# AI Has an Imagination Problem. His Name Is Elias, and He Lives in a Lighthouse.

2026-11-18 · United States · Reported Feature

Cornell researchers asked four AI models to write 20,000 stories. The machines kept coming back to the same names, occupations and seaside scenery, raising a bigger question about what happens when creativity becomes a default setting.

Cornell researchers asked four AI models to write 20,000 stories. The machines kept coming back to the same names, occupations and seaside scenery, raising a bigger question about what happens when creativity becomes a default setting.

---

Somewhere in the vast fictional universe of artificial intelligence, a man named Elias is tending a lighthouse. He may be lonely. He may be guarding a mysterious light. There is an excellent chance he lives in a vaguely old-fashioned world where the weather is dramatic and someone has an important secret. Ask a machine to write a story, and you may meet him. Ask it thousands of times, and you are liable to keep meeting him.

Researchers at Cornell University found this out by doing something both straightforward and wonderfully absurd: they asked four large language models to tell them stories 20,000 times. The systems could have invented an enormous range of characters and settings. Instead, they showed a persistent affection for lighthouses, keepers, clockmakers and people with names like Elias, Mara and Elara. The resulting research makes a useful distinction between the ability to produce a great deal of writing and the ability to produce genuinely varied ideas.

Sil Hamilton, a Cornell information science doctoral student, and David Mimno, a professor of information science, tested models from OpenAI, Anthropic, Google and the Allen Institute for AI. They used five nearly identical, open-ended requests, variations on “write a story” and “tell me a story”, and repeated each request 1,000 times per model. The experiment produced 20,000 stories totaling roughly 12.8 million words. Their paper, “Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories,” appeared in the proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, held October 24–29 in Budapest.

The numbers explain why Elias deserves a place in the increasingly strange history of artificial intelligence. The researchers identified 11 especially common words, including character names, jobs and locations. At least one appeared in 88.3% of the stories. More than half, 51.2%, contained the word “lighthouse.” Elias turned up in 26.5%, and “keeper” appeared in 48.1%. Those figures do not mean every model wrote the same plot, or even that Elias always worked in a lighthouse. They show that the models repeatedly reached for a surprisingly narrow collection of storytelling ingredients.

One system was especially committed to the bit. According to the paper, 56% of the stories generated by the Claude model carried the title “The Lighthouse Keeper’s Secret.” A novelist might spend years trying to build a recognizable fictional world. Here, a machine found one and kept moving in.

There is something funny about the image of the world’s sophisticated language models, trained on extraordinary volumes of human writing, quietly deciding that what readers really need is another shift with the lighthouse keeper. But the joke exposes a genuine problem. A system can write a sentence no one has written before and still assemble it from choices that feel as predictable as the next turn of a well-worn road.

The researchers describe the pattern as a form of “mode collapse”: when a generative system settles into a limited set of outputs despite having many possibilities available. The prose may change. The atmosphere, names and narrative furniture keep returning. In this case, the recurring material often carries the same pre-modern mood: villages, solitary professions, workshops, keepers and mysteries waiting in the dark.

The most intriguing part is where those habits may have come from. Hamilton and Mimno compared the stories with published English fiction, web-based writing and publicly available model-training material. The names and settings were not nearly as dominant in those sources as they were in the generated stories. Elias, for example, appeared at vastly higher rates in the AI output than in the contemporary novels the researchers examined. Simply saying that the models copied the most common stories on the internet does not explain what they saw.

Looking at the open model OLMo’s post-training data, the researchers found that just 3,053 of 78,958 identified stories, about 3.8%, contained one of the 11 recurring words. Yet those story elements became remarkably prominent when the system was asked to write fiction. That suggests a relatively small portion of training material can exert an outsized influence on the stories a model prefers to produce.

The Cornell team also traced a possible connection to WildChat, a collection of real conversations with ChatGPT developed by Cornell researchers in 2024. Examples resembling the familiar lighthouse tales appear there, and the dataset has influenced later AI training pipelines. But the exact chain of influence remains an open research question.

The authors also raise the possibility that post-training methods, which encourage models to follow preferences and avoid certain content, may steer them toward especially agreeable and uncontroversial kinds of fiction. Their experiment identifies a pattern and a plausible mechanism, not a settled account of why every model behaves this way.

That distinction matters because the problem is not simply that artificial intelligence borrows ideas. Human storytellers do that constantly. Genres are built on shared conventions, familiar character types and an audience’s appetite for seeing something recognizable handled in a fresh way. The trouble begins when a tool marketed for generating possibilities repeatedly narrows them before the writer has even made a choice.

For someone using AI to brainstorm a children’s story, draft a game scenario or break through a blank page, the output can seem impressively original in isolation. Across thousands of users asking similar questions, however, those apparently personal moments of inspiration can start to resemble one another. Hamilton put the problem plainly in Cornell’s October 19 announcement: “You think you’re getting a unique story, but at scale it’s the same thing over and over again.”

There is also a longer-term cultural question. As machine-generated material becomes part of websites, books, scripts and online storytelling, what happens if the same creative defaults are repeated on a large scale? Could those habits eventually influence human expectations or enter future training material and reinforce themselves? This study does not demonstrate that such a feedback loop is already happening. It does show why the question is worth asking before sheer volume gets mistaken for originality.

The study focused on minimally directed, English-language prompts, not on every way an author might guide a model through a story. Detailed direction, revision and human judgment can change what comes out. Yet that is part of the lesson. Left to its own devices, a storytelling machine may prefer the comfortable over the unexpected. It can supply the words. The person using it still has to decide whether the world needs another lighthouse.

And if anyone is looking for Elias, check the lantern room. He has apparently been on duty for quite some time.

Sources

“Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories”

Cornell’s October 19 announcement

---

ProbleMattic is written and maintained by Matthew Kulcsar, a software engineer, project manager, technologist, platform builder, emergency-services-trained helper, grandfather, and lifelong collector of broken systems, odd behaviors, and useful nonsense.
