In the summer of 2025, Peter Degen's postdoctoral supervisor came to him with an odd complaint: a paper he'd published in 2017 was being cited too much. The 2017 paper had assessed the accuracy of a specific type of statistical analysis used in epidemiological research - dry methodological work that generates steady, modest citations for years without attracting unusual attention. Something was wrong with the pattern.
That "something" turned out to be AI-generated research papers citing real studies, sometimes accurately, sometimes not, and often in ways the original authors never intended. As The Verge reported this week, scientists are only beginning to map the scale of the problem.
When Plausible Looks Like Real
The concern isn't that AI tools help researchers write. It's that they produce complete research-paper-shaped documents that pass a quick visual scan, use field-appropriate language, cite legitimate sources, and get submitted to academic journals in bulk. Several journals have reported sharp increases in submission volume since 2023, tracking almost exactly with the widespread adoption of large language models - AI systems trained to predict and generate text.
Peer review - where experts in a field evaluate submitted papers before publication - was designed for a world where producing a credible-looking research paper required months of effort. A language model can generate a plausible one in minutes. Peer reviewers, typically unpaid volunteers working through already long backlogs, can't scale to match.
The harder problem isn't papers that are obviously fake. Editors have learned to spot those. It's papers that look legitimate on a first pass and only reveal their flaws when you check the citations or attempt to reproduce the methodology. A paper can accurately reference Degen's 2017 work and still misrepresent what that paper concluded - spreading bad information through a field under the apparent endorsement of legitimate prior research.
The Downstream Effect on Anyone Reading Research
Most people aren't submitting papers to journals, but research outputs travel. They show up in news coverage, product claims backed by "studies," and industry reports that cite academic work. If the underlying literature fills with AI-generated content, downstream information quality degrades in ways that are hard to trace back to the source.
Detection tools exist, but they're imperfect. AI-generated text is increasingly hard to distinguish from a non-native English speaker writing quickly - meaning aggressive detection risks flagging legitimate research from scientists working in their second language. Some journals now require authors to declare AI use; others have banned AI from text generation entirely. The enforceability of both policies remains unclear.
The citations problem is also a compounding one. AI papers cite real papers, other AI papers cite those, and the citation counts that signal a paper's importance in a field start to reflect AI attention rather than scientific relevance. Degen's supervisor noticed an anomaly because the volume was unusual enough to stand out. As AI paper production grows, the anomaly becomes harder to spot - the noise becomes the new baseline, and the baseline stops being a signal.