AI papers do not have to be right to waste science’s scarcest resource

The Verge frames the story through a strange signal: postdoctoral researcher Peter Degen encountered a paper that was being cited too much, and in the wrong way. In academia, citations usually act as reputation currency, but this 2017 paper on statistical analysis of epidemiological data began receiving hundreds of citations in a short window. That did not look like a natural surge of scholarly interest. It looked like the residue of mass-produced papers borrowing legitimate references to appear scientific.

The subject matter matters. According to the source brief, many of the citing papers analyzed the Global Burden of Disease dataset, a major research resource used in public health. The dataset itself is legitimate and valuable. That is exactly why it is attractive raw material for questionable papers: take a respected data source, apply a statistical frame, arrange the familiar sections of a journal article, and use a writing tool that can imitate academic tone.

This is not the same as saying every paper assisted by AI is fraudulent. The sharper problem is that generative tools lower the cost of producing papers that are superficially clean but weak, redundant or strategically misleading. The Verge also describes a Guangzhou-based company offering tutorials on producing publishable research in under two hours with AI writing assistance. That is not curiosity moving faster. It is a production line learning to sound scholarly.

Academic publishing was already dealing with paper mills: businesses and informal networks that manufacture articles for career credit, institutional pressure or publication quotas. The Committee on Publication Ethics treats paper mills as a systemic threat to publication integrity because editors and reviewers must distinguish flawed but genuine work from polished waste. AI does not merely increase the volume. It makes the boundary harder to see.

The real cost is not only a few bad papers entering databases. The cost is a peer-review system that is already stretched. Every hour an editor spends evaluating a semi-automated manuscript is an hour not spent on real science. If suspicious papers cite each other, citation metrics begin to measure a manufacturing network rather than the influence of ideas. If those papers use large public-health datasets, they can create a false sense of consensus around claims that may later bleed into policy, guidelines or public debate.

Transparency rules for AI writing help, but they are not enough on their own. Nature’s AI policy, for example, keeps responsibility with human authors and limits how AI tools can be credited in research outputs. The operational pressure is broader: journals need better anomaly signals, data checks, citation-network checks and editors with enough time to reject synthetic clutter. Generative AI did not invent academic noise. It made that noise faster, cheaper and clean enough to waste serious people’s time.