LLMs repeatedly hallucinate identical fictional personas across major providers
September 19, 2026
A Cornell University study of 20,000 AI-generated stories found that specific names and occupations, such as lighthouse keeper and clockmaker, appear in 88 percent of outputs. The findings show that models from OpenAI, Anthropic, and Google consistently invent the same non-existent personas like Elias Thorne.
HOW THIS AFFECTS YOU
●
researcherYou should investigate systemic biases in training data that lead to these repetitive semantic collapses.