hallucination
Output from an AI model that is fluent, specific and presented as fact, but is false or unsupported by any source.
- The six court decisions ChatGPT invented in Mata v. Avianca
- Bard’s false claim about the James Webb telescope’s first exoplanet photo
- A chatbot’s false cargo report that nearly launched a US military operation
- Fake references in nearly 3,000 biomedical papers, reported in 2026
- Earliest AI use
- 1986
- Pattern recognition, as a goal: a network generating fingerprint patterns.
- Current meaning
- 2018
- Machine translation, as a fault: output untethered from its source.
- Also called
- Confabulation
- The psychology term. Some researchers prefer it.
- Opposite
- Grounded output
- Every claim traceable to a source the model was given.
In plain words
In AI, a hallucination is a false statement that a model presents as fact, often with specific detail such as names, dates, figures or citations.
A language model writes by predicting the next word from patterns in its training. It does not look anything up unless it is given a way to. Where the answer is thin or missing, the likeliest-sounding continuation can be something that never happened.1
The output has the same tone and formatting as a correct answer, which is why hallucinations are hard to spot. They matter most where a source is expected: law, medicine, finance and code. In 2023 a US judge fined two lawyers over six court decisions that ChatGPT had invented.2 By October 2026 one public database listed more than 2,100 court decisions dealing with hallucinated material.3 Outside the courts, Google’s Bard chatbot made an error in its first advertisement in February 2023, and Alphabet shares fell. In 2026 a chatbot’s false report on a ship’s cargo nearly set off a US military operation against a Chinese vessel.
The word is a metaphor. The model perceives nothing and intends nothing; it produces text that is well formed and wrong. Some researchers prefer confabulation for that reason.4 Hallucination is still the everyday word: in 2023 Cambridge Dictionary added the AI sense of hallucinate and named it word of the year.
A hallucination is different from an outdated answer, which was true when the model was trained and has changed since. That limit comes from the model’s knowledge cutoff. It is also different from a lie, which requires an intent to mislead.
How it happens
At each step the model ranks possible next words and picks from the top. Nothing in that step checks the facts. Testing makes it worse: a model scored only on accuracy does better by guessing than by saying it does not know.5
The problem starts earlier, in pretraining. If a model cannot reliably tell a false statement from a true one, ordinary statistical pressure will make it produce some false ones.6 Spelling follows consistent patterns, so larger models stop making spelling errors. A rare, arbitrary fact such as a pet’s birthday follows no pattern, so the model is left to guess.5
More capable models do not always hallucinate less. In 2025 The New York Times reported that newer reasoning systems from OpenAI and others scored worse on hallucination tests than older ones.
Supplying documents to answer from, known as grounding, does not remove the risk. In a 2020 study of automatic summaries, human raters found substantial hallucinated content even though each model had the source document as its input.
…the leading case on tolling under the Montreal Convention is
Origin and first known use
From Latin alucinari, “to wander in the mind”. Computer vision borrowed it first, as a compliment. Eric Mjolsness’s 1986 thesis used “fingerprint hallucination” for a neural network that generated fingerprint patterns. In 2000 Simon Baker and Takeo Kanade used “hallucinating faces” for filling in plausible detail when enlarging a low-resolution image. Machine translation researchers turned it into a warning in 2018.
- 1640s Enters English from Latin alucinari, “to wander in the mind”.Online Etymology Dictionary, “hallucination” Etymonline →: source for 1640s
- 1986 Eric Mjolsness’s Caltech PhD thesis uses “fingerprint hallucination” for a neural network that generates fingerprint-like patterns.Mjolsness, “Neural Networks, Pattern Recognition, and Fingerprint Hallucination”, Caltech PhD thesis, 1986 caltech.edu →: source for 1986
- 2000 Baker and Kanade’s “Hallucinating Faces” uses it approvingly: inventing plausible detail when enlarging a face was the goal.Baker and Kanade, “Hallucinating Faces”, IEEE FG 2000 cmu.edu →: source for 2000
- 2018 Google researchers use it for translations “completely untethered from the source material”. The negative sense starts here.Lee et al., “Hallucinations in Neural Machine Translation” research.google →: source for 2018
- 2020 Maynez and colleagues judge automatic summaries against their source documents and report substantial hallucinated content, meaning material unfaithful to the source.Maynez et al., “On Faithfulness and Factuality in Abstractive Summarization”, ACL 2020 arXiv 2005.00661 →: source for 2020
- 2022 OpenAI launches ChatGPT and lists as a limitation that it “sometimes writes plausible-sounding but incorrect or nonsensical answers”.OpenAI, “Introducing ChatGPT”, 30 November 2022 openai.com →: source for 2022
- 2023 Cambridge Dictionary adds the AI sense of hallucinate, to produce false information, and names it Word of the Year for 2023.University of Cambridge, “Cambridge Dictionary names ‘Hallucinate’ Word of the Year 2023” cam.ac.uk →: source for 2023
- 2025 Kalai and colleagues argue that models guess because training and benchmark scoring reward a guess over admitting uncertainty, and propose changing how benchmarks are scored.Kalai et al., “Why Language Models Hallucinate”, arXiv 2025 arXiv 2509.04664 →: source for 2025
In the wild
-
“…these models are highly prone to hallucinate content that is unfaithful to the input document.”
