One in 458 papers now cites a source that does not exist
No single study established this trend. A biomedical audit found it first. A review of NeurIPS 2025 submissions found it separately. So did a documented case out of Hong Kong. Three groups, working from different material, arrived at the same conclusion: papers containing at least one fabricated reference are showing up more often, and the pace of the increase is not small.
In 2023 it was one paper in 2,828. By early 2026 the rate had climbed sixfold.
Three sources, one trend
A study covered by STAT News found the base rate of papers with at least one fabricated reference rising from one in 2,828 in 2023 to one in 458 in 2025, and to one in 277 during the first seven weeks of 2026. Separately, the NeurIPS analysis found over a hundred fabricated citations across fifty one accepted papers. Separately again, a paper published out of Hong Kong was found to contain twenty fabricated references out of sixty one, caught not by peer review but by a reader on social media after publication.
None of these three groups was working from the others' data. The trend is worth taking seriously precisely because it shows up three times, from three angles, without anyone coordinating the finding.
What checking at scale actually looks like
A citation reads as one fact but behaves like several: a title, a set of authors, a venue, a year, an identifier, all bundled together. A fabrication often gets some of these right while getting others wrong, a real sounding title attached to an identifier that points somewhere else, or a real journal that never published the specific article being claimed. Merlise checks the whole bundle against several scholarly indices rather than treating any single field as sufficient on its own.
This also means the system does not jump straight to "fabricated" the moment one index comes back empty. It tries alternate spellings, checks for a mismatched identifier, and only reaches a strong fabrication verdict when multiple independent checks agree that nothing resolves.
The evidence ledger
The same claim by claim view the product shows on a live document, built from this case.
How it resolves
The resolving record
The STAT News coverage of the Lancet linked base rate study, the GPTZero NeurIPS analysis, and independent reporting on the Hong Kong case together establish the trend from three unconnected sources.
STAT News, May 2026; GPTZero NeurIPS 2025 analysis; reporting on a Hong Kong retraction case