Explainer

Why AI hallucinates — and how to catch it before it catches you

Chatbots sometimes state things that are simply untrue — with total confidence. Here’s why it happens, where it’s most likely, and a practical routine to protect yourself.

Key points

  • A “hallucination” is when an AI produces information that sounds plausible but is false or unsupported.
  • It happens because language models generate likely text rather than look up verified facts — and training tends to reward a confident guess over “I don’t know”.
  • Risk is highest for specific details: numbers, dates, quotes, citations, names, niche topics and recent events.
  • Ask for sources, allow “I don’t know”, ground answers in documents you provide, and verify anything that matters.

What is an AI hallucination?

An AI hallucination is output that is fluent and confident but factually wrong, made up, or not supported by the sources the model was given. Classic examples include a citation to a research paper that doesn’t exist, a statistic with no real source, a quote attributed to someone who never said it, an invented feature of a product, or a summary of a document that includes a point the document never makes.

Researchers have studied the problem for years. A widely cited 2023 survey by Ji and colleagues distinguishes between output that contradicts the source material and output that adds information the source can’t verify. Both matter to everyday users: the first gets your document wrong, the second slips in “facts” from nowhere.

Why it happens

1. Models predict plausible text, not truth

A large language model works by predicting the next token based on patterns learned from training data (see What is an LLM?). It doesn’t consult a database of verified facts before answering. When the pattern for “a citation looks like this” is strong but the specific knowledge is weak, the model can produce something that has the right shape — author, year, journal — without being real.

2. Training rewards guessing

A 2025 paper by researchers including Adam Kalai, titled “Why Language Models Hallucinate”, argues that standard training and evaluation methods reward guessing over admitting uncertainty. If a test gives points for a correct answer and nothing for “I don’t know”, a model that always guesses scores better than one that abstains — so systems learn to bluff. The authors suggest changing evaluations so that confident errors are penalised more than expressions of uncertainty.

3. Rare facts are harder

Facts that appear many times in training data — the capital of France — are learned reliably. Facts that appear rarely or only once, such as the birthday of a minor public figure or the details of a small company, are much easier to get wrong. The same paper links hallucination rates to how often a fact appeared in training.

4. The question contains a false premise

Ask “Why did the Eiffel Tower move to London in 1990?” and a model trying to be helpful may invent an explanation rather than correct you. Leading questions invite hallucinations.

5. Long or messy context

When a conversation is very long or a document is large and complex, a model can mix up details, attribute a point to the wrong section, or lose track of what was said earlier. See Context windows explained.

Where hallucinations are most likely

Higher riskWhyWhat to do
Citations, references, legal casesEasy to fake the formatOpen every source; search the exact title
Numbers, statistics, pricesSpecific and often rareFind the original source; check the date
QuotesPlausible wording is easy to generateSearch the exact quote in reliable sources
Recent events and product featuresAfter the knowledge cutoffUse web search; check official pages
Niche topics and peopleLittle training dataTreat as a lead, not an answer
Medical, legal, financial specificsHigh stakes, nuanced rulesConfirm with a qualified professional

Warning signs in an answer

  • Very specific details (exact figures, page numbers, dates) that you didn’t provide and that come without a source.
  • Sources that are vague (“studies show”, “experts say”) or that you can’t find.
  • An answer that agrees enthusiastically with a premise in your question.
  • Different answers when you ask the same question again or in a new chat.
  • Confident claims about things that happened recently, without web search being used.

How to reduce hallucinations

Allow the model to say “I don’t know”

Anthropic’s documentation on reducing hallucinations recommends explicitly giving the model permission to admit uncertainty. Add: “If you don’t know or aren’t sure, say so. Don’t guess.” It sounds simple, and it measurably changes behaviour in many cases.

Ground the answer in a source you provide

Paste the document and instruct: “Answer using only the text below. If the answer isn’t in the text, say so.” Ask for direct quotes supporting each claim — another technique Anthropic recommends — so you can check them quickly. This is the everyday version of retrieval-augmented generation (RAG), a technique introduced in 2020 to connect language models with external knowledge.

Use web search for current facts

For anything that changes — prices, product features, news, laws — use a chatbot mode that searches the web and shows sources, then open those sources. Our guide to the best AI for research compares options.

Ask the question neutrally

Instead of “Why is X true?”, ask “Is X true? What’s the evidence for and against?”

Cross-check

Ask the same question in a fresh chat or a different chatbot. Agreement doesn’t prove truth — models share training data — but disagreement is a strong signal to dig deeper.

Editor’s note — add your own experience: Add one real hallucination you’ve encountered (e.g. a fake citation or wrong product feature), with a screenshot, what the correct information was, and how you caught it.
Remove this box by adding your content here, or hide all notes with "showEditorNotes": false in site.config.json.

A five-minute safety routine

  1. Decide the stakes. Brainstorming a gift idea? No checks needed. Citing a statistic in a report, taking medication, signing a contract? Verify everything.
  2. List the checkable claims — names, numbers, dates, quotes, sources.
  3. Check each one at a primary source: the official website, the original paper, the government page.
  4. Note the date of the source; information changes.
  5. Remove or fix anything you couldn’t verify.

Our full guide, How to fact-check AI answers, expands this routine with examples.

The bigger picture

Organisations such as the US National Institute of Standards and Technology (NIST) include the accuracy and reliability of AI output in their risk management frameworks for good reason: confident errors can cause real harm when people rely on them. For individuals the lesson is practical rather than alarming. Use AI for what it does brilliantly — drafting, explaining, organising, brainstorming — and keep a human in charge of the facts.

Frequently asked questions

Do newer AI models still hallucinate?

Yes. Newer models generally make fewer errors on many tasks, and web search helps, but no current chatbot is free of hallucinations. Treat any important fact as unverified until you check it.

Are AI citations reliable?

Not automatically. Chatbots have produced citations to papers, court cases and articles that don’t exist, or real sources that don’t say what the chatbot claims. Open every source and check the quote.

Does web search stop hallucinations?

It reduces them for factual questions because the model can base its answer on retrieved pages, but it can still misread or misquote a source, or rely on a poor one.

Is it a hallucination if the AI is outdated?

Strictly, outdated information is a knowledge-cutoff problem rather than a hallucination, but the effect is similar: a confident answer that’s wrong today. Ask for sources and dates.

Which chatbot hallucinates least?

It changes with every model update and depends on the task, so we don’t publish a ranking without our own controlled testing. Tools that show sources next to answers make errors easier to spot.

Sources

  1. Kalai et al. (2025), Why Language Models Hallucinate (opens in a new tab)
  2. Ji et al. (2023), Survey of Hallucination in Natural Language Generation (opens in a new tab)
  3. Anthropic — Reduce hallucinations (opens in a new tab)
  4. Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (opens in a new tab)
  5. NIST — AI Risk Management Framework (opens in a new tab)

Facts, prices and features were checked against these sources on November 19, 2026. AI products change quickly; if you spot something out of date, please tell us.