Ji et al.'s 2022 NLG survey explicitly documents that 'hallucination' originated in computer vision with a positive meaning before acquiring its negative NLP sense

aillm-hallucinationterminologycomputer-visionnlp-historycitation-genealogycross-domain-bridge

Ji, Lee, Frieske, et al.'s "Survey of Hallucination in Natural Language Generation" (ACM Computing Surveys, arXiv:2202.03629) states in footnote 2, attached to the introduction's first use of "hallucination" as the name for the phenomenon ("Researchers started referring to such undesirable generation as hallucination [177]"; the word itself appears earlier, in the title and abstract): "The term 'hallucination' first appeared in Computer Vision (CV) in Baker and Kanade [9] (claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage) and carried more positive meanings, such as superresolution [9, 159], image inpainting [69], and image synthesizing [310]. Such hallucination is something we take advantage of rather than avoid in CV."

The same footnote continues past the sentence usually quoted from it: "Nevertheless, recent works have started to refer to a specific type of error as 'hallucination' in image captioning [17, 222] and object detection [8, 119], which denotes non-existing objects detected or localized incorrectly at their expected position. The latter conception is similar to 'hallucination' in NLG." On the survey's own account, then, the negative sense did not originate in natural-language generation either: it appeared first inside computer vision, and NLG's usage is described as resembling that later CV sense rather than as a separate NLP invention. The positive-to-negative flip is a shift within CV that NLG inherited, not a flip that happened at the border.

The significance here is not just the underlying genealogy but that a widely-cited technical survey states it as settled background, in one footnote, rather than treating the CV-to-NLP shift as something requiring argument. The history was never hidden or contested; it was simply not read by later citers of the term (claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination).

Correction history.

• 2026-09-11 — The source_quote ended at "Such hallucination is something we take advantage of rather than avoid in CV," and the body said the survey "goes on to document the negative sense — ungrounded or unfaithful generated text — as a later, distinct development specific to natural-language generation." Footnote 2 does not stop there. Its final two sentences record that the negative sense had already appeared inside computer vision — in image captioning and object detection, for "non-existing objects detected or localized incorrectly" — and that "the latter conception is similar to 'hallucination' in NLG." The negative valence is therefore not NLG-specific on the survey's own telling. The quote field now carries the whole footnote and the body states the corrected reading. Separately, "at the survey's first use of the word" was corrected: the word appears in the title and abstract before the footnote, which is attached to the introduction's first use of it as the name for the phenomenon. Re-read against arXiv:2202.03629, sha256 2c095a10… (identical to the capture's). Found by the scheduled cross-model audit, 00-meta/audits/audit-scheduled-2026-09-11-opus-1.md.