Why the same fake experts keep appearing in AI-generated papers

Large language models don't generate names randomly; they produce consistent pairs and trios of fictional characters whose co-occurrence rates far exceed chance, with patterns that differ between Claude, Gemini, and GPT. These "ghost couples" have infiltrated academic publishing: researchers found 1,655 papers with fake authors and nonexistent journals registered on Zenodo with legitimate DataCite DOIs, timestamped evidence of backdating, and traces across ResearchGate. The fingerprints are model-family and version-specific, effectively dateable.