A phantom term born from a 1950s typo has 'infected' neural networks: how an OCR bug became a marker of AI-generated content

A debate has erupted in the scientific community over the fictional term "vegetative electron microscopy," which has no physical or biological basis whatsoever, yet appears in at least 22 peer-reviewed scientific publications and, more alarmingly, is consistently reproduced by modern language models. According to Google Scholar, the phantom term has become so firmly embedded in English-language scientific discourse that automated text-checking systems now use it as one of the markers of content generated by neural networks. The story of this "ghost" in the scientific literature began long before the era of machine learning — back in the 1950s, when two articles from the journal Bacteriological Reviews were being digitized.

How a typo became a term: the birth of the phantom in the 1950s

The mechanism behind the anomaly turned out to be extremely simple and characteristic of the early era of optical character recognition (OCR). While digitizing two articles from Bacteriological Reviews, the algorithm accidentally "stitched together" the word "vegetative" from one column with the word "electron" from an adjacent column. As a result, a combined term appeared in digital databases that had never existed in the original printed editions. For decades this artifact went unnoticed: no researcher paid attention to the odd phrase, as it was buried in digitized archives to which the scientific community had limited access.

The Persian bridge: how the term entered international circulation

The first noticeable "resurrection" of the phantom term occurred in 2017 and 2019, when it began appearing in the English-language abstracts of Iranian scientific publications. The cause was a subtlety of the Persian (Farsi) language: the words corresponding to the Russian "vegetative" and "scanning" differ by only a single diacritical mark. A minor error in machine or even human translation launched the phrase into international scientific circulation, after which it began to be copied from article to article through the classic citation mechanism. To date, according to Google Scholar, the meaningless term has been recorded in at least 22 peer-reviewed scientific works.

Neural networks absorbed the bug: language model testing results

The researchers who discovered the anomaly decided to test how deeply it had penetrated the architecture of modern AI systems. Testing revealed a mixed picture: when fragments of the original articles were fed as context, the GPT-3 model automatically completed the sentence with the phrase "vegetative electron microscopy," demonstrating a persistent tendency to reproduce the phantom term. Meanwhile, earlier systems — GPT-2 and BERT — showed no such "memory." However, the persistence of the anomaly was confirmed in newer generations of models as well: both GPT-4o and Anthropic's Claude 3.5 also reproduce the fictional term in the appropriate context. The main channel of "infection" was the massive open web-data archive Common Crawl, on which the vast majority of modern neural networks are trained: it was through this archive that the 1950s OCR artifact ended up in the training samples.

Contradictory data

The available sources contain a number of nuances that require careful interpretation. First, the exact number of publications containing the phantom term is given as "no fewer than 22" — this is a lower bound according to Google Scholar, not exhaustive statistics; the actual scope may be wider, especially given non-indexed repositories and preprint versions. Second, the degree of "contamination" of the models is uneven: if GPT-2 and BERT do not reproduce the term, this does not mean they do not contain it in their embeddings — perhaps the activation threshold for these architectures is simply higher. Third, the precise reconstruction of which exactly two Bacteriological Reviews articles were digitized with the error and which specific OCR process was used in the 1950s is based on circumstantial evidence and linguistic analysis, rather than on direct archival documents from that period. These caveats do not cast doubt on the very fact of the anomaly's existence, but they indicate that the full chronology and statistics of its spread may be refined as further research is conducted.

Systemic vulnerability and a call for transparency

The case of "vegetative electron microscopy" vividly demonstrates the systemic vulnerability of AI algorithms to errors accumulated over decades of digitization. As long as automated text-checking tools can only detect anomalies that a human has already identified, the cycle of self-reproduction of fake data remains unclosed. Experts are calling on neural network developers to ensure the openness of training datasets, so that the scientific community can audit the sources of "contamination," and on publishers to strengthen peer review for phantom terms inherited from digitized archives. Without such measures, researchers warn, every new OCR bug or machine-translation error risks becoming yet another "ghost" that neural networks will confidently cite as an established scientific concept.