hallucination
Hallucination is when artificial intelligence confidently states something untrue or simply made up in a plausible form.
In plain terms
Imagine a student who did not prepare for the recitation but speaks so fluently and confidently that it is not immediately obvious they are improvising half of it. Language models sometimes behave similarly. They present non-existent books, incorrect years, or fabricated quotes in just as smooth sentences as the correct answer.
A hallucination is not a lie, because the machine has no intent to deceive. The model does not retrieve facts from a database but writes the continuation that seems most probable. If it lacks enough knowledge about the question, it fills the gap with plausible-sounding text.
A traditional software bug is usually visible: the program freezes or displays an error message. A hallucination, however, is a well-formulated, convincing answer, making it much harder to detect.
How it works
- A language model learns from vast amounts of text which words usually follow each other, not what is true and what is not.
- When you ask, it assembles the answer word by word: it always chooses the continuation that seems probable based on learned patterns.
- If it has seen little, contradictory, or outdated data on the topic, it will still write something, because it is tuned to answer, not to remain silent.
- The response format is flawless in such cases: it includes a name, a year, and a reference, but none of these may actually exist.
- Developers reduce errors by giving the model a search engine or reliable documents for the answer and training it to say: "I do not know this".
An everyday example
You ask a chatbot to recommend literature for your thesis, and it provides five well-formatted citations with author, year, and journal. When you search in the library, you discover that two of them never existed: the machine assembled them from elements that sounded real.
Who is affected
A common misconception
Limits and risks
- It cannot be completely eliminated today: better models make fewer mistakes, but none are flawless.
- Always verify names, numbers, dates, quotes, and references from independent sources, as errors are most common there.
- The rarer, fresher, or more local the topic (for example, Hungarian legislation or small-town data), the higher the chance of hallucination.
- If an agent acts independently based on incorrect information, the error no longer remains just in the text but also manifests in actions, such as sending a wrong email or recording wrong data.
Similar, but not the same
- disinformation – Disinformation is spread intentionally to deceive, whereas hallucination is an unintentional machine error.
- bias – Bias is a systematic distortion arising from one-sided training data. Hallucination, by contrast, is a fabricated or incorrect statement that can occur on any topic.
- deepfake – A deepfake is a false image, audio, or video created or commissioned by a human with AI assistance, usually for deceptive purposes. A hallucination, by contrast, is the model's unintentional error.
Related concepts
generative AI, trustworthy artificial intelligence, AI risk
fabricated citation, factual error
large language model, retrieval-augmented generation (RAG), fact-checking, prompt, knowledge cutoff
A brief history
In machine learning, the word "hallucination" first appeared in image processing, where it carried a neutral meaning. Researchers began using it for errors in language models from the late 2010s, initially in the context of machine translation. The general public became familiar with it after the launch of ChatGPT in November 2022. In 2023, the Cambridge Dictionary chose the verb "hallucinate" as its Word of the Year, precisely due to its new meaning linked to artificial intelligence.
Where things stand
By the end of 2026, hallucination remains a well-known, everyday weakness of language models. Systems working with search and source citations make fewer mistakes, but the phenomenon has not disappeared. It happens that a cited source does not support the claim, so verification remains your responsibility.
In the news
- New US sanctions over fake AI-generated citations as public database passes 2,200 casesNov 3, 2026 · Jog
- UK courts and Law Society urge checks on AI use in the legal professionNov 3, 2026 · Jog
- AI hallucinations in court filings are slowing, but judges still sound alarmsOct 19, 2026 · Jog
- Companies may be held liable when AI harms clients; tax court declines to sanction AI hallucinationOct 14, 2026 · Jog
- Lawyer's AI-hallucinated citations cost client the case; attorney must pay nearly $39,000Oct 12, 2026 · Jog
- Courts grapple with hallucinated citations: pro se litigants, sanctions and lawyers pushing backOct 5, 2026 · Jog
- Hallucination-laced TRO prompts alarm over AI use by courts as judicial AI rules take shapeOct 5, 2026 · Jog
- Legal AI tools: LawVu Lens, Harvey–LexisNexis motion to dismiss workflow, LAB-AA benchmarkOct 5, 2026 · Jog
Also known as
hallucination · AI hallucination · AI hallucination · hallucination · hallucination · AI hallucination · confabulation · confabulation
Sources
- Wikipedia: Hallucination (artificial intelligence)
- NIST AI 600-1: Artificial Intelligence Risk Management Framework – Generative AI Profile
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.