The ghost authors

Ask Claude, Gemini, or ChatGPT to invent a researcher and the same names keep surfacing: Elena Vasquez, Aris Thorne, Elara Voss. That LLMs fall back on a small cast of fictional people is old news. What two researchers at Samsung's Warsaw lab and the University of Warsaw did was follow those ghosts off the open web and into real publishing infrastructure.

The phantom Elena Vasquez appeared on 77 of the 1,655 records — and nobody had searched for her name.

They searched Zenodo — the CERN-run repository that mints genuine, citable DOIs — not by author but by the names of journals the AI texts kept citing. The journals don't exist: no 'Journal of Functional Materials' or 'Journal of Computer Engineering' appears in Crossref. The query returned 1,655 records. The phantom Elena Vasquez was listed on 77 of them, even though nobody had searched for her — she simply fell out of the fake-journal haul.

The damning part is in the timestamps. Each record claims a publication date of 2020 to 2023, but Zenodo's DOIs carry an immutable server timestamp showing when they were actually registered: March and April 2026, 991 in March alone, roughly 25 a day for sixty straight days. The papers were generated now and backdated to look established — and the discrepancy is provable by anyone, because the registration date is queryable through the public DOI API.

Scholarly aggregators, citation indexes, and lit-review tools ingest DOI metadata on trust. These records are, in the authors' words, harvestable by any of them. At the observed pace, from a single repository found via just two fake journal titles, the contamination is a leak into the plumbing every researcher draws on — and it leaves a datable fingerprint, since which ghost names a document carries reveals which model wrote it.