Explained simply · Free to read · 4 min read

AI hallucination explained simply: why AI makes things up and how to check answers

An AI hallucination is a false statement that a chatbot phrases as convincingly as if it were true, for example a wrong date, an invented quote or a source that does not exist. It happens because language models string together likely words. An OpenAI study also explains part of the problem by the fact that common training and evaluation methods reward guessing more than an honest "I don't know". How often it happens depends heavily on the task and the model: in a large study by European public broadcasters, 45 per cent of answers to news questions had at least one significant issue. You should therefore always check important information against the original source (as of October 2026).

Published 11 October 2026

In short

AI hallucination: An AI hallucination is a false or invented statement that an AI chatbot phrases as confidently as if it were true, such as a wrong date, a made-up quote or a source that does not exist.^1,2,4; Example: You ask a chatbot for a book on the history of your home town. It gives you a title, author and publisher, but when you ask at the library, it turns out the book does not exist.; Also called: hallucination · AI makes up sources · invented facts · KI-Halluzination

What hallucination means

When experts say an AI hallucinates, they mean it gives an answer that is factually wrong but sounds confident. OpenAI's help pages list typical examples: wrong definitions, dates or facts, invented quotes, studies or references, and overconfident answers to ambiguous questions.

As OpenAI's researchers point out, this has little to do with human hallucinations. The term has stuck anyway.

Germany's consumer advice centre, the Verbraucherzentrale, notes that chatbot answers are not checked for accuracy. They can be inaccurate, incomplete or inappropriate and are no substitute for a reliable source or professional advice.

Why AI makes things up

A language model learns from huge amounts of text which word is likely to come next. It only sees fluently written sentences, with no label saying whether a statement is true or false. That lets it reproduce common patterns well. Rare details, such as the birthday of a little-known person, cannot be derived reliably from such patterns.

According to an OpenAI study from September 2025, there is a second reason: most tests used to rate models only count correct answers. It is like a multiple-choice exam where guessing can earn points but "I don't know" never does. OpenAI's example: anyone who simply guesses an unknown birthday has a 1 in 365 chance of being right. Models are effectively trained to guess rather than admit they don't know.

In addition, a model only knows things up to a certain cut-off date unless it searches the web. And according to OpenAI, confident wording does not mean the answer is correct.

How often it happens: three dated measurements

There is no single error rate for all chatbots. The figures depend heavily on what is being tested. Three examples show the range.

News: in a study published by the European Broadcasting Union (EBU) on 22 October 2025, 22 public service media organisations from 18 countries, including Germany's ARD, ZDF and Deutsche Welle, checked more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity. 45 per cent had at least one significant issue. 31 per cent had serious sourcing problems and 20 per cent major accuracy issues, including invented details and outdated information.

Summaries: the company Vectara measures how often models add made-up content when summarising texts, even though the text is right there. It uses more than 7,700 articles of roughly 50 to 24,000 words. In its ranking of 22 September 2026, rates range from 1.8 to 24.2 per cent depending on the model. One limitation: the rating is not done by people but automatically by Vectara's own evaluation model, HHEM-2.3.

Knowledge questions: OpenAI itself reports from a test with difficult factual questions that an older model was wrong 75 per cent of the time and a newer thinking model 26 per cent, because it declined to answer in just over half of the cases. These are vendor figures.

Our assessment: the numbers cannot be compared directly. But they show that errors still occur regularly with current models.

How to check answers in everyday life

OpenAI advises treating a chatbot answer as a first draft, not a final source. Always check quotes, figures, technical details and references to documents. Let the chatbot search the web and open the links it gives you yourself. Does the page really say what the chatbot claims? Does the page exist at all? The EBU study shows that source references in particular are often missing or misleading.

The Verbraucherzentrale recommends evaluating information before you use or share it. For news and claims, fact-checking sites, where journalists verify reports, can help. Germany's tagesschau advises relying on established media for news and being especially careful with politics, health and money.

You can also help prevent hallucinations. Anthropic recommends explicitly allowing the AI to say "I don't know". If you upload your own text, you can ask the AI to back up its statements with word-for-word passages from that text. If it cannot find a matching passage, it should drop the statement.

Sources

Created with AI assistance and editorially approved before publication.