The short answer
AI hallucination is when an AI model produces false or made-up information and presents it as fact, such as an invented statistic, quote, source or legal case. NIST's Generative AI Profile calls it confabulation: the production of confidently stated but erroneous or false content by which users may be misled. Google's Machine Learning Glossary defines it as plausible-seeming but factually incorrect output. It happens because language models predict likely text rather than look facts up, and no current model is free of it.
What AI hallucination means, in plain English
A hallucination is an answer that sounds right and is wrong. The model does not know it is wrong. It writes a fluent sentence with a date, a name or a figure that fits the pattern of a good answer, whether or not it is true.
The US National Institute of Standards and Technology (NIST), in its Generative Artificial Intelligence Profile (NIST AI 600-1, July 2024), uses the term confabulation and defines it as the production of confidently stated but erroneous or false content, known colloquially as hallucinations or fabrications, by which users may be misled or deceived. Google's Machine Learning Glossary defines hallucination as the production of plausible-seeming but factually incorrect output. Anthropic's documentation describes it as text that is factually incorrect or inconsistent with the given context.
Note the second half of Anthropic's wording. A hallucination is not only a wrong fact about the world. It can also be a summary that adds a point your document never made.
Why AI models hallucinate
- They predict, they do not retrieve. A large language model generates the most likely next token. Without a search or a document to read, it fills gaps with plausible text.
- Their knowledge stops at training. Ask about last month and a model without web search may invent rather than say it does not know.
- Long inputs blur. Anthropic notes that accuracy and recall degrade as the context window fills, so a detail on page 80 of a long report is easier to get wrong.
- Prompts push them. "Give me five studies that prove X" invites the model to produce five studies, whether or not five exist.
Why hallucination matters at work
NIST points out that the risk comes from people believing false content, often because of the confident tone, and acting on it. Its own example is a confabulated summary of patient information leading doctors to a wrong diagnosis. In office work the usual versions are smaller but still costly: a made-up figure in a board deck, an invented clause in a contract summary, a non-existent source in a report, or a wrong date in a client email.
The makers say it plainly too. Anthropic's guide on reducing hallucinations ends with the note that its techniques significantly reduce hallucinations but do not eliminate them, and that critical information should always be validated. Our ChatGPT guide makes the same point: confident errors still happen with GPT-6 models.
Where it bites hardest
- Citations, quotes and references can be invented or attributed to the wrong source.
- Numbers such as statistics, prices, dates and percentages are easy to get wrong.
- Niche or recent topics with little training data are a weak spot.
- Legal, medical and financial details need extra checking.
- In summaries of long documents, the model may add or merge points.
Try it: catch a hallucination yourself
- Pick a topic you know very well, such as your own company or a niche part of your field.
- In a chat with web search off, ask: "Give me three statistics about [topic], each with the source and year."
- Look up each source. Check whether it exists, whether it says that, and whether the year is right.
- Now repeat the question in a tool that cites sources by default, such as Perplexity, where every answer includes inline citations, and click through.
Most people who do this once never again paste an unchecked AI statistic into a report.
Five ways to reduce and catch hallucinations
These come from Anthropic's published guidance and work in any assistant.
- Give permission to say "I don't know". Anthropic says explicitly allowing uncertainty can drastically reduce false information.
- Ground it in your documents. Attach the source and tell the model to use only that material. Tools built for this, such as Gemini Notebook (formerly NotebookLM), answer only from your notebook sources with citations. That approach is retrieval-augmented generation.
- Quotes first. For long documents, ask for word-for-word quotes before the analysis.
- Make it cite and self-check. Ask for a supporting quote or link for every claim, and to remove any claim it cannot support.
- Ask twice, or ask two tools. Anthropic calls running the same prompt several times best-of-N verification: inconsistent answers are a warning sign. You can also compare two assistants; see how to compare AI answers.
For a full process, read how to reduce AI hallucinations and how to check if AI is telling the truth.
Common misconceptions about AI hallucination
- "Newer models have solved it." Newer models make fewer errors on many tasks, but every maker still tells you to verify important facts.
- "If it gives a source, it is true." Models can cite sources that do not exist or do not say what is claimed. Open the link.
- "Web search removes hallucinations." It helps a lot, but the model can still misread a page or merge two sources.
- "The AI is lying." Lying needs intent. NIST notes that some commenters object to "hallucination" itself because it gives software human traits. The model is generating plausible text, not deceiving you.
- "It only happens with facts." A summary that drops a key caveat or adds a point that is not there is also a hallucination.
Related terms
- RAG means grounding answers in retrieved documents.
- Deep research covers multi-step, cited research modes.
- Reasoning model covers models that work step by step, which can help with logic errors.
- Context window explains why long inputs raise the risk.
- AI safety is the wider field that studies these risks.
Prompts to copy
Answer only from my document
Answer using only the document below. If the document does not contain the answer, say "Not in the document" instead of guessing. After each sentence, add the exact quote that supports it in brackets. Question: [your question]. Document: [paste or attach]
You get: An answer where every sentence is tied to a quote you can check, and a clear flag when the document is silent.
Self-check before you send
Review your last answer. List every factual claim (names, numbers, dates, sources). For each one, label it confident, unsure or unable to verify, and say why. Remove or mark any claim you cannot support.
You get: A checklist of claims sorted by confidence, so you know exactly what to verify by hand.