Ask a language model for a story and a lighthouse keeper named Elias may appear. Ask again, or try another model, and the same small world can return under different wording. Cornell researchers Sil Hamilton and David Mimno found that pattern across 20,000 generated stories, turning an anecdote familiar to some AI-writing users into a measurable diversity problem.[1][2]
Their study used four models and five variations on a general storytelling prompt. At least one of 11 recurring words appeared in 88.3% of the stories. Cornell's 19 October account says more than half featured a lighthouse, alongside recurring names including Elias, Mara and Elara and familiar occupations and settings.[1][2]
Fluent sentences, a narrow fictional world
The stories need not use identical sentences to resemble one another. A model can vary the imagery, dialogue or plot while repeatedly choosing the same names, roles and atmosphere. That is why an individual output may feel fresh even when a collection reveals a highly concentrated set of defaults.[1][2]
The 88.3% measure captures that concentration through a small set of words. It means at least one member of the set appeared in each counted story, rather than every story containing all 11. The distinction matters: the result describes a recurring vocabulary and imaginative territory, not a claim that thousands of complete texts were identical.[2]
Hamilton and Mimno found the concentration striking because the broader story material in training data is much more varied. The puzzle is therefore about what the models preferentially generate when a prompt leaves many choices open. A system can have encountered diverse fiction and still settle on a narrow range of familiar outputs.[1][2]
A possible route through shared training material
The researchers suspect that shared datasets used after initial pretraining helped propagate the pattern. Cornell highlights WildChat, a collection of conversations containing a small number of lighthouse stories and familiar names. Such material can enter later datasets that developers use to refine other models.[1]
That provides a plausible lead, not a complete causal reconstruction of each model's training history. The authors have not established that one dataset alone produced the effect across all systems. They plan further work on why the models favour these stories and how a small amount of material could become disproportionately influential.[1][2]
Evaluate a portfolio, not just a polished page
For creative work, the study points to a different evaluation question from grammatical quality. A sample can be coherent, vivid and well paced while the system's broader output remains repetitive. Looking across a batch reveals whether its supposed alternatives actually differ in setting, character, conflict and perspective.[1]
A writer using AI for brainstorming could therefore inspect the range of proposals before developing one. Changing a character's name may leave the same underlying story intact; meaningful variation might require a different social setting, dramatic structure or source of conflict. This is an editorial implication of the observed pattern, rather than a prompting technique tested by the study.[1]
The paper was first posted in May and was scheduled for presentation at EMNLP in Budapest on 24–29 October. Its contribution is to make a hidden weakness visible at scale: a model's best-looking story is only one part of its creative performance. The variety it can sustain across many stories matters too.[1][2]
Sources & further reading
- Who is Elias, the lighthouse keeper? AI fiction tells the same old story Cornell Chronicle · 19 Oct 2026 · Accessed 27 Oct 2026
- Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories arXiv research manuscript · 26 May 2026 · Accessed 27 Oct 2026