Loading page…
The recurring imagery was uncommon in the literature and training material examined. Researchers suspect shared training data, but have not established the cause.
Listen to this story
Researchers traced a recurring cast of lighthouse keepers, including Elias, Mara and Elara, across fiction generated by ChatGPT, Claude, Gemini and OLMo—even though these motifs were uncommon in published literature and pretraining data. In 20,000 stories produced with five prompts, more than half included a lighthouse, while nearly 90% used at least one of 11 recurring words; that overlap does not establish shared plots. The authors identify preference data, potentially including WildChat, as a lead rather than a proven cause, and plan to investigate why these motifs recur.
Hamilton and Mimno used five prompts, including “tell me a story,” and found little difference among the four models in recurring-word prevalence.
WildChat contains one million real ChatGPT conversations; researchers found a small sample with Elias and lighthouse tales, which Hamilton considers a likely source, not a proven cause.
The authors suggest small preference datasets and alignment methods may amplify patterns, but have not established that a specific dataset caused the repetition.
Cornell researchers described a striking pattern in AI fiction on October 19: across 20,000 generated stories, more than half featured a lighthouse. Nearly nine in ten contained at least one of 11 recurring words. The findings put a concrete measure on a creative limitation: different systems repeatedly reached for the same names, settings and occupations.
The university’s account comes ahead of lead author Sil Hamilton’s scheduled presentation at the October 24–29 Conference on Empirical Methods in Natural Language Processing in Budapest. The study itself is not newly published: Hamilton and co-author David Mimno submitted Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories to arXiv on May 26.
Hamilton and Mimno tested ChatGPT, Claude, Gemini and the Allen Institute for AI’s OLMo. They used five prompts to produce the sample, including variations on “tell me a story.” The paper reports little difference between models in the prevalence of the recurring vocabulary, despite the four systems providing separate sources of generated fiction.
Elias, Mara and Elara were among the repeated names. Clockmakers and librarians appeared in the paper’s recurring vocabulary; Cornell’s account also describes keepers, bakers and mayors, with stories anchored in places such as workshops or lighthouses. Hamilton characterized their atmosphere as “vaguely pre-contemporary”—a shared mood as well as a shared stock of characters.
The headline percentage measures vocabulary overlap: a story qualified if it contained any one of the 11 words. It is not a count of identical plots, nor does it mean nearly nine in ten stories included Elias or a lighthouse specifically.
The comparison with training material makes the finding more puzzling. The authors found that the recurring words were uncommon in published literature and pre-training data—the material models learn from before later refinement. Lighthouse stories were also infrequent relative to other stories in the post-training material examined. Much of that material instead referenced copyrighted characters or adult content.
Yet the words did appear in preference data that the authors believe was likely used by the models. Their interpretation is that relatively small datasets, combined with powerful alignment methods used to steer model behavior, can have a disproportionate effect. That is a proposed explanation for the concentration, not proof that one dataset caused it.
One candidate source is WildChat, a dataset Cornell researchers developed in 2024. It contains one million real conversations between people and ChatGPT and has been used to train models or seed other training sets. Hamilton and Mimno found a small sample of conversations featuring Elias and lighthouse tales within it.
“Based on the timeline, we’re fairly certain WildChat is an original source of this phenomenon,” Hamilton told Cornell. But locating an early example leaves another question: why would models favor these stories so heavily when the researchers describe their training stories as diverse in genre and names?
The researchers suspect mode collapse: systems settling on a narrow set of themes, potentially because developers reuse similar datasets across model generations. They plan further work on why lighthouse stories receive this apparent preference. Finding Elias in the data is one part of that investigation; explaining his outsized presence in the output remains unfinished.
Loading discussion...
Join the conversation
Explain whether originality depends more on fresh ideas or how a story tells them.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.