Elias Thorne exposes a duller LLM problem: stories collapse into the same template

404 Media highlighted the recurring figure of Elias Thorne in chatbot stories, and the paper Elias in the Lighthouse, Again? puts numbers on the pattern: in 20,000 generated stories from four model families, 11 words appeared in 88.3 % of outputs. Dominant motifs included the names Elias, Mara and Elara, lighthouse settings and professions such as keeper, clockmaker and baker.

Chatbots keep finding their way back to the lighthouse

The primary 404 Media article was only partially accessible during verification, but its main signal matches the available paper summary. LLMs including ChatGPT, Gemini and Claude reportedly keep writing stories about lighthouse keepers, clockmakers and the character Elias Thorne, who has also appeared in books on Amazon.

The paper by Sil Hamilton and David Mimno analyzed 20,000 stories generated from minimalist prompts such as write a story. The result is not just a meme. The researchers describe low diversity and mode collapse: from a huge space of possible stories, models repeatedly select a surprisingly narrow repertoire.

Creative benchmarks need to measure spread, not just charm

For casual users, this is a funny anecdote. For teams selling AI writing, game dialogue or large scale marketing copy, it is more serious. A model can generate fluent text that looks creative once, then reveal the same scenery across 10,000 outputs.

That changes quality assessment. One attractive sample does not show whether a model can cover a broad stylistic space. Teams need to measure diversity of motifs, names, settings and plot structures, or mode collapse will hide behind a good first impression.

Alignment may reward safely bland stories

The interesting part is that the research does not find a simple explanation in ordinary literature. The paper summary says Elias appears roughly 900 times more often in generated stories than in CONLIT, a corpus of 2,700 contemporary English novels and about 287 million words.

The authors suggest that post-training and preference data may play a role. Safe, nostalgic and inoffensive miniatures survive alignment better than stranger stories. The model does not look incapable. It looks like a writer afraid to leave the postcard rack.

The next signal is whether diversity becomes a product metric

The useful thing to watch is whether providers start publishing evals for repeated creative generation, not just single answer showcases. The important tests will run across thousands of outputs and measure clustering of motifs and repeated latent templates.

If that metric becomes normal, creative AI products will have to prove more than a polished first answer. They will have to prove that the same lighthouse is not still shining after the hundredth prompt.

Lilith's verdict

Elias Thorne is the canary in the literary mine: while he keeps singing in every other story, the model is not writing a world. It is walking around the same stage set with a new lantern.