Researchers found that large language models repeatedly invent the same character names when asked to fabricate bios or bylines, and those names tend to travel in correlated groups across the web. The pattern offers a tentative new tool for flagging AI generated content.
When a volcano expert named Elena Vasquez appears on three unrelated websites, each bio plausible, none of her real, something has gone wrong. The names were not copied from a shared directory. They were conjured independently by the same large language model, which keeps returning to the same handful of invented names across completely separate contexts.
That pattern — names that travel in correlated groups rather than as singletons — sits at the core of arXiv preprint 2606.02184, a paper by Michał Brzozowski and Neo Christopher Chung of Samsung AI Center and the University of Warsaw, posted to arXiv on June 1, 2026. The central claim: large language models have model-specific, version-specific priors over character names, and those priors are strong enough to act as a weak forensic signal for AI-generated text.
The signal works in ensembles, not in isolation. A single hallucinated byline is unremarkable; any model can produce a "Dr. Sarah Mitchell" if asked. What the paper documents is that models tend to produce the same handful of names when they invent, and those names show up together. Elena Vasquez and Marcus Chen recur as a pair across contexts as different as volcano expert bios, podcast host rosters, thriller protagonists, and the bylines of 1,655 fictional academic papers on Zenodo, a CERN-operated repository. Add a third correlated name — Amara Okafor — and the signal sharpens further. The authors documented a collage of three unrelated sites whose invented experts share the same trio of first and last names and the same AI-stock-photo face treatments. The full ghost trio peaks at 20% co-occurrence in one Claude checkpoint and is fully suppressed in later versions.
The technical mechanism is a model-diffing method called CDD (Contrastive Decoding Diffing), originally developed in related work. The name-priors result emerged as a side finding, then grew into its own paper. The broader pattern is well known in machine learning research: training data skew and tokenizer statistics create predictable defaults, and a model will fall back to those defaults whenever a prompt does not push it elsewhere. Names are unusually legible defaults because they are high-entropy tokens that get reused.
That is what makes the finding useful in practice. Journalists and fact-checkers already triage suspicious bylines by hand. Platforms trying to flag AI-inflated author lists on academic repositories or guest-post networks have had to rely on writing-style heuristics, which are easy to defeat with a light paraphrase. Name priors are a different kind of tell, harder to clean out without changing the entire content, and the paper offers a way to measure them.
The limits are real. The paper explicitly states that finding Elena Vasquez and Marcus Chen together on a website is a strong tell for Claude-generated content — but that tell is version-specific: the pair co-occurred in 23% of responses in one checkpoint, decaying to 0% in the latest Sonnet release, with active suppression visible at release boundaries. Any detection system built on name ensembles will need to be re-baselined as models are updated. The 1,655 ghost-authored records on Zenodo — 991 of them registered in a single month, March 2026 — represent the downstream harm the signal helps reveal: real researchers whose names have been borrowed, and editorial workflows that have to sort the invented work from the genuine.
What to watch next: independent replication on closed-weight commercial models, a published version of the CDD method, and any platform integration that uses name-ensemble checks alongside other signals. The practical form this would take is a tool that scores a byline for prior likelihood and surfaces that score to a human reviewer, who then decides whether the byline is worth a closer look.