An AI hallucination is a confident, fluent response generated by an artificial intelligence system that contains fabricated, inaccurate, or unverifiable information presented as fact. The term is most commonly applied to large language models (LLMs) and other generative AI systems, which can produce text, images, citations, or numerical claims that sound authoritative but have no grounding in real data. Because these systems are designed to generate plausible output rather than retrieve verified truth, hallucinations are a structural property of how they work, not an occasional bug.
How AI hallucination works
Large language models generate output one token at a time by predicting what is statistically likely to come next, based on patterns learned during training. The model has no built-in notion of a verified fact database; it only knows which words and phrases tend to follow others. When a prompt asks for specific details such as a citation, a date, or a person's biography, the model may produce text that looks like the right shape of an answer without any constraint tying it to real-world facts. If the training data is sparse on a topic, or if the question is unusual, the model effectively fills in the gaps with its best guess.
A simple example: ask an LLM for a paper title and author, and it may invent a realistic-sounding title attributed to a real researcher who never wrote it. The model is not lying deliberately; it is doing what it was trained to do, which is produce a likely continuation. Retrieval-augmented generation (RAG), tool use, and constrained decoding are common techniques used to reduce, though not eliminate, this behavior.
Why it matters
Hallucinations matter because users frequently treat AI output as authoritative, especially when it is presented in a confident, polished format. In domains such as medicine, law, journalism, and software engineering, a single fabricated citation or invented API can cause real harm. This is why most enterprise and production deployments pair generative models with retrieval systems, fact-checking layers, and human review. Understanding that hallucination is a default behavior, rather than a rare failure, is essential for anyone shipping or relying on AI-generated content.
Common types of AI hallucination
- Factual hallucination: confidently stated facts that are wrong, such as invented statistics, dates, or historical events.
- Fabricated citation: references to papers, articles, or URLs that do not exist or attribute real work to the wrong author.
- Intrinsic vs. extrinsic: an intrinsic hallucination contradicts the source material the model was given (for example, a provided document), while an extrinsic hallucination introduces information not supported by any source at all.
- Reasoning hallucination: logical steps that look valid but contain flawed inferences, such as correct-looking math that arrives at a wrong answer.
Hallucinations are an inherent trade-off of today's generative AI: the same fluency and flexibility that make models useful also make them capable of inventing plausible falsehoods. Treating model output as a draft to be verified, rather than a finished source of truth, remains the most reliable defense. Tools that ground models in retrieved documents, enforce structured outputs, or check claims against trusted databases can meaningfully reduce the rate of hallucinations, but they do not remove the underlying cause.