Hallucination: From a Medical Term to an Everyday AI Complaint

The Entry: Hallucination

Definition: In AI discussions, a hallucination is information generated by an artificial intelligence system that is false, fabricated or insufficiently supported by the available source material, but presented in a way that can appear credible. The term is particularly associated with large language models, although related usage exists in other areas of computing.

Earliest documented computing use: Research published in 2024 under the title Evolution of Hallucination in NLP traces an early machine-learning use to image-resolution research by Baker and Kanade in 2000. In language-model discussions, Andrej Karpathy used the metaphor in a 2015 discussion of recurrent neural networks generating nonexistent URLs. The dates refer to documented research uses, not a claim that either was the first person ever to use the word this way.

Variants: AI hallucination, model hallucination, factual hallucination, ungrounded generation and confabulation. These expressions overlap, but they are not interchangeable in every technical discussion.

Constructed example: An assistant supplies a convincing quotation and attributes it to a book that does not contain those words. The quotation is a hallucination because its apparent specificity does not establish that the source exists.

• Factual hallucination: An incorrect statement about an identifiable fact.
• Fabricated citation: An invented source, author, publication or reference.
• Ungrounded generation: Content that cannot be adequately supported by the supplied evidence.
• False attribution: A genuine statement assigned to the wrong person or source.

A Word with a Much Older History

Hallucination did not originate in computer science. The English noun is recorded from the 1640s, according to the Oxford English Dictionary's historical tradition and the etymological account published by Etymonline. It derives from Latin hallucinatio, associated with hallucinari, a verb concerning wandering in the mind or mistaken perception. The exact deeper etymology has been debated.

The term acquired medical significance as European writers developed more systematic descriptions of unusual perceptual experiences. An influential historical account is associated with the physician Thomas Browne, whose Pseudodoxia Epidemica appeared in 1646. The work discusses errors in perception and received beliefs. Later medical writing developed more specialised distinctions between hallucinations, illusions and other unusual experiences.

In contemporary clinical language, a hallucination generally concerns a perceptual experience occurring without the corresponding external stimulus. Its causes and significance depend on the individual and clinical context. This meaning is important because the AI usage is metaphorical. A language model does not have human sensory perception, and its generation of an unsupported sentence is not evidence of a clinical experience.

How the Word Entered Artificial Intelligence

The transition into computing happened before the public popularity of chatbots. In image processing, researchers used hallucination to describe the generation of plausible visual detail when increasing the resolution of an image. Baker and Kanade's 2000 work on super-resolution is an example of this earlier technical usage. The word described detail reconstructed beyond what the original low-resolution image directly contained.

The metaphor subsequently became useful in language processing. Karpathy's 2015 discussion of recurrent neural networks illustrated how a model could generate apparently plausible material, including URLs that did not exist. Later natural language processing research used hallucination to discuss generated text that departed from an input, source or factual reference.

These developments matter because the modern AI vocabulary did not emerge in one sudden naming event. Different technical disciplines already had reasons to describe generated content as hallucinated. The arrival of widely accessible conversational systems brought those specialist concerns into ordinary discussion.

What Counts as an AI Hallucination?

The category covers several different problems. A chatbot might invent a historical date, misidentify an organisation, attribute a quotation to the wrong author or produce a bibliography containing publications that cannot be found. In document-based tasks, it might introduce a detail that does not appear in the supplied material. Each example involves a different relationship between the output and the evidence.

Consider a simple worked example. A user provides a paragraph stating that a fictional library opened in 1984. The assistant summarises the paragraph and adds that the library was designed by a named architect, although no architect appears in the source. The summary contains a hallucinated detail even if its description of the opening date is correct.

This distinction helps avoid treating an entire response as either reliable or unreliable. One answer can combine accurate information, reasonable interpretation and fabricated details. Accuracy must be considered claim by claim, with attention to the evidence supporting each statement.

• Check names, dates, quotations and numerical claims independently.
• Open cited documents and confirm that the relevant passage exists.
• Distinguish information in an uploaded source from additional background claims.
• Ask for uncertainty to be identified when evidence is incomplete.

Hallucination, Confabulation and Fabrication

Several neighbouring words help clarify the discussion. Fabrication describes the invention or creation of information, particularly when that information is presented as genuine. Confabulation has a history in psychology and clinical discussion and is sometimes proposed as an alternative metaphor for language-model errors. A 2023 article in PLOS Digital Health, Hallucination or Confabulation? Neuroanatomy as Metaphor in Large Language Models, discusses the conceptual differences and argues that confabulation may better describe some generated inaccuracies.

The terminology remains contested. Hallucination is widely recognised in AI research and public discussion, while fabrication can be more direct when a specific source or quotation has been invented. Unsupported generation is useful when the central issue is a lack of grounding, even if the output happens to be factually correct.

Editors should identify the particular failure before choosing a label. A false date is an inaccuracy. An invented source is a fabricated citation. A response that adds details absent from its source may be described as ungrounded. Hallucination remains a convenient umbrella term, but it does not explain the mechanism behind every error.

Why the Distinction Matters for Readers

The popularity of the term has created a useful shorthand, but it can also obscure differences between errors. A model may reproduce a misleading claim from its training material, misinterpret an instruction, lack access to current information or generate an unsupported detail. These situations can produce similar-looking answers while requiring different responses.

The word also carries a human-centred metaphor that can suggest intention, perception or experience. That implication should be handled carefully. Describing an AI system as hallucinating is a familiar way of discussing unreliable output, not a diagnosis or an explanation of what the system experiences.

For journalists, teachers and librarians, the practical question is always what supports a particular claim. The next time an AI-generated answer contains a precise name, quotation or date, locate the original source and compare the wording before incorporating it into published or educational material.

Questions readers ask