Does scientific fraud operate through organized networks or individual actors?

Richardson et al. argue that scientific fraud at scale is organized as cooperating networks, not as the work of lone defectors. Their case studies go past fake papers to what they call "brokerage roles in a widespread network of editors and authors." One group of editors in PLOS ONE, affiliated with institutions in four countries, "sent most of their submissions to one another over other editors," and more than half of the articles that group accepted have been retracted under nearly identical notices about "authorship, competing interests, and peer review." A network of 2,213 articles flagged for duplicate images, linked by 4,188 duplication observations, splits into 20 connected components, and only 34.1% of those articles have been retracted. The authors also describe an entity, ARDA, that advertises "Conferences and Meetings," "Journal Publication" and "Thesis/Article Writing."

The proposed mechanism is a working hypothesis about batch production. If paper mills publish "in large batches," papers in a batch "could use a fixed bank of images," and each batch should appear as a module in the image-sharing network, concentrated in particular publishers and years. The same model predicts "journal hopping": when a journal is deindexed by WoS or Scopus, the mill's clientele stops crediting it, so the mill moves on. The authors call this "adaptive behavior," but they document it in one entity. Across the five ARDA journals they inspected, between 34.0% and 98.7% of articles fell outside the journal's stated scope, and 10.1% of articles had authors from multiple countries, which the authors read as support for mills selling authorship slots.

Against the nearest notes, the picture sharpens. Do language models leak their training through fictional names? treats 1,655 ghost-authored DOIs as evidence that generated names have reached scholarly infrastructure. This source describes the channel through which fabricated papers reach indexed journals, though the excerpt never connects the two. The conclusion is blunter about AI: large language models and "machine scientists" are "not yet able to distinguish quality science from poor quality or fraudulent science," a task that "only becomes more difficult as the number of fraudulent scientific publications increases." That places the failure earlier than the reviewer-rule question in Does banning LLM use in peer review change review outcomes?. The ICML trial measures whether reviewers obey LLM rules. Richardson et al. describe editors and conference organizers, small teams who "could also accept an excessive number of problematic articles."

The excerpt does not establish the title's scale claim. Its growth section, "Scientific Fraud Is Growing Much Faster than the Scientific Enterprise As a Whole," has no body text here, so no growth rate or comparison with legitimate science is shown. The counts are also floors. The authors say their "analyses rely on the instances of scientific fraud that have been reported," that "the vast majority of paper mill products have not been detected," and that apparent growth may reflect shifting detection effort. The batch and journal-hopping models remain hypotheses, and the module test is described without its result. The defensible reading is that coordinated networks are documented in case studies, while the claim of rapid, large-scale growth is still open.