Definition
An AI hallucination is when an artificial intelligence states false or nonexistent information as if it were true, in a smooth and confident voice. The AI isn't trying to lie. It happens because the AI strings together the most likely next words based on probability, which makes it one of the best-known limits of today's AI.
Why does AI lie so confidently?
Imagine playing a word-chain storytelling game with a friend. Each person adds the word that fits most naturally after the last one.
Generative AI (a large language model, or LLM) writes in basically the same way. It doesn't deeply understand your question the way a person does. Instead, it draws on the huge amount of text it learned from and calculates, "After a question like this, these words usually come next." Then it adds them one at a time.
Here's the problem: when the AI doesn't have accurate information, it tends to produce a plausible answer instead of saying "I don't know." By default, it has no built-in step to check whether its answer is true. It just focuses on building the most believable-sounding sentence, with smooth grammar and a good fit for the context. So it can invent fake information that sounds real.
A closer look: the probability trap
Hallucination doesn't happen because AI is dumb. It comes from the math behind how language is generated, so it's very hard to avoid in principle.
A large language model reads a massive amount of text, like web pages, books, and documents, and learns which words are likely to follow a given context. For example, it knows statistically that "The capital of the United States is" is very likely followed by "Washington, D.C." But there are gaps in the statistics when a topic barely appeared in the training data, or never did, when a question mixes up similar pieces of information, or when a question is built on a false premise.
To fill those gaps, the AI blends together many expressions it has learned and creates a brand-new answer. And the better a model is at writing, the more it wraps wrong answers in smooth logic, which makes people easier to fool.
Efforts to reduce hallucinations
Hallucination is the biggest roadblock to trusting AI. In fields where accuracy is everything, like law, medicine, and finance, it can cause serious harm.
That's why AI is no longer asked to pull every answer purely from its own memory. A technique called retrieval-augmented generation (RAG) is now widely used. It has the AI look up trusted, up-to-date documents or databases first, and then write its answer. It's like letting a student open the textbook during the test instead of relying only on what they memorized.
Still, no matter how far the technology goes, AI works on probability, so it's hard to remove hallucinations completely. That means you shouldn't blindly trust important numbers or expert claims from AI. Always check the sources yourself and cross-verify.
Common misconceptions
AI deliberately lies to trick people.
AI has no bad intent and no intention to lie. It simply chains together likely next words, and sometimes that produces a combination that doesn't match the facts.
Once an AI model gets big enough, hallucinations disappear completely.
As long as AI generates words by probability, it's structurally very hard to get rid of made-up knowledge 100%.
Where you meet it
You ask about a historical event that never happened, and the AI answers as if it were an official entry in the history books, inventing dates and people that sound believable.
You ask for the source of a paper or news article, and the AI recommends a fake paper with a convincing title and a made-up scholar's name that don't exist anywhere.
In one sentence
A hallucination is a believable but false answer that AI invents when it chains together probable next words, not because it understands your question.