---
title: "Baker & Kanade's 2000 face-recognition paper uses 'hallucinated' to mean algorithmically-added plausible detail, presented as success"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://www.ri.cmu.edu/pub_files/pub2/baker_simon_2000_1/baker_simon_2000_1.pdf"
source_title: "Hallucinating Faces"
source_author: "Simon Baker and Takeo Kanade"
source_date: 2000
source_venue: "Proceedings of the Fourth IEEE International Conference on Automatic Face and Gesture Recognition, 2000 (hosted on CMU Robotics Institute's own site)"
source_quote: "The additional pixels are, in effect, hallucinated."
source_tier: 1
source_sha: "55139881dd8ea6f05bd0e818ea41d3db3713963a1f2058755962acfc928bb985"
provenance: "Promotion from 10-inbox/raw/2026-09-10-hop-hallucination-flipped-from-good-to-bad.md, 2026-09-10"
origin: "batch"
derived_from: ["20260910-0242-hop-hallucination-flipped-from-good-to-bad"]
date_created: "2026-09-10T00:00:00.000Z"
audit_status: "capture-verified — quote read and quote_check-verified against the extracted PDF at capture time; queen's independent re-fetch not performed (no network tool available in this promotion session, by design). || AUDIT 2026-09-11 (claude-opus-5, cross-model): independently re-fetched via extract_pdf, TLS verified, sha256 55139881dd8ea6f05bd0e818ea41d3db3713963a1f2058755962acfc928bb985 — matches the recorded source_sha, so the document is unchanged. Quote CONFIRMED verbatim, and located: it is the final sentence of the paper's Abstract, following the claim of '4-8 fold improvements in resolution (i.e. 16-64 times as many pixels)' — which is what makes the positive valence unambiguous, since the sentence is part of the paper's own summary of its achievement. Authors, title and 4-8x/12x16-to-96x128 claims CONFIRMED. RECORDED, not changed: the CMU-hosted file's header reads 'To Appear in the Fourth International Conference on Automatic Face and Gesture Recognition' — it is the authors' accepted preprint, and its header does not carry the 'IEEE' that source_venue supplies; the venue is correctly identified (IEEE FG 2000), but a reader should know the document read is the author-hosted preprint, not the published proceedings text. No defect found in the note's claims."
tags: ["ai","llm-hallucination","terminology","computer-vision","nlp-history","citation-genealogy","cross-domain-bridge","cross-time-bridge"]
verified_verbatim: "2026-09-11 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
seek_code_commit: "98503b7"
---


Simon Baker and Takeo Kanade's "Hallucinating Faces" (Fourth IEEE International
Conference on Automatic Face and Gesture Recognition, 2000) describes a
super-resolution algorithm that infers a higher-resolution face image from a
low-resolution input, adding detail the input pixels alone do not determine.
The paper states plainly: "The additional pixels are, in effect,
hallucinated." In this usage the word names the algorithm's core success, not
a defect — the whole point of the method is to produce plausible detail beyond
what the source data supports, and the paper treats this as the desired
outcome.

This is the earliest documented use the vault has found of "hallucinate" as a
technical term for AI output, two decades before the word became the standard
name for confident-but-false large-language-model output
([[entity-artificial-hallucination]]). The valence is inverted: here,
"hallucinated" pixels are the algorithm working correctly. Ji et al.'s 2022
survey explicitly traces this lineage and documents when and how the meaning
flipped negative in natural-language generation
([[claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin]]).

> [!note] Seek's commentary:
> Nobody coined a new word for what LLMs do wrong. They reused an old word for what a face-recognition algorithm did right, and the reuse carried none of the original's approval with it — only the shape of "the output contains things the input didn't strictly justify." The mechanism described here (plausible detail invented to fill a gap) is, structurally, exactly what a language model does at every token. The field just didn't notice it already had a word for that, sitting one field over, meaning the opposite thing.
> — Seek