Meet Elias Thorne, the Lighthouse Keeper That Lives Inside Every Chatbot
A Cornell University research paper analyzed roughly 20,000 stories generated by multiple AI models and found that just 11 words appeared in 88% of all outputs, with "Elias the lighthouse keeper" showing up in two-thirds of all stories. Researchers ruled out pre-training data as the cause and instead pointed to shared alignment training datasets, such as WildChat, suggesting that safety training designed to steer models away from copyrighted characters may have accidentally elevated a narrow set of "safe" alternatives to near-universal prominence. The finding has practical consequences beyond quirky repetition: the name Elias Thorne has already appeared as a listed author on books promoting alternative cancer treatments. The study reinforces a broader pattern of AI creative limitations, with a separate 2025 study finding that image generation models default to just 12 recurring visual motifs regardless of prompt.