A hallucination is output that is fluent, confident and false: a fabricated citation, an invented statistic, a plausible event that never happened. The term is universal in the field and it is a bad one, because it implies a malfunction. Nothing is malfunctioning. The system is doing exactly what it does, which is produce probable text, and the probable text is sometimes not true.
The answer, in one line
Output that is fluent, confident and false: a fabricated citation, an invented statistic, a plausible event that never happened. It arrives in the correct format, with the register and structure of accurate information, and carries no internal marker distinguishing it from output that happens to be right.
Definition#
Hallucination: generated content presented as factual that is not supported by the model's training data, the provided context or reality. It arrives in correct form, with the register and structure of accurate information, and carries no internal marker distinguishing it from output that happens to be right.
The case for retiring the word#
It implies a fault state. Hallucination in a person is a departure from normal perception. In a language model there is no equivalent normal state to depart from: correct and incorrect output are produced by the identical process. Calling the wrong ones hallucinations suggests a bug that could be fixed, rather than a property of how the thing works.
It anthropomorphises, and that changes behaviour. A system that "hallucinates occasionally" sounds like a mostly-reliable colleague having a bad moment. That framing invites exactly the reduced scrutiny described in automation complacency.
It collapses distinct failures. Fabricating a source, misremembering a real fact, over-generalising from thin evidence and confidently answering an unanswerable question are different problems with different responses. One word for all of them obstructs thinking.
Better alternatives exist and are used in the technical literature: confabulation, which is closer to the human analogue that actually fits, or simply fabrication and factual error, which say what happened.
Why it happens#
A language model produces the most probable continuation of a text. The most probable continuation of "the leading study on this is" is a plausible-looking citation, whether or not one exists. Fluency is generated by the same process as accuracy and is not connected to it, which is the subject of why AI sounds so confident.
The form is the dangerous part. A fabricated citation has authors, a year, a journal and a volume. A wrong figure has the right number of digits. Readers scan form first, so these errors survive review by people who are paying attention.
The case that matters#
In January 2025 the High Court in Pietermaritzburg dealt with counsel who had cited authorities that did not exist. The judge tested one citation by asking ChatGPT, which confirmed the non-existent case was real and then confirmed it addressed a point it could not have addressed. The judgment was referred to the Legal Practice Council.
Two lessons, and the second is the one people miss. Fabrications reach professional output. And you cannot verify a machine's output with another machine, because the second system is producing probable text about the first system's probable text rather than checking anything.
Rates are falling, and by how much is unclear#
Rates are falling, retrieval-grounded systems fabricate less, and citations to real documents are increasingly checkable automatically. Whether the failure mode is reducible to negligible or is intrinsic to the architecture is genuinely contested among researchers, and anyone confident in either direction is going beyond what is established. This page describes systems as they behave in 2026.
Why the word sets the response#
The word matters because it sets the response. "Hallucination" invites waiting for a fix. Fabrication invites a verification process, which is what an organisation actually needs and what regulation now expects.
And the practical burden falls in a predictable place. Detecting a fabrication requires enough domain knowledge to know the cited thing does not exist, which means verification is expertise applied rather than administration. That is the argument in who owns verification. It is why the problem gets worse in organisations that have automated away the work which built the expertise.
Related SuperSkills research#
On the mechanism, why AI sounds so confident. On detection, how do I know when AI is wrong. On the tendencies it exploits, automation bias and automation complacency. On who catches it, who owns verification.
Key sources
- Mavundla v MEC: COGTA KwaZulu-Natal [2025] ZAKZPHC 2. High Court of South Africa, Pietermaritzburg.
- Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
- Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier.
About this definition#
Hallucination is the established field term and is not a SuperSkills coinage. The argument for retiring it is the author's interpretation and is marked as such. On a 90-day review cycle, because fabrication rates move.
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.
Explainer · SS-2026-073 · Graded against the published rubric
Hirji, R. (2026). What is an AI hallucination?. The SuperSkills evidence base, SS-2026-073. https://thesuperskills.com/research/what-is-an-ai-hallucination. Last reviewed 26 August 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work