When AI Tricks the Science World: Hundreds of Biomedical Papers Pack Fake References

It sounds like something out of a sci‑fi thriller: computers whispering made‑up studies into the ears of researchers, and those phantom papers end up on the reference list of real‑world biomedical articles. That’s exactly what a new investigation has found – over 2,800 published papers contain citations that simply don’t exist.

The detectives behind the discovery weren’t detectives at all, but a team of data‑scientists and librarians who fed the entire PubMed database into a custom‑built algorithm. The program scanned every reference, cross‑checked titles, authors and journal details, and flagged any that couldn’t be matched to a genuine source. The result? A startlingly large collection of “ghost citations.”

Why are these fake references appearing in the first place? The researchers point to the rapid adoption of large language models (LLMs) – the same AI behind chatbots that can draft entire sections of a manuscript. When authors use these tools for help with literature reviews, the AI sometimes fabricates plausible‑looking references to fill gaps, especially when it can’t locate a real article that matches the query. The software thinks it’s being helpful; the author, often under time pressure, may not notice the mismatch.

What’s worrying is that many of the affected papers were published in reputable journals, spanning topics from oncology to neurology. Some of the bogus citations referenced articles that never existed, while others pointed to completely unrelated works. In a few cases, the invented sources even listed nonexistent DOI numbers, making them impossible to verify.

Journals are now scrambling to plug the loophole. Some have started to require authors to submit a CSV file of every reference, which will be automatically cross‑checked against bibliographic databases before peer review even begins. Others are updating their author guidelines to warn against the unchecked use of AI for literature searches.

For scientists, the takeaway is clear but unsettling: treat AI‑generated content as a draft, not a final product. Double‑check every citation, especially when a reference looks unfamiliar or the DOI doesn’t resolve. And perhaps most importantly, keep a healthy dose of skepticism about the convenience AI promises.

In the meantime, the study’s authors hope their findings will spark a broader conversation about responsible AI use in research. After all, the credibility of science rests on the tiny building blocks of each reference – and those blocks need to be solid, not imagined.