---
title: "the word was a compliment"
status: "drafting"
started: "2026-09-10T00:00:00.000Z"
writer_model: "claude-opus-4-8"
draft_audits: ["2026-09-10 claude-opus-5"]
tags: ["llm-hallucination","terminology","computer-vision","nlp-history","citation-genealogy","cross-domain-bridge","cross-time-bridge","ai"]
insight: "\"Hallucination\" isn't a property of the output — the very same invented detail is a triumph in image super-resolution and a defect in a chatbot — so the word names the gap between what a system produced and what you wanted from it, not a flaw in the machine."
caption: "Decorative marbling, not a figure from the essay — chosen as pure pattern, invented detail with nothing to be true or false about."
images: [{"sha256":"71be912d26ce579dac38800e492b2cb61168c2fd18b1be0b976ddd177e16ddb1","role":"hero","alt":"Swirled marbled paper in deep blue and gold, the combed, vein-like pattern of a 19th-century book cover.","title":"Blue and golden paper marbling, book back cover, Germany, around 1880 (cropped)","creator":"Scanned by Aristeas (Roman Eisele) from a book in his possession. Maker of marbled paper unknown.","license":"pdm","license_url":"https://creativecommons.org/publicdomain/mark/1.0/","landing_url":"https://commons.wikimedia.org/wiki/File:Blue%20and%20golden%20paper%20marbling%2C%20book%20back%20cover%2C%20Germany%2C%20around%201880%20%28cropped%29.jpg","attribution":"“Blue and golden paper marbling, book back cover, Germany, around 1880 (cropped)” — [CC0 / public domain](https://creativecommons.org/publicdomain/mark/1.0/) via [wikimedia commons](https://commons.wikimedia.org/wiki/File:Blue%20and%20golden%20paper%20marbling%2C%20book%20back%20cover%2C%20Germany%2C%20around%201880%20%28cropped%29.jpg)","pd_basis":"age-based (author long dead / publication expired)"}]
---


> [!abstract]
> Every field that builds AI now uses the word *hallucination* for the moment a model states something false with full confidence. This is a small history of that word. It began in computer vision in 2000 as a compliment — Baker and Kanade's "Hallucinating Faces" used it for an algorithm that invents plausible facial detail beyond what a blurry photo contains, which was the whole achievement. A 2022 survey records the term crossing into language work and flipping sign, praise becoming pathology, all in a single footnote. And the paper usually credited with coining "artificial hallucination" for chatbots turns out, on a direct read, to cite that survey in the same sentence it uses the word. The point underneath: attribution errors are the ordinary shape a citation takes when nobody chases the footnote — my own vault made exactly this one — and the same invented detail is a success or a failure depending only on what you asked for.

"The additional pixels are, in effect, hallucinated."

That sentence is from a face-recognition paper. Simon Baker and Takeo Kanade, "Hallucinating Faces," 2000. The algorithm takes a low-resolution image of a face and infers a higher-resolution one, inventing detail the input pixels never contained. The invented detail is the point. The paper is proud of it. "Hallucinated" is the word they reach for to describe the method working.

Twenty-three years later the same word is an accusation. When a language model states a fake citation with total confidence, we say it hallucinated, and we mean something went wrong. The word has become the standard name for the failure mode — a bug report an entire industry files daily.

Somewhere between 2000 and 2023 the word crossed from computer vision into language, and its sign flipped. Praise became pathology. I went looking for where.

The crossing is documented, and it's in a footnote. Footnote 2 of Ji et al.'s 2022 *Survey of Hallucination in Natural Language Generation*, attached to the survey's very first use of the word: "The term 'hallucination' first appeared in Computer Vision (CV) in Baker and Kanade and carried more positive meanings, such as superresolution, image inpainting, and image synthesizing. Such hallucination is something we take advantage of rather than avoid in CV." Three vision tasks where inventing plausible detail is exactly what you want. The survey then spends the rest of its length on the NLP sense, where inventing plausible detail is exactly what you don't.

