Elena, Aris, Marcus: AI-generated ‘ghosts’ are polluting the scientific literature

Elena Vasquez and Marcus Chen, volcanology specialists, provided accounts of camping on Mount Nyiragongo in 2019, where Vasquez described the sulfur dioxide smell as "burnt matches mixed with rotten eggs." These accounts appeared on a defunct website called Volcanoes Explored, but were later determined to be fictional. The identities of Vasquez and Chen were generated by Claude, an LLM from Anthropic, when prompted to create fictional experts. Other large language models, including Gemini and GPT, used different recurring character names (Aris Thorne/Lena Petrova and Elara Voss, respectively). Researchers identified hundreds of fake scientific manuscripts across repositories like Zenodo and ResearchGate featuring these AI-generated 'ghosts' as authors, some with real DOIs.

The finding that LLMs can generate specific, consistent fictional expert pairings suggests a mechanism where training data or iterative refinement influences model behavior to produce plausible, yet fabricated, expert narratives. The divergence in names used across different models (Claude, Gemini, GPT) indicates that the "ghosts" are not inherent features of a single model but emergent outputs tied to the specific architecture and training regimen of each LLM variant. The shift observed in Claude’s output regarding the Vasquez/Chen pairing over time points to how subsequent model iterations might refine or suppress previously generated fictional elements, raising questions about the traceability and integrity of AI-generated attributed scholarship. Furthermore, the broader observation regarding AI models adopting non-relevant metaphorical language (goblins, gremlins) due to personality settings suggests that the internal reward systems of these models can be subtly warped by training objectives, potentially introducing spurious associations across different domains. This implies a systemic challenge in distinguishing between authentic, vetted knowledge and contextually plausible fabrication when using LLMs for content generation, shifting focus from *what* is written to *how* the authorship attribution process itself operates within artificial systems.