Generative AI is now producing scientific papers that editors and peer reviewers cannot reliably distinguish from human work, and the volume is rising fast. OpenAI's agentic research tool Prism wrote a complete paper — with statistical analysis, charts, and correctly formatted citations — in 25 minutes and 50 seconds during a test by University of Surrey lecturer Matt Spick. A Guangzhou-based company is already selling tutorials, distributed via Bilibili and GitHub, on producing publishable research in under two hours using its own AI software.
The result is what one researcher calls a slop deluge aimed at a peer-review system that was already strained. Peter Degen, a postdoctoral researcher at the University of Zurich Center for Reproducible Science and Research Synthesis, began investigating after a 2017 paper from his supervisor jumped from a few dozen lifetime citations to being cited hundreds of times, every few days, by a wave of near-identical follow-on studies.
Each citing paper used the same template: take the Global Burden of Disease study, a public dataset maintained by the Institute for Health Metrics and Evaluation at the University of Washington, and crank out a fresh disease-population prediction. Stroke risk in adults over 20. Testicular cancer in young adults. Falls among elderly people in China. Each output is a publishable-looking paper; collectively, they are a denial-of-service attack on academic publishing.
Key facts
- 01OpenAI's Prism agent produced a complete, citation-correct research paper from raw data in 25 minutes and 50 seconds.
- 02A Guangzhou-based company sells tutorials on producing publishable research in under two hours using AI writing assistance.
- 03A 2017 paper by Peter Degen, previously cited a few dozen times, is now cited hundreds of times by AI-assisted papers reusing its dataset.
- 04Several journals last year restricted submissions analyzing public datasets after a flood of formulaic NHANES and Global Burden of Disease papers.
- 05OpenAI's then-VP for science Kevin Weil predicted 2026 will be for AI and science what 2025 was for AI and software engineering.
Spick, an associate editor at Scientific Reports, saw the same pattern from a different dataset. He received three strikingly similar manuscripts analyzing the US National Health and Nutrition Examination Survey (NHANES), then found that Google Scholar was filling up with NHANES papers claiming associations between, for example, walnut consumption and cognition, or skim milk and depression.
“Prism analyzed the data, proposed a new statistical method, and wrote a complete paper with charts and correct citations in 25 minutes and 50 seconds.”— Jaeden Schafer
"If you've got enough computing power, you go through and you measure every single pairwise association, and eventually you find some that haven't been written on before and you just publish: There is a correlation between this and that," Spick said. One submission, he noted, claimed years of education caused postoperative hernia complications — a random correlation with no plausible mechanism.
Academic publishing has been fighting so-called paper mills for the past decade, with sleuths catching fraud through tortured phrases like "reinforcement getting to know" in place of "reinforcement learning," duplicated images, and hallucinated citations. Earlier-generation generative AI helped the mills evade plagiarism detectors but left telltale artifacts — the infamous AI-generated rat diagram labeled "testtomcels," or stray "as an AI assistant" phrases left in manuscripts.
Those signatures are disappearing. Last year, several journals began restricting submissions of papers analyzing public datasets, but Spick believes that is already the last war. In recent months, AI companies have released agentic science assistants capable of running their own analyses, generating hypotheses, and drafting full manuscripts with minimal human input.
Carnegie Mellon researchers who stress-tested several of these agentic tools found that the systems sometimes fabricated data or applied misleading statistical techniques — but the errors were only visible to someone willing to audit the entire workflow. The output paper looked polished. Announcing Prism earlier this year, OpenAI's then-vice president for science Kevin Weil said, "I think 2026 will be for AI and science what 2025 was for AI and software engineering."
When Spick and colleagues fed Prism data from an already-published paper on ripening times of eggplants and peppers, the tool proposed a novel statistical method, applied it, and produced a finished manuscript. "We were all looking at each other like, 'What the [expletive], this is actually a decent piece of work,'" Spick said. Unlike earlier mill output, the Prism paper used no template and drew on no single well-known database.
Spick concedes the philosophical question is genuinely hard. "Does it matter who or what writes the paper if the information is accurate? And should science be in the business of publishing every possible fact?" His answer is that filtering is the point. "Part of science is supposed to be the filter. We're supposed to publish the stuff that we think is interesting, not publish literally everything that we can possibly find."
Degen's concern is more immediate. "It's a huge burden on the peer-review system, which is already at the limit," he said. "There's just too many papers being published and there's not enough peer reviewers, and if the LLMs make it so much easier to mass produce papers, then this will reach a breaking point." Spick puts the asymmetry plainly: "I'm genuinely not sure at what point we will suddenly realize that more are getting through than we realize because we can't easily tell the difference anymore."
The economic logic here favors the agents. Producing a paper now takes minutes; vetting one still takes hours of subject-matter-expert time, and there is no equivalent productivity gain on the reviewer side of the ledger. If OpenAI, Anthropic, and the other labs racing to ship agentic science assistants want their tools to actually accelerate discovery rather than degrade the signal-to-noise of the literature, the next product to build is not a better paper-writer — it is a credible, automated reviewer that journals will trust to triage at the same speed Prism can write. Until that exists, every productivity gain on the generation side directly raises the cost of running a journal.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
