Why AI makes things up: hallucinations explained without jargon

Short answer: An AI chatbot does not look facts up in a database. It predicts the most likely next words based on patterns it learned. Most of the time those patterns produce correct answers, but when the model does not really know, it often fills the gap with something that sounds right. That confident filler is called a hallucination. It is not lying, because there is no intent, but the effect on you can be the same.

How a chatbot actually produces an answer

A large language model is trained on huge amounts of text and learns which words tend to follow which. When you ask a question, it writes the answer one small piece at a time, each time choosing a likely continuation.

That works remarkably well for things that appear often and consistently in its training data: how photosynthesis works, how to structure a cover letter, the plot of a famous novel. It works much worse for:

• rare facts (a small town's history, an obscure person's birthday)
• precise details (exact numbers, dates, quotes, page numbers)
• anything after its training data ends, unless it can search the web
• things that do not exist, such as a book you made up, which it may happily summarize

In those cases there is no strong pattern to follow, but the model is still built to produce fluent text. So it does.

Why it does not just say "I don't know"

In September 2025, researchers at OpenAI published a paper arguing that the way AI models are trained and tested actually rewards guessing. Most benchmarks score an answer as right or wrong; "I don't know" earns zero, the same as a wrong answer. If a model guesses someone's birthday, it has roughly a 1 in 365 chance of being right. Over thousands of test questions, a model that always guesses looks better on the scoreboard than a careful one that admits uncertainty.

The researchers propose changing how models are evaluated, so that confident errors cost more than honest uncertainty. Newer models hallucinate less, especially when they take time to reason or can search the web, but no chatbot is free of it.

Why it sounds so convincing

A hallucination is written in the same calm, fluent, confident style as a correct answer. There is no change in tone, no hesitation, no "um". Humans use confidence as a signal of accuracy, and chatbots do not give us that signal.

Hallucinations also tend to be plausible. A made-up study will have a realistic title, a believable journal and authors with ordinary names. A fake legal case will be formatted exactly like a real one.

When it goes wrong in real life

• Fake court cases. In 2023, lawyers in a New York federal case against the airline Avianca filed a brief citing court decisions that ChatGPT had invented, complete with fake quotes. The judge fined them $5,000.
• A chatbot's promise a company had to keep. In 2024, a Canadian tribunal ruled that Air Canada was responsible for its website chatbot wrongly telling a grieving customer he could claim a bereavement discount after his flight. The airline argued the chatbot was responsible for its own words; the tribunal disagreed and ordered a refund.

The pattern in both cases: a person trusted an answer that looked authoritative and did not check it.

How to catch hallucinations

1. Be most careful with specifics. Names, numbers, dates, quotes, citations and links are where errors hide.
2. Ask for sources, then open them. If the chatbot gives a link or a study, check that it exists and actually says what the chatbot claims.
3. Ask "How sure are you?" It is not foolproof, but models will often flag uncertainty when asked directly.
4. Use search-connected modes for current facts. Answers grounded in web results with visible sources are easier to verify.
5. Cross-check important answers. Ask a second chatbot or a fresh chat, and compare. Disagreement is a red flag.
6. Never rely on AI alone for high-stakes decisions. Medical, legal, financial and safety questions deserve a qualified human or an official source.

Is it lying?

Not in the human sense. Lying means knowing the truth and saying something else on purpose. A chatbot has no intention and no internal list of true facts to betray. It is closer to someone who has read a lot, remembers patterns rather than details, and never likes to leave a question unanswered. That is useful company for brainstorming and explaining, and risky company for facts you cannot afford to get wrong.

Based on published research and court records as of October 2026.