Using AI well

Why does AI hallucinate? And how to catch it

AI chatbots make things up because they are built to produce likely-sounding text, not to look facts up. Here is why hallucinations happen, when they're most likely, and how to reduce them.

4 min read

Short answer

AI hallucinate because they are trained to produce likely-sounding text, not to look facts up. When a model hasn't learned a fact well, it can still write a fluent, confident answer that fits the pattern of a right one. Training and testing have also tended to reward a confident guess over saying "I don't know". You can cut the risk by giving the AI sources, asking it to quote them and checking anything important.

What an AI hallucination looks like

A hallucination is when an AI states something false as if it were true. It isn't lying: it has no intention to deceive. It's producing text that looks right. Common examples:

  • A book, research paper or web link that doesn't exist.
  • A quote someone never said.
  • Wrong dates, statistics or prices, given with .
  • A confident summary of a document that adds points the document never made.

This has real consequences. In 2023 a US federal judge fined two lawyers who had filed a brief citing court cases that ChatGPT had invented.

Why it happens

A chatbot is a large language model that writes by predicting the next token. It learned from a huge amount of text, but it doesn't keep a tidy database of facts it can look up. Its knowledge is spread across billions of numbers, a bit like a very well-read person's fuzzy memory.

  • Plausible beats true. The model is trained to continue text in a likely way. A made-up citation can look exactly like a real one.
  • Rare facts are weak. Things that appeared often in training, like the capital of France, are solid. A niche fact seen once or twice is easy to blur or invent.
  • Guessing gets rewarded. In a September 2025 article, OpenAI argued that common tests score a confident guess better than "I don't know", so models learn to guess.
  • No live knowledge. Without a search tool, the model knows nothing after its training cut-off, but may still answer as if it does.

When hallucinations are most likely

  • Questions about obscure people, places or small companies.
  • Exact figures, dates, quotes and references.
  • Recent news, unless the tool searched the web.
  • Long, multi-step answers, where one early mistake carries through.
  • Questions with a false assumption built in, like asking why something happened when it didn't.

How to reduce hallucinations

  1. Give it the source. Paste the document or data and ask it to answer only from that. This is the idea behind RAG, retrieval-augmented generation, which you can try in the RAG and agents lesson.
  2. Ask for quotes and links. Ask it to quote the exact passage it used. Then open the link or find the quote yourself.
  3. Give it permission not to know. Add "If you're not sure, say so" to your .
  4. Turn on search for current facts. A tool that searches the web and shows its sources is far safer for recent events.
  5. Ask again, differently. If two phrasings give different answers, at least one is wrong.
  6. Check what matters. Health, money, legal and anything you'll publish deserve a reliable second source.

More prompting tips are in prompt engineering for beginners and the lesson on using AI safely.

How AI companies are fixing it

Developers ground answers in search results or company documents, train models to express uncertainty and decline when unsure, and measure hallucination rates with evals, test sets with known answers. Reasoning models that check their own work can also help on some tasks. Hallucinations have become less frequent in newer models, but no model is free of them yet, so checking is still your job.

Key takeaways

  • A hallucination is a confident answer that is false or made up.
  • It happens because models predict plausible text rather than retrieve facts.
  • Rare facts, exact numbers, quotes, citations and recent events are the danger zones.
  • Give the model sources, let it say "I don't know", and verify what matters.

Learn it properly, free

Terms in this post

Questions

What is an AI hallucination?

An AI hallucination is when an AI system, usually a chatbot, states something false or made up as if it were fact, such as a fake quote, a wrong statistic or a source that doesn't exist.

Why does ChatGPT make up sources?

Because it predicts what a plausible reference looks like rather than looking one up. A made-up title, author and year can fit the pattern perfectly. Always open and check any source a chatbot gives you.

Can AI hallucinations be completely fixed?

Not yet. Grounding answers in real documents, web search and better training reduce them a lot, but every current language model can still be wrong. Checking important facts remains essential.

Which AI hallucinates the least?

It changes with every new model release and depends on the task. Models that search the web or answer from documents you provide, and that show their sources, are usually the most reliable for factual questions.

Sources and further reading

Read next

Spotted a mistake? Suggest an edit