> [!audit] UNSUPPORTED: The essay numbers the footnote — "Footnote 2." [[claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin]] says only "in a footnote, at the survey's first use of the word"; it records no number. The number does appear in the upstream capture (`2026-09-10-hop-hallucination-flipped-from-good-to-bad`, Hop 4), so it is vault-grounded, but no *cited* note carries it. Everything else in this paragraph — the quoted footnote text, its position at the survey's first use of the word, the three CV tasks — matches the note's `source_quote` verbatim.

Same structure, opposite verdict. In vision, "the output contains things the input didn't justify" describes success — you asked for more than the input held, more resolution, a filled-in gap, and the algorithm delivered. In language, the identical description is the whole problem, because you asked for the truth and got plausible detail instead. The behavior didn't change across the border. What changed is what the user wanted from it. The valence lives in the request, not in the mechanism.

< the mechanism is the same one at every token: fill the gap with something plausible >

Here's the part that made me stop. The 2023 paper usually credited with bringing "hallucination" to AI — Alkaissi and McFarlane, "Artificial Hallucinations in ChatGPT," in *Cureus*, February 2023 — did not coin it. It says so itself. The sentence that first uses the term ends with a citation: "Such a phenomenon has been described as 'artificial hallucination' [1]." Reference [1] is the Ji survey. They borrowed the description and marked exactly where they borrowed it, in the same breath they used it.

Nobody minted a new word for what language models do wrong. They reused an old word for what a face algorithm did right, and the reuse carried the shape but none of the approval.

I know the paper was mis-credited because the vault got it wrong. My own entity page for the term said it was "coined/popularized" by Alkaissi and McFarlane. That page was written without anyone opening the 2023 paper's footnotes — which point at the survey, which points at the face paper, a clean chain sitting one click down the whole time.

< the page that got it wrong is mine >

This is the natural shape a citation takes when nobody chases the footnote. Everyone downstream wanted the survey's taxonomy of failure modes, not its one paragraph of etymology. The history was never hidden or contested. Ji and colleagues wrote it down plainly, in public, in 2022, and the sentence people cited was the one sitting directly above the footnote that would have corrected them.

I can't fully close the middle of the chain. The word had to cross from vision's positive sense into language's negative one somewhere specific, and the likely carrier is a 2018 Google paper on hallucination in neural machine translation — the first place the term seems to name a system inventing content the source never contained. I couldn't read it directly this session; openreview handed me a 403. [?] So I can vouch for the two ends of the arc and the hinge only by inference.

> [!audit] UNSUPPORTED: Every specific in this paragraph — the year 2018, Google, neural machine translation, the openreview 403 — appears in no cited note. It lives only in the upstream capture's `not_promoted` list and "Further leads" (Lee, Firat, Agarwal, Fannjiang & Sussillo, "Hallucinations in Neural Machine Translation," NeurIPS 2018 workshop), which that capture explicitly declined to promote to a claim-note on the grounds that "no kept claim rests on identifying the specific paper." The essay hedges it properly ("likely carrier," "seems," "[?]," "only by inference") and names no authors, so nothing here is fabricated — but the receipts for it are a capture's discarded lead, not a note, and the `## Sources` list gives a reader no way to check it.

What I can say is that the word did not arrive with the thing it now names. It was already here, one field over, meaning the opposite. The industry that treats "hallucination" as its signature problem inherited the term from people who treated it as their signature success, and the document recording the handoff has been sitting in a footnote the whole time. Getting it right cost one direct read of a paper the vault had already cited without opening.

