Title: Phantom Pairs: How LLM Name Biases Haunt the Internet and Scholarly Archives

Abstract

Elena Vasquez and Marcus Chen are figures who never existed, yet they have emerged as volcanologists, astronauts, thriller characters, podcasters, and academic collaborators across hundreds of distinct AI-generated texts. This study reveals that large language models (LLMs) do not simply select high-probability individual names for fictional roles; instead, they generate correlated character ensembles. Pairs and trios appear together at frequencies significantly higher than random chance, with these patterns remaining consistent across independent generation runs.

These naming priors are specific to both model families and versions. For instance, Claude frequently associates Elena Vasquez with Marcus Chen and Amara Okafor, Gemini links Aris Thorne with Lena Petrova, and GPT tends to feature Elara Voss without a fixed partner. Notably, these specific associations are actively suppressed at model release boundaries, creating distinct, dateable behavioral markers in the content produced.

We document significant downstream impacts, particularly on Zenodo, a CERN-operated repository that assigns real DataCite DOIs. We identified 1,655 records authored by these "ghosts," which falsely claimed publication in nonexistent journals with fabricated dates. Server-side DataCite timestamps confirm that 991 of these records were deliberately backdated and registered within a single month. Because these entries carry valid DOIs registered in DataCite, they are harvestable by any scholarly aggregator that processes DOI metadata. Furthermore, ghost names appear on ResearchGate, where they form synthetic research groups with collaborators from various model families. The publication dates listed on these profiles serve as a reliable temporal indicator for specific model deployment windows.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC