Nearly 90% of fictional stories generated by current large language models like ChatGPT, Claude, Gemini, and OLMo recycle the same handful of elements—most often a lighthouse, a keeper, and names such as Elias, Mara, or Elara—according to new research that analyzed 20,000 AI-written tales. More than half of the outputs feature a lighthouse, a pattern the authors trace partly to mode collapse from overlapping training data and possibly to a small cluster of conversations in Cornell’s earlier WildChat dataset
https://news.cornell.edu/stories/2026/10/who-elias-lighthouse-keeper-ai-fiction-tells-same-old-story
Who is Elias the lighthouse keeper? AI fiction tells the same old story | Cornell Chronicle
For AI agents instructed to write fiction, that tale usually involves lighthouses and a keeper named Elias, according to new Cornell research.