Why AI storytelling patterns recur across models: an emergent literary prior and what it means for builders

A new study reveals that large language models (LLMs) from different makers converge on the same set of narrative words when writing short fiction. This emergent literary prior in AI storytelling has practical implications for anyone building or using AI for creative writing tasks.

What happened

Researchers Sil Hamilton and David Mimno published a study titled "Elias In The Lighthouse, Again? Diagnosing Low Diversity In LLM Stories" on arXiv. They generated 20,000 short stories across four different LLMs, using five short prompts per model, each run 1,000 times. The total corpus was about 12.8 million words, with an average story length of roughly 640 words.

Across all models and prompts, eleven specific nouns appeared repeatedly and prominently. The words include "lighthouse" and "Elias", along with other archetypal fiction nouns. The researchers call this an emergent literary prior: a set of statistical anchors that the models gravitate toward when generating narrative text.

Why AI builders should care

This convergence matters because it challenges the assumption that LLMs compose stories from a neutral, random starting point. The same eleven words appeared across four independently built models, suggesting a shared bias in the foundational layers of modern AI: similar training data, similar algorithms, and similar tuning practices.

For AI builders, this means that any application relying on LLMs for creative writing, content generation, or narrative tasks inherits this bias. If you are building a story generator, a game narrative engine, or a marketing content tool, your outputs will tend to cluster around these narrative attractors unless you actively counteract them.

This also affects AI storytelling reliability and prompts. If you evaluate a model's creative range using short, open-ended prompts, you may overestimate its diversity. The model is not being creative in a broad sense; it is sampling from a narrow conditional distribution shaped by its training.

Practical implications

Prompt design matters more than you think. The study used very short prompts that gave the AI no directional guidance. When you use short prompts in production, the AI fills in defaults, and those defaults are the emergent literary prior. To get more diverse outputs, you need to write detailed prompts that specify setting, character, tone, and constraints.

Cross-model checks can reveal hidden bias. If you are using multiple LLMs for content generation, run the same prompt across models and compare the outputs. If they converge on the same words or narrative structures, you have evidence of a shared bias that your application needs to handle.

Consider stochastic methods for creative tasks. The article suggests using random number generators or seed-of-thought prompting techniques to break out of the default narrative attractors. This is especially relevant for game narrative, interactive fiction, and any application where variety is a feature.

Evaluation frameworks should account for convergence. If you are benchmarking LLMs on creative writing tasks, your evaluation should measure not just quality but also diversity and originality. A model that scores high on coherence but always writes about lighthouses and librarians may not be suitable for your use case.

Caveats

The findings are reported through a Forbes article summarizing the arXiv study. The exact list of eleven words, the specific model names, and the full experimental methodology are not reproduced here. For precise details, consult the original study by Hamilton and Mimno.

The study used short prompts and short stories (about 640 words each). It is unclear whether the same convergence would appear with longer stories or more detailed prompts. The article speculates that longer stories might still show the pattern, but this has not been tested.

The words appeared prominently in the stories, but the study does not fully distinguish between words used as central plot elements and words used as throwaway references. The practical impact on narrative quality may vary.

Finally, the study covers fictional short stories only. The same convergence is unlikely to appear in factual or non-fictional writing, where real-world constraints guide word choice.