They are volcano experts, astronauts, therapists and podcast-ers. Their names repeatedly surface in over 1600 academic records.
Also, they don’t exist.
A new study has revealed that popular artificial intelligence models repeatedly generate the same fictional names when creating scientists, experts and other characters, leaving behind what researchers describe as identifiable “behavioral fingerprints” across the web.
Take for example an “author” named Lyra Emberlyn, who has single handedly authored 88 Amazon books including “The Alchemist’s Cipher”.
According to the pre-print study titled “The ghost Authors couple: Correlated LLM name priors and their haunting of the web and academic publishing”, available on arXiv, AI models do not simply invent fictional names. Instead they tend to produce recurring combinations of fictional identities. The study has not been peer reviewed yet.
“These names do not exist. Elena Vasquez and Marcus Chen have appeared as volcano experts, astronauts, thriller protagonists, podcast hosts, and academic co-authors across hundreds of independently produced AI-generated documents, never having lived,” the study mentions.
“We show that large language models do not merely default to high-probability individual names when generating fictional experts: they produce correlated character ensembles: pairs and trios whose co-occurrence rates far exceed chance and are consistent across independent generations,” it stated.
For example the study noted that certain versions of Claude frequently generated the fictional pair “Elena Vasquez” and “Marcus Chen”, while Google’s Gemini models preferred “Aris Thorne” and “Lena Petrova” and GPT based models produced “Elara Voss” with no fixed partner.
The study argued that the rapid increase of LLM-generated content on the web has raised questions about content provenance and authenticity.
During the research, the authors identified 1,655 records on Zenodo, a research repository operated by CERN, which as per the study listed nonexistent journals, fabricated publication dates and fictional authors while still receiving real DOIs registered in DataCite.
The researchers also claimed that 991 of these records were uploaded in March 2026 and stated that records could potentially be harvested by scholarly databases that ingest DOI metadata.
A DOI (Digital Object Identifier) is like a permanent identification number for scholarly work such as research papers or reports while DataCite is one of the organisations which officially issues and registers DOIs such as for research data sets, repository records etc.
The study also disclosed that 436 ResearchGate records have recurring fictional author names.
“Ghost Authors names additionally appear on ResearchGate forming synthetic research groups with collaborators drawn from multiple model families; publication dates on these records provide a reliable temporal proxy for model deployment windows,” the study claimed.
It highlighted that the rapid increase of LLM-generated content on the web has raised urgent questions about content provenance and authenticity.
Meanwhile the researchers tested multiple public versions of Claude, GPT and Gemini using several prompts requesting fictional experts, researchers, or protagonists.
Across repeated generations they found that certain names appeared at rates far higher than would be accepted by chance.
The researchers also noted that newer Claude releases showed a gradual decline in the recurring names, suggesting that later model updates might have reduced the behaviour.
“In every case we document, our search found no individual with the stated name active in the stated field at the stated affiliation: the expertise, institution, and name co-occur only in AI-generated content,” the study maintained.
The study was authored by Michał Brzozowski of Samsung AI Center Warsaw, and Neo Christopher Chung of Samsung AI Center Warsaw and the University of Warsaw.
Cover Pic Credit: Better Images of AI, Leo Lau & Digit
Also Read: ACL 2026 Rejects Over 100 Accepted Papers for Hallucinated Citations
