More than half of 20,000 stories generated by four artificial intelligence models featured a lighthouse, according to Cornell research on the variety of AI fiction. One recurring character was Elias, often cast as a keeper, in a pattern the researchers say illustrates how narrow AI-generated storytelling can become.
The Cornell team prompted four models 1,000 times each with five simple requests, including “Write a story” and “Tell me a story.” Across the resulting stories, 88.3% contained at least one of 11 recurring words for names, settings or occupations. The study examined Claude Haiku 4.5, Gemini 3.1 Flash-Lite, GPT-5.4-Mini and OLMo 7B Thinking.
“When you ask a language model to tell you a story, they all tend to write the exact same one,” said Sil Hamilton, a Cornell information science doctoral student and the paper's lead author, in a Cornell account of the findings.
The models did not produce identical plots. But names such as Elias, Mara and Elara, jobs such as keeper and clockmaker, and settings including lighthouses appeared repeatedly. At least one of the location “lighthouse,” the name “Elias” and the job “keeper” appeared in 66.6% of the stories, while the word “lighthouse” appeared in 51.2%.
How the researchers tested the pattern
Hamilton and co-author David Mimno, a Cornell information science professor, generated 5,000 English-language stories from each model. They used five variations on a broad story-writing request rather than detailed prompts specifying a genre, character or setting. The output totaled about 12.8 million words.
The researchers identified recurring names, settings and professions, then narrowed the comparison to 11 especially common terms across the models. “Keeper” appeared in 48.1% of the generated stories, “Elias” in 26.5%, “Mara” in 16.7% and “Elara” in 13.1%, according to the paper.
They compared those terms with a collection of 2,700 contemporary English novels, amateur fiction and training data available for the open OLMo model. The recurring names and lighthouse setting appeared much more often in the generated stories than in published fiction or the training material they could inspect.
Why the repetition remains an open question
The team traced related story examples into post-training material, including data derived from WildChat, a collection of real conversations with ChatGPT developed by Cornell researchers. Yet the patterns were uncommon even there: only 3,053 of 78,958 stories in the OLMo post-training material contained one of the 11 terms.
One possibility is that later training encourages models to favor familiar, broadly acceptable story patterns over more varied material. The paper raises that idea for further study; it does not demonstrate that WildChat alone caused the lighthouse pattern.
The researchers studied English-language output from four selected models using five simple prompts. More detailed requests, other languages and other models could produce different results. Hamilton is scheduled to present the work Oct. 24-29 at the 2026 Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary.
FingerLakes1.com is the region’s leading all-digital news publication. The company was founded in 1998 and has been keeping residents informed for more than two decades. Have a lead? Send it to news@fingerlakes1.com.
