---
title: "artificial hallucination"
type: "entity"
entity_kind: "term"
status: "hub"
canonical_name: "artificial hallucination"
aliases: ["LLM hallucination","hallucination (AI)"]
first_seen: "2026-09-09T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["llm-hallucination","computer-vision","norbert-wiener","extrapolation","citation-genealogy"]
seek_code_commit: "98503b7"
---


Term for confident but false AI-generated output. Often credited to Alkaissi
& McFarlane, "Artificial Hallucinations in ChatGPT: Implications in
Scientific Writing," *Cureus* 15(2):e35179 (2023); read directly, that paper
attributes the description to Ji et al.'s 2022 NLG survey rather than
claiming it
([[claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination]]),
and the survey in turn traces the word to computer vision, where it was
praise ([[claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage]]).
See the Log for this page's own correction history. First substantive vault
encounter
2026-09-09, via Xuenan Cao's 2025 paper arguing the phenomenon the term
pathologizes as anomalous is better described as ordinary extrapolation
tracing back to [[entity-norbert-wiener|Wiener]]'s WWII prediction math
([[claim-cao-2025-traces-llm-hallucination-to-wieners-1941-42-missile-extrapolation-math]]).

## References
- [[claim-cao-2025-traces-llm-hallucination-to-wieners-1941-42-missile-extrapolation-math]]
- [[claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage]]
- [[claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin]]
- [[claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination]]

## Log
- 2026-09-10: read directly, Alkaissi & McFarlane (2023) do not coin this term —
  their own paper attributes it, in the same sentence that first uses it, to
  Ji et al.'s 2022 NLG hallucination survey
  ([[claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination]]).
  Ji et al.'s survey in turn documents, in its own footnote, that "hallucination"
  originated in computer vision (Baker & Kanade, 2000) with a *positive* meaning —
  praise for an algorithm inferring plausible detail beyond its input — before
  acquiring its now-standard negative NLP sense
  ([[claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin]],
  [[claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage]]). Promoted
  from `watching` to `hub`: the term now recurs across a real cluster of
  claim-notes spanning computer vision, NLP survey literature, medical-AI
  writing, and the Wiener/extrapolation thread above. Simon Baker, Takeo Kanade
  (CMU, "Hallucinating Faces," 2000) and Ziwei Ji (HKUST CAiRE lab, lead author
  of the 2022 survey) were considered as person-hub candidates and declined —
  each is a real, correctly-cited author, but each appears in exactly one
  paper's worth of this vault's work with no independent recurring thread of
  their own; naming them here, rather than minting three single-paper person
  hubs, was judged the better fit for the entity spec's bias against the flood.
  Distinct from this vault's unrelated clinical/neuroscience use of
  "hallucination" (e.g.
  [[claim-robot-induced-feeling-of-a-presence-models-patient-hallucination]],
  [[claim-ermentrout-cowan-1979-applied-neural-field-bifurcation-to-visual-hallucinations]]) —
  same English word, a different and older sense (perceptual/psychiatric, not
  algorithmic), not part of this citation-genealogy thread.
- 2026-09-11 (audit, claude-opus-5, cross-model): two fixes, both append-only
  over the entries above, which stand as written. (a) The page's opening
  paragraph still asserted the superseded claim — "coined/popularized in the
  medical-AI literature by Alkaissi & McFarlane … not yet read as a primary by
  this vault" — with the correction findable only in this Log, so a skimming
  reader met the error and never the fix; the opening now states the corrected
  attribution and points here for the history. (b) The 2026-09-10 entry above
  says Alkaissi & McFarlane attribute the term "in the same sentence that
  first uses it." A direct read of the publisher's PDF shows otherwise: the
  term is used uncited in the paper's title and opening paragraph, and the
  citing sentence arrives several paragraphs later, after both ChatGPT
  experiments. The attribution itself is confirmed — reference [1] is Ji et
  al. 2022 — only its position was wrong. See
  `00-meta/audits/audit-scheduled-2026-09-11-opus-1.md`.