- AI tools can now mass-produce plausible scientific papers from real data.
- A nonsensical phrase in 20+ papers reveals how AI propagates errors.
- AI-generated microscopy images fooled 250 scientists in controlled testing.
Sabine Hossenfelder has been watching the scientific literature fill with noise for years. In a recent video, the physicist traced how generative AI is accelerating a crisis that includes everything from mass-produced papers to the bizarre case of "vegetative electron microscopy," a nonsensical term now embedded in dozens of published studies.
The scale is hard to dismiss.
In January 2025, a group of economists showed that GPT and Claude could analyze stock data and generate 288 complete finance papers following a reasonable analytical protocol. These were not gibberish - the economists ran real data through real methods. The same approach could work in cosmology or particle physics.
Key figure
288
Complete finance papers generated by AI in a single study using real stock data
How Vegetative Electron Microscopy Became a Digital Fossil
The term has appeared in more than 20 published scientific papers, including work in journals from Springer Nature and Elsevier. It means nothing.
The phrase traces back to a 1959 article where "vegetative" appeared in one column and "electron microscopy" in the adjacent column. An optical character recognition scan merged them. The error spread through AI training data, compounded by the fact that in Farsi, the words for "scanning" and "vegetative" differ by a single dot.
Queensland University of Technology researchers have since coined the term "digital fossil" for errors like this, embedded so deeply in language model training data that they may be permanently irremovable.
The future of science isn't about finding truth, just about generating statistically plausible sentences about it.
Sabine Hossenfelder, physicist and science communicator
When Even Scientific Images Become Unreliable
The problem extends beyond text.
A 2025 study published in Nature Nanotechnology demonstrated that AI tools can produce atomic force and electron microscopy images of nanomaterials that 250 scientists could not distinguish from real ones. The researchers even generated images of fictional "nano-cheetos" that looked entirely credible.
Image integrity consultant Mike Rossner noted that AI-generated scientific images lack the telltale signs of traditional Photoshop manipulation. Automated screening tools like Proofig AI, now used by Springer Nature and the Science family of journals, represent perhaps the only realistic defense.
AI-generated papers appear most frequently in fish-related research, though nobody has explained why.
What is a digital fossil?
A "digital fossil" is an error preserved in AI training data that resurfaces unpredictably in language model outputs. Unlike human errors that can be corrected at the source, digital fossils persist because they are distributed across massive datasets used to train multiple AI systems.
Corruption Beyond the Page
The integrity crisis reaches beyond papers and images.
Hossenfelder described how PubPeer, a platform where researchers flag problematic studies, has itself become a target. Researcher Sylvain Bernes received a message demanding he help retract a rival's papers, followed by a threat to post false complaints about his own work. The tactic resembles leaving fake one-star reviews to sabotage competitors.
AI and Science Papers
The Paper Mill Problem: Science's Fraud Industry Is Growing Faster Than Peer Review Can Handle
Paper mills, predatory journals, and AI-generated content are overwhelming peer review systems. Global retractions exceeded 10,000 in 2023, with Hindawi alone withdrawing over 8,000 articles.
→Citation gaming compounds the problem. A cat acquired academic citations through fake papers uploaded to ResearchGate and indexed by Google Scholar. In a separate incident, ResearchGate recommended that Geoffrey Hinton read a paper it claimed he co-authored with Yann LeCun. Neither had written it.
A March 2026 Northwestern University study found that organized fraud networks involving paper mills, brokers, and compromised journals now produce fake research at industrial scale.
This crisis might force a reckoning with academia's "publish or perish" culture.
What is clear is that the tools for generating plausible science now far outpace the tools for detecting it.
Sources
- Primary Source: AI Slop Is Spreading In Science, Too (Sabine Hossenfelder, YouTube)
- Additional Context:
- Vegetative electron microscopy as paper-mill fingerprint (Retraction Watch)
- Fake microscopy images indistinguishable from real (Chemistry World)
- Organized fraud networks spreading faster than real science (ScienceDaily / Northwestern University)
Fact Check: Claim-by-Claim Verification Verified
The article accurately reports on verified cases of AI-generated content and errors in scientific literature, matching primary sources like Sabine Hossenfelder's video and Retraction Watch.
Commentary
- "Digital fossil" term specifically coined by Queensland University of Technology researchers; article attributes correctly but lacks direct peer-reviewed source here.
- AI papers frequent in fish research noted in Hossenfelder's video without explanation, as stated.
- Proofig AI (likely "Proofig") used by Springer Nature and Science journals for image screening, per video context.
Sources used for verification
Academic/Peer-reviewed:
Other reliable sources:
Fact-checked by Perplexity Sonar Pro on 2026-03-10
