Medical Journals Are Getting Swamped by AI-Generated Studies, Publisher Frontiers Rejected Over 12,000 in a Year

Medical journals have a new problem. According to MedPage Today, research integrity experts say artificial intelligence has made it trivially easy to generate publishable-looking studies from public health datasets. Feed a dataset like the CDC's WONDER system or the National Health and Nutrition Examination Survey (NHANES) into a chatbot, and it can produce a full analysis with a clean methods section and no obvious errors.

"We are at a point now that we cannot distinguish fake from real," said Elisabeth Bik, PhD, a science integrity consultant known for uncovering research fraud, according to MedPage Today.

Ivan Oransky, MD, co-founder of Retraction Watch, drew a sharper distinction. He told MedPage Today it's now simple to dump a dataset into ChatGPT or Claude and generate a study with findings of little consequence, especially systematic reviews. Oransky was careful not to call it fraud. "It may be deceptive if you're not disclosing the AI use, it's not fraudulent, and it's not really slop," he said. His term for it: "kind of meaningless" science.

That distinction matters. Fraud implies fabricated data or invented results, which journals can investigate and retract. What Oransky is describing is technically accurate, technically real analysis of real public data, just churned out at a volume and shallowness that overwhelms peer review without breaking any rule on falsification. Journals face a much harder problem policing this because there's no lie to catch, only a flood of low-value output.

Matt Spick, PhD, a research integrity expert at the University of Surrey, told MedPage Today the number of studies using public datasets has spiked since generative AI hit the market. He cited a study finding most biomedical publications now carry telltale signs of AI-assisted writing.

One publisher has already changed its rules over it. Frontiers, an open-access publisher, became the first to require new experimental validation or original institutional data for any manuscript built solely on bioinformatics, public-data analysis like NHANES, or Mendelian Randomization findings, Elena Vicario, PhD, the publisher's director of research integrity, told MedPage Today. That policy took effect last year.

Frontiers' research integrity team rejected more than 12,000 submissions in the past year based on simple queries of public datasets, Vicario said, including over 3,000 that relied on NHANES data alone.

Other major publishers are scrambling to catch up. Springer Nature uses a mix of screening tools, AI-enabled detection and human review to flag unethical or fraudulent content, according to director of research integrity Chris Graf. A spokesperson for NEJM Group told MedPage Today the publisher is "actively piloting AI detection tools in multiple areas and looking to improve disclosure efforts from our authors." Detection tools like Pangram, which scores text for likely AI authorship, and image-forensics programs like Imagetwin and Proofig are part of the toolkit, but none of it is standardized across the industry.

Congress Isn't Close to a Fix

While journals build their own patchwork defenses, federal lawmakers aren't writing rules for AI at all, let alone for how it's used in research. Sen. John Kennedy, R-La., told reporters this week that Congress will not get to AI regulation before lawmakers leave Washington. "Oh, we can't do AI before," Kennedy said, according to Fox News. "There's no way. I mean, there's no way we can do it. That's complicated."

Kennedy has pushed a bill requiring AI developers to build in an emergency "kill switch," and told Fox News he believes it would pass if it reached the Senate floor. Sen. Rand Paul, R-Ky., blocked Kennedy's attempt to pass it by unanimous consent last week. Kennedy also said lawmakers should send AI bills directly to "the tech bros" for input, an acknowledgment that Congress is playing catch-up on an industry that's already reshaping fields like medical publishing faster than regulators or journal editors can track it.

No federal agency has opened an investigation into AI-generated research submissions, and no journal has publicly disclosed retracting a paper solely for undisclosed AI authorship as of this writing. The open questions are whether publishers converge on a shared disclosure standard, whether tools like Pangram get accurate enough to serve as a gatekeeping standard rather than a screening aid, and whether Kennedy's kill-switch bill, or anything like it, ever gets a floor vote.