Guillaume Cabanac got a strange notification earlier this year. Google Scholar told him his work had been cited in a dental journal. The problem? His research on spotting fabricated papers has nothing to do with dentistry. The citation title looked vaguely like one of his preprints, but the journal was listed as Nature and the DOI led nowhere. Cabanac, a computer scientist at the University of Toulouse, immediately suspected AI hallucination. He was right to be worried.
A Nature analysis suggests tens of thousands of publications from 2025 probably contain invalid AI-generated references. Surveys of computer science conferences found that 2-6% of papers now include hallucinated citations, up sharply from previous years. Alison Johnston, a political scientist at Oregon State University and co-lead editor of the Review of International Political Economy, rejected 25% of roughly 100 submissions in January because of fake references. "We're going to see a flood of fake references," she told Nature.
These aren't typos. They're complete fabrications. Mohammad Hosseini, who studies research ethics at Northwestern University, says the problem has shifted from minor inaccuracies to active fabrication. Joe Shockman, co-founder of Grounded AI, calls them "Frankenstein" citations because AI stitches together fragments of real publications into references that look authentic but don't exist. One study found that when GPT-4o generated literature reviews, nearly 20% of citations were fabricated.
For a junior researcher trying to build on existing work, a phantom citation can waste months. They might hunt down a paper that doesn't exist, or worse, cite it themselves and spread the error further. In fast-moving fields where researchers race to publish, bad references compound quickly.
This creates a nasty feedback loop. Future AI models trained on scientific literature will ingest these fake references, treating them as real. Researchers call this "model collapse," where errors amplify rather than correct themselves. Publishers including Springer Nature, Elsevier, and Frontiers are developing screening tools, but the volume of contaminated submissions may overwhelm current quality controls.