses an AI assistant to draft answers to client questions about deadlines.

At first, the assistant answers from memory. One draft gives a filing date that sounds right but is a year out of date. A partner spots it before it is sent. The firm changes the setup: the assistant now answers only from the firm's own up-to-date guidance notes, quotes the line it used, and says "I can't find this in our notes" when the answer is not there. Staff still check every draft, but now they can see where each answer came from.

Frequently asked questions

Why does AI make things up?

AI language models learn to predict likely-sounding words, not to check facts. When they hit a gap, they often fill it with a plausible guess, partly because the tests used to rate them have rewarded guessing over admitting uncertainty.

Can AI hallucinations be fully stopped?

No. Both OpenAI and Anthropic say hallucinations can be reduced but not removed entirely. Checking important answers against trusted sources is still necessary.

How can I tell if an AI answer is a hallucination?

You often cannot tell from the wording, because guesses sound as confident as facts. Ask the AI for its source, check names, numbers, quotes and links yourself, and be most careful with rare or very specific details.

Does giving the AI my own documents help?

Yes. Asking it to answer only from documents you provide, and to quote the part it used, makes wrong answers rarer and easier to spot. It does not remove the need to check.

Related explainers

• What is an LLM? Explained with a new hire who read the whole library
• What is RAG? Explained with an open-book exam
• What is human-in-the-loop? Explained with a manager signing cheques

The DevDrift builds AI automation with approvals and audit trails: see how.

Sources

• OpenAI: Why language models hallucinate (September 2025)
• Claude Platform docs (Anthropic): Reduce hallucinations