Ask several leading AI models to tell a story and the lighthouse may turn up before anything especially original does. Cornell researchers found that nearly 90% of the stories they generated contained at least one of 11 recurring words, with more than half featuring a lighthouse.
Watch Desk analysis
What happened
The researchers prompted ChatGPT, Claude, Gemini and the Allen Institute for AI’s OLMo with variations on “tell me a story”. Across 20,000 generated stories, they found recurring names including Elias, Mara and Elara, along with jobs such as keeper, baker or mayor and settings such as workshops and lighthouses. Cornell published its account on 19 October, describing results from research due to be presented at a conference later this month.
The team found familiar themes were not common in existing literature or the training data they examined. It suspects that a small number of lighthouse-themed conversations in Cornell’s WildChat dataset may be part of the explanation, because that dataset was used to train or seed later datasets. The researchers say they cannot yet explain definitively why the pattern recurs.
Why it matters
This is a concrete example of a distinctive output becoming common across different models. It gives readers a way to think about AI-generated writing beyond whether a passage sounds fluent: ask whether it is actually varied, and where its apparent originality comes from.
The findings concern stories produced under these prompts, not every kind of writing or every output from the four models. Still, a repeated lighthouse is a memorable clue that models can sound inventive while drawing on surprisingly narrow patterns.
Our read
A lighthouse keeper named Elias is an entertaining motif; finding him on the literary night shift across four models is more interesting. The study makes a useful case for testing creative systems on variety, not just polish. The researchers’ proposed link to WildChat is a hypothesis to investigate, not a settled explanation.
What to watch
- The paper’s presentation at the Empirical Methods in Natural Language Processing conference, scheduled for 24–29 October.
- Whether follow-up work confirms how WildChat or other shared datasets contribute to the pattern.
- Whether similar repetition appears in other genres, prompts and languages.
Discussion spark: Should AI writing tools be judged on how varied their output is, even when users ask only for a single polished answer?
Sources and evidence
- Source update (19 October 2026, 14:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.