# Pseudo-lexical labels Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: pseudo-lexical-label - Category: Semantic (semantic) - Subcategory: non-language - Also known as: diagram non-words, text-shaped strings that are not words - Status: Active. Currently signals low-effort writing. - Severity: high - Evidence grade: corroborated (named independently by multiple credible secondary sources) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/pseudo-lexical-label ## Description Strings shaped like technical vocabulary that are not language. They have the right length, the right suffixes and the right position in a diagram, and they are not words. This index is otherwise text-only, and this entry is its single declared exception, because the strings are legible text carried inside an image. ## Why it reads as machine-written An image generator producing letterforms is drawing the texture of technical writing rather than spelling anything. The published case is a paper retracted three days after publication in February 2024, whose figures carried labels that looked like signalling-pathway vocabulary and were not words in any language. The authors had disclosed the tool in the text. It passed review anyway, which is the part worth remembering: the labels were visible to anyone who looked at the figure. ## Detection Type: judge (a rubric for a lightweight model judge) Rubric for a lightweight model judge: ```text Task: read every string inside figures, diagrams, charts, screenshots and generated illustrations, including axis labels, legends, node names and watermarks. Step 1. Transcribe each token exactly as rendered. Step 2. Classify each token as one of: dictionary word; known technical term; proper noun; abbreviation whose expansion appears in the surrounding text; identifier such as a gene, chemical, part or product code; none of these. Step 3. Tokens landing in "none of these" are pseudo-lexical candidates. Check the near-miss cases: a real word with one letter wrong is a typographic error, while a string with plausible morphology and no lemma behind it is the pattern. Escape hatches: optical character recognition of a scanned page, where errors cluster and follow the scan quality; font substitution failure, which mangles a whole block uniformly rather than one label; deliberate nonsense, placeholder text and lorem ipsum; a language or script the reviewer does not read, which returns UNKNOWN; hand-drawn figures photographed at low resolution. Decision. FLAG when two or more pseudo-lexical tokens appear in one figure, or when one appears in a figure title or a heading. Otherwise PASS. Output: the tokens, their classification, and the verdict. Note that this check reads images, which the rest of this index does not. ``` ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > Figure 2 labels the three stages of the pipeline as Preproccenting Layer, Feature Extructation and Ouput Nomalisation, and the caption calls the diagram self-explanatory. After, repaired: > Figure 2 labels the three stages as tokenisation, feature extraction and normalisation. The caption names who drew it and from which run of the data, and the alt text repeats the three labels so a screen reader gets the same words a sighted reader does. Note: The strings in the first version are invented for this specimen, in the shape the failure takes. The repair also fixes the accessibility hole that let the labels go unread. ## False positives Who legitimately writes this way. Scanning and font handling produce identical strings for entirely mechanical reasons. Optical character recognition on a poor scan invents morphology that looks exactly like this, and a missing font substituted at print time can turn a whole diagram into plausible-looking rubbish. Both are print-production failures with nothing to do with how the text was written. Specialist identifiers also defeat a naive reader: gene names, chemical registry numbers, part codes and internal build labels are not dictionary words and never were. Reviewers outside the field should record what they saw and ask the authors rather than score the figure, and any check that reads images should say so, since it is a different kind of evidence from reading prose. ## Model attribution Image generators rather than text models. The retracted paper disclosed the image tool in its own text, which is unusual and makes the attribution unusually solid for this dataset. ## Sources 1. Science Integrity Digest, the Frontiers rat figure retraction https://scienceintegritydigest.com/2024/02/15/the-rat-with-the-big-balls-and-enormous-penis-how-frontiers-published-a-paper-with-botched-ai-generated-images/ (tier: press; accessed 2026-08-14) 2. Gizmodo, science journal retracts AI-generated images https://gizmodo.com/science-journal-rat-dck-ai-generated-images-retracted-1851297606 (tier: press; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Active: A journal retraction three days after publication, covered independently, with the mangled figure labels reproduced in the coverage. The entry is kept despite falling outside the text-only scope, and it declares that exception in its own description. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.