-
“A robot’s $100 billion error: Alphabet shares tank after its ChatGPT rival makes a mistake in its very first ad”
-
“Two US lawyers fined for submitting fake court citations from ChatGPT”
The Guardian theguardian.com →: The Guardian, 23 Jun 2023 -
“The result is known as ‘hallucination’, where a system produces highly plausible but incorrect results.”
Judge Anne Redston, Harber v HMRC, quoting the Solicitors Regulation Authority Decision →: Judge Anne Redston, Harber v HMRC, quoting the Solicitors Regulation Authority, 4 Dec 2023Archived →: copy of Judge Anne Redston, Harber v HMRC, quoting the Solicitors Regulation Authority, 4 Dec 2023 -
“…hallucination is all LLMs do. They are dream machines.”
-
“The production of confidently stated but erroneous or false content (known colloquially as “hallucinations” or “fabrications”)…”
NIST, defining confabulation in its generative AI profile nist.gov →: NIST, defining confabulation in its generative AI profile, 26 Jul 2024Archived →: copy of NIST, defining confabulation in its generative AI profile, 26 Jul 2024 -
“A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse”
The New York Times nytimes.com →: The New York Times, 5 May 2025 -
“Large language models sometimes produce confident, plausible falsehoods (‘hallucinations’), limiting their reliability.”
-
“AI hallucinations are infiltrating expert work—and entering the permanent body of knowledge”
-
“AI hallucination nearly triggers US military operation”
Not to be confused with
-
Confabulation
A near synonym from psychology: filling a gap in memory with invented detail, with no intent to deceive. Many researchers prefer it for the AI sense.
-
Lying
Lying needs a belief and an intent to mislead. A hallucinating model has neither; it is producing likely text, not hiding a true answer.
-
An outdated answer
Correct as of the training data and wrong now. The fact existed and then changed. A hallucinated fact never existed.
Questions
Why do language models hallucinate?
They are trained to produce likely text, not checked text. Where the right answer is rare or missing in the training data, a plausible guess still looks like an answer. A 2025 OpenAI paper adds that benchmarks scored only on accuracy reward guessing over saying “I don’t know”.
What is an example of an AI hallucination?
In 2023 a US judge fined two lawyers and their firm $5,000 over a brief that cited six court decisions invented by ChatGPT. Earlier that year, Google’s Bard chatbot made an error in its first advertisement, and Alphabet shares fell. In 2026 a chatbot’s false report on a ship’s cargo nearly set off a US military operation against a Chinese vessel.
Is hallucination the same as confabulation?
They are near synonyms. Confabulation is the psychology term for filling a gap in memory with invented detail, with no intent to deceive. Some researchers prefer it for AI because a model has no sensory experience to misperceive. NIST’s 2024 generative AI risk profile names the risk confabulation and calls hallucinations the colloquial term.
Can AI hallucinations be prevented?
Not fully. Supplying source documents does not prevent hallucinations on its own: in a 2020 study, summarization models given a document still produced substantial hallucinated content. In a 2026 Nature paper, Kalai and colleagues argue that pretraining leaves rare facts error-prone and that accuracy-only scoring rewards guessing. They propose benchmarks that state how wrong answers are penalized, so that guessing no longer pays.
When was hallucination first used in AI?
The earliest known AI use is Eric Mjolsness’s 1986 Caltech thesis, which used “fingerprint hallucination” for a neural network that generated fingerprint patterns. In 2000 Simon Baker and Takeo Kanade used “hallucinating faces” for adding plausible detail to enlarged images. The sense of an error dates from 2018, when Google researchers applied it to machine translations that ignored their source.
Watch
Cite this entry
“Hallucination.” SUPERR_dictionary, Superrduperr, 2 Oct. 2026, superrduperr.com/
Superrduperr. (2026, October 2). Hallucination. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Hallucination.” SUPERR_dictionary. October 2, 2026. https://superrduperr.com/