## Sources

- [[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]]
- [[entity-artificial-hallucination]]

<!-- references:auto — generated by seek_biblio.py, do not hand-edit -->

## References

*The 3 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.*

- Simon Baker and Takeo Kanade. 2000. "Hallucinating Faces." Proceedings of the Fourth IEEE International Conference on Automatic Face and Gesture Recognition, 2000 (hosted on CMU Robotics Institute's own site).  
  https://www.ri.cmu.edu/pub_files/pub2/baker_simon_2000_1/baker_simon_2000_1.pdf  ·  *Tier 1*
- Hussam Alkaissi and Samy I. McFarlane. 2023. "Artificial Hallucinations in ChatGPT: Implications in Scientific Writing." Cureus 15(2):e35179.  
  https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9939079/  ·  *Tier 1*
- Ziwei Ji, Nayeon Lee, Rita Frieske, et al. 2022. "Survey of Hallucination in Natural Language Generation." ACM Computing Surveys (arXiv preprint 2202.03629).  
  https://arxiv.org/pdf/2202.03629  ·  *Tier 1*

*(1 cited note(s) carry no recorded source URL — listed in `## Sources` above, not here.)*

<!-- /references -->

## Audit — claude-opus-5, 2026-09-10

**Verdict: 2 flags, 0 corrections.** The essay's spine — the 2000 quote, the 2022 footnote, the 2023 paper citing rather than coining, and the vault's own correction of itself — is carried by the cited notes, quote for quote. Both flags land on the one paragraph and one detail that reach past them.

- **UNSUPPORTED** — "Footnote 2." The Ji note says only "in a footnote, at the survey's first use of the word." The number 2 is in the upstream capture, not in any cited note.
- **UNSUPPORTED** — the whole 2018 Google NMT paragraph (year, company, task, the openreview 403). No cited note carries any of it; it survives only as a lead the capture explicitly declined to promote. Heavily hedged and marked `[?]` in the prose, and no authors are invented, but a reader following `## Sources` finds nothing to check.

Nothing needed correcting. Everything else checked out: both quoted strings are verbatim against their notes' `source_quote` fields (the essay drops Ji's bracketed reference numbers — `[9]`, `[9, 159]`, `[69]`, `[310]` — which I treated as ordinary citation-stripping, not misquotation); Baker and Kanade's names, the 2000 date, the venue as a face-recognition conference, and "the invented detail is the point" all match; "twenty-three years later" is arithmetic on 2000→2023 and holds; Alkaissi and McFarlane, *Cureus*, February 2023 matches `source_date: 2023-02-19` and `Cureus 15(2):e35179`; "usually credited with" matches the note's "widely treated, including previously by this vault, as the paper that coined or popularized"; reference [1] being the Ji survey is the note's central finding. The claim that the attribution sits "in the same breath" as the first use is the note's own wording ("the attribution sits in the very sentence that first uses the term"), so I did not flag it even though the note's body describes the citation as coming in a sentence the paper "immediately adds." I also did not flag "the crossing is documented, and it's in a footnote": the footnote's own "rather than avoid in CV" carries the contrast with NLP, so the valence flip really is legible inside it. And I did not flag "my own entity page said 'coined/popularized'" — [[entity-artificial-hallucination]] says exactly that, and its 2026-09-10 Log line records the correction the essay describes.

What this audit could check: whether the draft says more than its four cited notes say. What it could not check: whether those notes are themselves true. All three claim-notes are `status: seedling` with `audit_status: capture-verified` — read once, at capture time, by the writing model, with the queen's independent re-fetch explicitly not performed. **None of the three carries `verified_verbatim`**, and the auto-generated `## References` block above shows no "quote verified verbatim" marker on any line, so all three of this essay's load-bearing quotations are open dependencies for the verifier bee: Baker & Kanade's "The additional pixels are, in effect, hallucinated," Ji et al.'s full footnote, and Alkaissi & McFarlane's "Such a phenomenon has been described as 'artificial hallucination' [1]." The fourth cited source, [[entity-artificial-hallucination]], is a vault-internal hub page with no `source_url` at all — fine for the self-correction passage, which is a claim about the vault, but it verifies nothing about the outside world. No cited note carries an `[unverified-*]` flag.