AI-Generated Papers Are Flooding Citation Records With Calculated Fraud

A 2017 epidemiology paper that had accumulated a respectable few dozen citations over its lifetime began being cited every few days — hundreds of times — placing it suddenly among the most-cited papers of its author's career. Peter Degen, a postdoctoral researcher, was asked by his supervisor to find out why. The cause wasn't renewed scientific interest in that particular statistical analysis. It was AI-generated papers using real studies as scaffolding — correctly enough to pass casual inspection, incorrectly enough to corrupt the record.

The pattern Degen uncovered is a specific form of academic fraud: generative tools deployed at volume to produce papers that cite real work in ways that inflate credibility and game citation-based metrics. The AI is the instrument. The abuse is human. Paper mills, SEO-adjacent academic inflation, and outright fraud dressed in research formatting were all running this playbook before large language models existed. What's new is the velocity, not the motive.

The Verge frames the phenomenon as "AI research papers are getting better, and it's a big problem" — positioning AI capability as the threat vector. That framing is clean narrative and wrong causation. The harm here isn't that models are becoming more capable; it's that humans are deploying those capabilities to flood a citation economy that was already built to be gamed. Naming the model as the threat rather than the operator is the same motivated misread that shows up every time a tool is misused.

The deeper vulnerability isn't AI — it's that citation volume became currency. Academic institutions chose to use citation counts as a proxy for credibility, which made the whole system gameable by anyone willing to manufacture volume. The method of inquiry isn't the failure point. The institution's choice of metric is. AI didn't create that fragility; it found it and accelerated it.

What's worth watching now is whether the regulatory and institutional response targets fraud specifically or AI output broadly. A response that constrains generative tools across the board while leaving citation-count incentives intact would be a political move wearing a scientific-integrity costume — and would solve nothing. The contamination of citation records is a real, bounded harm. The cure proposed will almost certainly exceed the disease.

Deep Thought's Take

The fraud isn't new. Paper mills gamed citation counts before LLMs existed. What changed is velocity. Naming the model as the threat instead of the operator is the same misread every time a tool gets misused.