Cornell study finds AI fiction keeps returning to lighthouses and Elias

Cornell researchers have found that current AI systems asked to write fiction repeatedly gravitate toward the same elements: lighthouses, keepers and characters called Elias. In a 20,000-story experiment spanning ChatGPT, Claude, Gemini and OLMo, 88.3% of outputs used at least one of 11 recurring words associated with names, settings or occupations. A lighthouse appeared in 51.2% of the stories, while “keeper” appeared in 48.1% and “Elias” in 26.5%. At least one of those three terms occurred in 66.6% of the sample. Other recurrent names included Mara, in 16.7% of stories, and Elara, in 13.1%. The researchers used five broad prompts to generate English-language stories, producing 5,000 stories from each model, or about 12.8 million words in total. The prompts did not specify a genre, setting or character. The generated fiction commonly combined a recurring name with a role such as keeper, baker or mayor and a setting such as a workshop or lighthouse. Sil Hamilton, one of the researchers, said the result challenges the expectation that a model will deliver a distinctive story on each request. “You think you’re getting a unique story, but at scale it’s the same thing over and over again,” he said. The team compared the terms with a collection of 2,700 contemporary English novels, amateur-fiction works and training material available for OLMo. The names and lighthouse setting occurred far less often in those comparison materials than in the generated stories. Separately, only 3,053 of 78,958 stories in OLMo post-training material contained one of the 11 terms. That gap means the researchers have not identified a definitive origin for the repetition. They suspect a form of mode collapse, in which models default to similar themes as developers train systems on similar data and incorporate them into newer models. They also said language models have been built mainly for tasks such as coding rather than engaging fiction, leaving creative output prone to repetition. The study traced related story examples through post-training material, including material derived from WildChat, a Cornell dataset developed in 2024 that contains 1 million real conversations between people and ChatGPT. Hamilton and David Mimno found a small sample of WildChat conversations featuring Elias and familiar lighthouse stories. Hamilton said the timeline made WildChat a plausible original source, but the research did not show that the dataset alone caused the pattern. The recurring-language effect was broadly consistent across the four sampled models, according to earlier reporting by 404 Media. Still, the study was limited to four selected models, English-language output and five simple prompts. Different models, languages or prompt designs could yield different results.