A recent preprint published arXiv describes a potentially useful new signal for identifying AI-generated content: recurring combinations of apparently fictional names produced by large language models. The study reports that some models do not simply favor individual names when asked to generate fictional experts or researchers; they repeatedly produce particular pairs or groups. In controlled prompting, for example, one model family showed a substantial tendency to generate the same two names together, with the frequency changing across model versions. The authors then traced these name combinations into websites and purported academic records.
The implications become more significant when these synthetic identities move beyond generated prose and into systems that researchers routinely use to discover and assess scholarly work. The preprint reports hundreds of records on ResearchGate associated with apparently nonexistent researchers, as well as a much larger collection of records on Zenodo that claim publication in nonexistent journals. In the latter case, the study identifies 1,655 records carrying Zenodo DOIs and reports a striking discrepancy between user-supplied publication dates and later DOI registration timestamps. The reported concentration—991 registrations in March 2026 alone—suggests that the problem is not simply isolated hallucinated citations, but potentially automated production of metadata that can enter the scholarly discovery infrastructure.
For researchers, this raises a familiar but increasingly consequential distinction between the existence of a record and the validity of the research it purports to describe. DOI assignment, repository indexing, author profiles, and search-engine visibility are useful components of scholarly infrastructure, but none necessarily establishes that a claimed author, journal, institution, or study actually exists. The episode also illustrates a limitation of conventional AI-content detection: fabricated scholarly identities can persist even when the underlying text is not obviously machine-generated. Conversely, recurring names should be treated as a forensic clue rather than proof of fabrication, particularly because real people can share common names.
Looking ahead, scholarly infrastructure may need stronger provenance checks alongside existing metadata and indexing mechanisms. Repository systems and aggregators could potentially cross-reference claimed journals, affiliations, author identities, publication dates, and DOI relationships before records acquire the appearance of established scholarship. For researchers, the practical response is likely to be less about learning to recognize a particular set of “ghost names” than about reinforcing habits already central to rigorous literature assessment: verifying primary sources, checking author and institutional affiliations, examining DOI provenance, and distinguishing discoverability from validation. As generative systems become increasingly capable of producing not only papers but the surrounding apparatus of scholarly credibility, those checks may become an essential part of research practice.
Click here for the Japanese version.