The AI ‘Ghosts’ Contaminating Academic Publishing

A new preprint from Samsung and the University of Warsaw says large language models are not just inventing people, they are recycling them. The paper, “The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing,” tracks names like Elena Vasquez and Marcus Chen, which show up again and again as if they were real academics, co-authors, or experts. They’re not. They’re model habits turned into fake identities.

The researchers say the pattern goes beyond one-off hallucinations. Different models keep reaching for different names, and some of those names tend to travel together. Claude leans toward Elena Amara Okafor. Gemini keeps producing Aris Thorne and Lena Petrova. ChatGPT often lands on Elara Voss. The team says this creates “correlated character ensembles,” a phrase that sounds academic and slightly cursed in equal measure.

That matters because the names are not staying inside chat windows. The paper says Zenodo, the CERN-operated repository that issues real DataCite DOIs, contains 1,655 ghost-authored records with fake journal names and fabricated publication dates. Anyone with a free account can create a DOI there, and once the record exists, it can be picked up by scholarly aggregators. The paper says these synthetic papers are already appearing on ResearchGate, and being indexed without verification by Google Scholar and Semantic Scholar.

There’s also a feedback loop here. The same names that models spit out can be used to guess which model produced a piece of junk in the first place. Brzozowskim, the paper’s lead author, says Elena Vasquez was especially common in content made by Claude Sonnet 4. But that trick may not last long if AI-generated text keeps flooding the web and gets scraped back into training data. At that point, the ghosts stop being a bug and start becoming part of the source material.