AI “digital fossils” are already polluting the world. They seem authentic and complex, but in fact, they don’t make sense. Kind of like this AI-generated image.
In 2023 and 2024, as AI text generators started to become mainstream, a curious trend emerged: the word “delve” began appearing in a suspicious number of science papers. It became a kind of calling card for AI-generated content — but it’s far from the weirdest one.
Let us introduce you to: “vegetative electron microscopy.”
Vegetative what?
If you know basic science, you’re already raising an eyebrow. “Vegetative electron microscopy” doesn’t make sense — and that’s because it isn’t a real thing. It’s what researchers call a “digital fossil” — a strange, erroneous term born from a mix of optical scanning errors and an apparent translation mistake, then preserved in the data used to train artificial intelligence.
Remarkably, this nonsense phrase appears to have emerged independently through two unrelated errors.
Back in the 1950s, two papers in the journal Bacteriological Reviews were scanned and digitized. In one of them, the word “vegetative” appeared in one column and “electron microscopy” in the adjacent one. The OCR software mistakenly merged the two — and so, the fossil was born.
Then, in 2017 and 2019, two papers used the term again. Here, this appears to be a translation error. In Farsi, the words for “vegetative” and “scanning” differ by only a single dot. So instead of scanning electron microscopy, you got vegetative electron microscopy.
All of this came to light thanks to a detailed investigation by Retraction Watch in February. But this wasn’t the end of the story.
Why this matters
You’d think this weird glitch wouldn’t matter — but it turns out, it kind of does.
×
Thank you! One more thing…
Please check your inbox and confirm your subscription.
The term has now appeared in at least 22 different papers. Some have been corrected or retracted, but by then, the damage was done. Even El País, one of Spain’s leading newspapers, quoted it in a story in 2023.
Why? Blame AI.
Modern AI systems are trained on vast troves of data — essentially everything they can scrape. Once “vegetative electron microscopy” appeared in several published sources, the AI models treated it like a legitimate term. So when researchers asked these systems to help write or draft papers, the models sometimes spat it out, blissfully unaware that it was gibberish.
According to Aaron J. Snoswell and colleagues, who published a deep dive on The Conversation, the term began polluting the AI knowledge pool after 2020 — after those two problematic Farsi translations. And it’s not just a one-time fluke: the error persists in large models like GPT-4o and Claude 3.5.
“We also found the error persists in later models including GPT-4o and Anthropic’s Claude 3.5,” the group write in a post on The Conversation. “This suggests the nonsense term may now be permanently embedded in AI knowledge bases.”
“Permanently” may be too strong in a literal sense. AI developers can filter outputs, revise training sets and replace old models. But correcting an error buried inside billions or trillions of learned statistical relationships is much harder than fixing a line in a conventional database. Removing the original phrase from the internet wouldn’t necessarily remove the association from models already trained on it.
AI-assisted scientific writing is no longer a marginal phenomenon
This bizarre example is more than a fun anecdote — it highlights real risks.
Since the vegetative-electron-microscopy story emerged, researchers have begun measuring AI’s influence across entire bodies of scientific literature.
In 2025, researchers analyzed more than 15 million biomedical abstracts indexed in PubMed. They tracked vocabulary that became unusually common after the release of ChatGPT and estimated that at least 13.5 percent of abstracts published in 2024 had been processed with a large language model. Rates varied widely between fields, journals and countries, sometimes reaching about 40 percent.
Another study, published in Nature Human Behaviour, examined more than 1.1 million papers and preprints. By September 2024, the researchers estimated that large language models had modified 22.5 percent of computer-science abstract sentences and 19.6 percent of introduction sentences. The estimated rates were lower in mathematics and Nature Portfolio journals but had still risen sharply after ChatGPT’s release.
In 2026, an analysis of 7.3 million articles published between 2020 and 2025 found evidence of some degree of LLM influence in slightly more than half of the papers from 2025. That did not mean that half of the papers were wholly written by chatbots. The category included everything from light linguistic editing to much more extensive generation.
Using AI to improve grammar or phrasing isn’t scientific misconduct. For many researchers, especially those writing in a second language, these tools can make papers clearer and more accessible. The problem arises when authors use them without checking the output, allow them to invent references or scientific claims, or conceal the extent to which a machine produced the paper’s reasoning.
An AI Arms Race
Researchers are trying to fight this and detect this sort of issue. The Problematic Paper Screener, for instance, is an automated tool that combs through 130 million articles every week. It uses nine detectors searching for new instances of known fingerprints or improper use of AI. They found 78 papers in Springer Nature’s Environmental Science and Pollution Research alone.
But it’s an uphill battle.
There’s already so much AI content everywhere that it’s almost becoming virtually impossible to detect it; and that’s just one part of the problem. Scientific journals are another problem.
Journals have every incentive to protect their reputation and avoid retractions, even if it means defending dubious content. Case in point: Elsevier initially tried to justify the use of “vegetative electron microscopy” before ultimately issuing a correction. They ultimately issued a correction but the response is telling.
The problem is that as long as tech companies aren’t transparent about their training data and methods, researchers have to play detective and look for AI needles in the publishing haystack. According to one estimate, there are close to 3 million papers published a year, and the use of AI in writing is becoming more and more common.
The real danger is that these kinds of accidental errors can become entrenched in our scientific record — and once embedded, AI systems will keep repeating them. Knowledge is incremental, and if we build on wrong foundations, the consequences can be severe.
Ultimately, it seems even nonsense, once digitized and published, can become immortal.