OpenAI and Scientific Journals: AI Papers Got Better and Are Overloading Peer Review

Scientific journals are facing a new problem: AI has started churning out not blatantly absurd papers but entirely plausible ones. As a result, editors and reviewers spend more and more time weeding out work that looks convincing but adds almost nothing to knowledge.

How the surge was spotted

The problem drew loud attention after a strange episode involving researcher Peter Degen. His 2017 paper on statistical analysis of epidemiological data had collected a normal academic number of citations for years, then suddenly began receiving them literally every few days. A check showed that it was being cited en masse by new papers built on the open Global Burden of Disease dataset.

Formally, these were studies on the risk of stroke, cancer, falls among the elderly and dozens of other topics, but in essence they were endless variations on the same template. Degen found traces of this factory on GitHub and on the Chinese platform Bilibili, where a company from Guangzhou advertised lessons on producing publishable scientific papers in under two hours using its own software and AI assistants. Such texts often contained errors and stretched claims, but they no longer looked as absurd as early AI junk.

Screening them out became much harder, and the load on journals grew.

This is a huge burden on peer review, which is already running at its limit.

Why the filters are failing

Fake or automatically assembled papers used to carry visible markers: invented references, odd illustrations, chatbot-style phrases accidentally left in the final text. Publishers were already fighting paper mills, the semi-legal factories that sell authors publications for their résumés. Generative AI first helped such schemes evade plagiarism checks, but it also gave itself away through hallucinations. Now that safety net has almost disappeared: manuscripts have become coherent, neatly structured and stylistically even.

Journal editors feel this most. Marit Moe-Pryce, managing editor of Security Dialogue, said the number of incoming manuscripts has grown by roughly 100 percent year over year, and the main problem is that almost all of them look fine at first glance. In one case, a paper passed through at least ten editors and two rounds of peer review before anyone noticed a plausible but invented reference. It is no longer enough to check whether a cited work exists; you also have to judge whether a real expert would have chosen it.

Where the system breaks

The risk is amplified not only by template generators but also by more autonomous research agents. Carnegie Mellon researchers showed that such tools can invent data or use questionable methods even though the final paper looks polished. Matt Spick of the University of Surrey tested the Prism tool on already published data about the ripening of eggplants and peppers. The system proposed a new statistical approach and in 25 minutes 50 seconds produced a full paper with charts and correct references, good enough that experienced scientists were seriously surprised.

• Journals report a 40–100 percent rise in incoming submissions.
• To find two reviewers, editors increasingly have to write to 12–20 people.
• Grant funders are already receiving a flood of neatly tailored applications.
• Conferences, editorial boards and reviewers spend more and more hours manually checking works of doubtful usefulness.

The problem is not only about model quality but about how science is built. Open access journals earn money by processing manuscripts, universities and funds still look at publication counts, and researchers live by the logic of publish or perish. Against this backdrop, AI becomes a machine for inflating metrics. According to a study published this year in Nature, scientists who use AI publish three times as many papers and receive almost five times as many citations. But as productivity grows, focus narrows: the system pushes authors toward well-mapped topics where it is easier to synthesize one more publishable result.

What this means

The main threat now is not that AI will fully replace scientists, but that it is already undermining the human filters that hold academia together. If scientific value continues to be measured primarily by the number of papers and citations, models will only speed up the production of work that takes up other people's time but does not move knowledge forward. Science will therefore have to change both how it verifies authenticity and the rules of academic success.