Hallucination
A hallucination refers to a response from an AI system that sounds convincing but is factually wrong or entirely made up. The problem isn't a glitch, but a consequence of how these programs generate text.
A chat program like ChatGPT answers questions in fluent, well-formulated sentences. Sometimes, however, the content of these sentences is simply invented. It cites a book that was never written, a court ruling that doesn’t exist, or an incorrect date of birth. Such made-up details are called hallucinations. The term is somewhat misleading, since the program doesn’t see ghosts and isn’t deliberately lying either. It simply has no built-in distinction between a true statement and one that merely sounds good.
Why invented facts are so dangerous
An obvious error is harmless because people notice it. Hallucinations are the opposite: they come in the same confident tone as correct answers. There’s no shaky tone, no hesitation, no warning. That’s exactly why people adopt them without checking.
In practice, this has already caused real harm. In the US, lawyers filed a brief in 2023 that cited several court rulings that didn’t exist. A chat program had invented them, complete with plausible case numbers. The court imposed a fine. Similar cases have occurred in medicine, journalism, and education.
For businesses, this is the main reason why AI is only used in sensitive areas with human oversight. An error rate of a few percent sounds low. But with a million queries a day, that’s thousands of incorrect answers.
Word probabilities instead of a knowledge database
A language model is a program that has learned which word best fits a given sequence of words. To do this, it has analyzed vast amounts of text. For every answer, it chooses the most probable continuation, word by word. It doesn’t look anything up—it calculates.
The comparison to a student works well here. Imagine someone who has read a great deal but isn’t allowed to use a book during the exam. When there’s a gap in memory, they write down something that fits with the rest. The result is an answer with the right style but the wrong content. That’s exactly what a language model does, just without the guilty conscience.
Hallucinations are therefore not a bug that can simply be removed. However, they can be significantly reduced. The most important technique is called RAG, short for retrieval-augmented generation: before answering, the model is given real documents and is only allowed to cite from them. Checking with a second model and citing sources also help. According to current knowledge, the problem doesn’t disappear entirely.
How to spot hallucinations in everyday use
They occur most often with detail-level questions. Typical triggers are page numbers, years, names of minor characters, studies, legal sections, or links. Anyone asking for a web address often gets one that’s structured like a real one but leads nowhere. Made-up bibliographic references in term papers are also a classic case.
In the news, you’ll usually encounter the term in two contexts. First, in product announcements: companies like OpenAI or Google advertise that a new model hallucinates less. Second, in regulation, for instance when it comes to liability for incorrect information provided by an AI.
Hallucination should be distinguished from two similar things. Outdated knowledge is not a hallucination, just old: the model was trained up to a certain cutoff date. And a bias in the model, often called just that, is a skew in evaluation, not an invented fact. Practical rule: for anything you intend to reuse, check numbers, names, and sources yourself.