You ask a chatbot a simple question, and it answers with total confidence. The tone is perfect, the details look real, and the source even has a title. Then you check, and none of it exists. You just met an AI hallucination.
This problem is one of the most important things to understand about AI tools, because it can cost you time, money and trust. This guide explains why it happens, shows real cases and gives you simple ways to catch it.
Quick answer: An AI hallucination is a confident answer that is false or not backed by any source, such as a made up fact, quote, link or court case. It happens because language models predict likely words, and many tests reward guessing over saying “I don’t know”. You can reduce it by giving the AI your sources, asking for citations and checking key facts yourself.
What is an AI hallucination?
An AI hallucination is content that sounds right but is wrong. It can be a fake statistic, a quote no one said, a book that was never written or a link that leads nowhere.
The US standards body NIST calls this “confabulation” in its Generative AI Profile. OpenAI describes hallucinations as plausible but false statements. Either way, the danger is the confidence. A wrong answer delivered calmly is easy to believe.
Why an AI hallucination happens
In September 2025, OpenAI published research called Why language models hallucinate. It points to two main causes.
1. Models learn patterns, not truth
A language model learns by predicting the next word across huge amounts of text. That text has no labels saying which sentences are true. So the model learns regular patterns well, like grammar and spelling. However, rare facts, such as a person’s birthday, follow no pattern. When the model meets one, it often guesses.
2. Tests reward guessing
Many benchmarks only measure accuracy. A lucky guess earns a point, while “I don’t know” earns zero. As a result, models learn that guessing pays.
OpenAI showed the effect with its own models on the SimpleQA test. The older o4-mini model answered almost everything and got 24% right, but its error rate was 75%. The newer gpt-5-thinking-mini declined to answer 52% of the time, scored 22% right, and its error rate fell to 26%. Slightly lower accuracy, far fewer confident mistakes.
Real cases where an AI hallucination cost money
These are not just funny chatbot errors. Real people have paid for them.
- Fake court cases: in the Mata v. Avianca case in New York, two lawyers filed a brief that cited cases ChatGPT had invented. In June 2023, the court fined them and their firm $5,000.
- A chatbot’s promise: in 2024, a Canadian tribunal ruled that Air Canada had to honor a refund policy its website chatbot made up. The airline argued it was not responsible for its chatbot. The tribunal disagreed and ordered a partial refund of CA$650.88 plus interest and fees.
- A growing list: researcher Damien Charlotin tracks court decisions involving hallucinated AI content. His database listed 2,246 cases worldwide on 2 November 2026.
The lesson is clear. If an AI tool speaks for you or your business, you own what it says.
Which AI hallucinates the least?
There is no single answer, and you should be wary of anyone who gives one. Hallucination rates depend on the task.
For example, Vectara’s public leaderboard measures how often models add false details when summarising a document. In its May 2026 update, scores ranged from under 2% to over 24% across models. In contrast, Stanford researchers found in 2024 that AI legal research tools gave wrong information more than 17% of the time, and general chatbots did far worse on legal questions.
So a model that does well at summaries may still invent facts in a niche topic. Treat every leaderboard as a snapshot, not a guarantee.
Free: AI Tools Cheat Sheet 2026
24 free tools sorted by task, a simple prompt formula and 10 prompts you can copy. We send the PDF straight to your inbox.
Sent! Check your inbox
We emailed the AI Tools Cheat Sheet to you. Not there in 2 minutes? Look in Spam or Promotions.
How to spot an AI hallucination
Watch for these warning signs:
- Very specific numbers, dates or quotes with no source.
- Citations, book titles or court cases you cannot find with a quick search.
- Links that lead to error pages or unrelated sites.
- Different answers when you ask the same question twice.
- Confident claims about recent events the model may not know.
- Rare details about small companies, local people or niche topics.
If you want to check whether a whole piece of text came from AI, our guide on how to tell if text is AI generated covers the tells.
How to reduce AI hallucinations as a user
You cannot remove hallucinations, but you can cut them sharply:
- Give the AI your source. Paste or upload the document and say: answer only from this, and say “not in the document” if the answer is missing.
- Ask for citations, then open them. A link that looks real is not proof.
- Allow “I don’t know”. Tell the model it is fine to say it is unsure.
- Use search modes for current facts. Do not rely on a model’s memory for recent news.
- Cross check important answers. Ask a second model or search the key claim.
- Be extra careful with law, health and money. Always confirm with an official source or a professional.
How builders reduce hallucinations
If you build AI apps or chatbots, a few methods help most:
- Retrieval augmented generation: let the model pull facts from your own trusted documents. Our guide to RAG explains how it works.
- Quote first answers: ask the model to pull exact quotes before it answers, then drop any claim without a quote.
- Better evaluation: OpenAI recommends tests that punish confident errors more than honest uncertainty.
- Human review: keep a person in the loop for anything high stakes.
Anthropic’s guide to reducing hallucinations lists more prompt techniques and notes that they reduce errors but do not remove them.
Also read: How to Use NotebookLM (Now Gemini Notebook): Step by Step Guide for 2026
FAQ
What is an AI hallucination in simple words?
It is when an AI gives a confident answer that is false or made up, such as a fake fact, quote or source.
Why does ChatGPT make things up?
It predicts likely words rather than looking up facts, and training often rewards guessing over admitting it does not know.
Can AI hallucination be fully eliminated?
Not with today’s methods. Anthropic says its techniques reduce but do not eliminate them, and OpenAI says models can avoid many by admitting uncertainty.
How do I stop AI from making up sources?
Give it your own documents, ask it to cite exact quotes, and open every source before you use it.
Which AI model hallucinates the least?
It depends on the task. Rates differ for summaries, short facts and legal or medical questions, so test on your own use case.
