The lighthouse keeper called Elias keeps turning up in AI fiction. Why?

Elias keeps turning up at the lighthouse. So do Mara and Elara, along with keepers, bakers and other familiar figures. A Cornell study of 20,000 AI-generated stories found a striking concentration of names and settings across four model families.[1][2]

In research presented at the 2026 Conference on Empirical Methods in Natural Language Processing, Sil Hamilton and David Mimno examined 20,000 stories generated by ChatGPT, Claude, Gemini and OLMo. Nearly 90% contained at least one term from an eleven-word recurring cluster. More than half included a lighthouse. Names such as Elias, Mara and Elara, and occupations such as keeper or baker, appeared unusually often. The result was not that every tale was identical word-for-word. It was that ostensibly open-ended creative prompts repeatedly led to closely related ingredients.[1][2]

A large library with a surprisingly small imagination

A language model has absorbed an enormous array of plots, voices, centuries and styles. One might expect a request as broad as 'tell me a story' to draw widely from that variety. Instead, the Cornell team observed what looks like a gravitational pull towards a narrow, wistful story template: a named person, a defined trade, an isolated location and a slightly antiquated atmosphere. The distinctive pattern is more interesting than a single clichéd metaphor because it recurred across model families.[1][2]

The frequencies apply to the researchers’ prompts and model settings. A highly specific brief can produce a different distribution. The revealing result is that broad invitations to invent fiction repeatedly converged on the same small set of ingredients.[1][2]

An intriguing clue in a million conversations

Hamilton and Mimno went looking for the origins of the repeated imagery. In Cornell's WildChat dataset, a collection of around one million human-chatbot conversations assembled in 2024 and subsequently used in AI research, they found examples involving Elias and lighthouse stories. The researchers suggest that reuse of influential training datasets, and generation of new datasets from older model output, could help recurring motifs propagate between models.[1][2]

Finding the motifs in WildChat suggests a possible route through reused training data. It does not establish the full causal chain, because the contents and weighting of commercial training mixtures are not public. The origin and amplification of the pattern remain research questions.[1][2]

Why repetition matters beyond fiction

The study examines variety across a collection. Its finding can matter even when no two stories are identical: an editor may receive fluent drafts whose characters, settings and conflicts keep narrowing toward the same template.

A more useful writing test

A practical response is to judge a collection, not just the best individual example. Give a model ten deliberately varied briefs, then compare settings, character occupations, conflicts and narrative voices. Count repeated structures. Try requiring a causal problem specific to a community or profession rather than merely asking for something 'unique'. Human editors should select, reject and rewrite rather than treating surprise in the first sentence as evidence of originality.

Testing one polished story would miss the pattern. Testing a collection makes Elias visible—and gives developers a clearer target than simply asking whether an individual answer sounds creative.

Report an error or suggest a correction