'Vegetative Electron Microscopy' Nonsense Phrase Contaminates Scientific Literature via AI (2020)

Incident Summary

The nonsense phrase “vegetative electron microscopy” — a term with no basis in any real scientific methodology — appeared in at least 22 scientific papers indexed on Google Scholar, undermining trust in parts of the scholarly record.^[1]

Investigations by Retraction Watch journalists and researchers Guillaume Cabanac (Université de Toulouse) and Cyril Labbé (Université Grenoble Alpes) traced the phrase’s origin to a 1950s paper in Bacteriological Reviews. When the paper was later digitized, OCR software confused the two-column layout, merging “vegetative” from the left column with “electron microscopy” from the right. The error entered digital databases, was reinforced by a Farsi near-homograph confusion (the Persian words for “vegetative” and “scanning” differ by a single dot), and was ultimately absorbed into AI training data.^[2]

Testing confirmed that GPT-3 consistently reproduced the phrase, and the error persists in GPT-4o and other current models. Researchers describe it as a “digital fossil” — an error now embedded in AI knowledge bases that is “nearly impossible to remove.” The phrase has become a recognized fingerprint for AI-generated or paper-mill-produced manuscripts, joining a list of approximately 4,000 “tortured phrases” tracked by Cabanac’s Problematic Paper Screener.^[3]

Key Facts

• Origin: 1950s OCR digitization error merging text across two columns of a Bacteriological Reviews paper
• Propagation chain: OCR error → digital databases → Farsi near-homograph confusion (2017–2019) → AI training data (GPT-3 onward) → paper mills and AI-assisted manuscript writing
• Scale: At least 22 scientific papers contain the phrase; one in a Springer Nature journal was subject to a contested retraction
• AI contamination: GPT-3 consistently generates the phrase; the error persists in GPT-4o and Claude 3.5
• Detection tools: Cabanac’s Problematic Paper Screener tracks approximately 4,000 similar “tortured phrases” across ~130 million articles weekly
• Related nonsense phrases: “counterfeit consciousness” (artificial intelligence), “bosom peril” (breast cancer risk), “kidney disappointment” (kidney failure)

Threat Patterns Involved

Primary: Misinformation and Hallucinated Content — A digitization error was absorbed into AI training data and reproduced in scientific manuscripts, creating a self-reinforcing cycle where AI-generated text contaminates the very sources future models are trained on.

Significance

1. Training data contamination loop — The incident demonstrates a concrete mechanism by which errors in digitized text propagate through AI training pipelines into generated outputs, which then re-enter the corpus as new publications, creating a self-reinforcing contamination cycle
2. Scientific integrity impact — The phrase’s presence in peer-reviewed journals published by Springer Nature and Elsevier reveals weaknesses in editorial screening processes, particularly as AI-assisted writing becomes more prevalent
3. Detection vs. decontamination asymmetry — While the phrase can be detected (and serves as a useful paper-mill fingerprint), removing it from AI training data is described as “nearly impossible,” highlighting a fundamental challenge in AI data quality
4. Broader pattern — The approximately 4,000 “tortured phrases” tracked by the Problematic Paper Screener suggest that “vegetative electron microscopy” is one visible example of a much larger AI-driven scientific integrity problem