✍️🤖 “The Lighthouse Keeper’s Secret”
From the NYT Books Newsletter for November 3, 2026:
According to a recent study, when researchers at Cornell used A.I. to generate 20,000 short stories, conspicuous motifs emerged. They tended to feature uncommon names like Mara, Elara and Elias. Nearly half of the stories contained a character who worked as a “keeper,” and more than half contained a lighthouse. Fifty-six percent of the stories generated by Claude bore the same title: “The Lighthouse Keeper’s Secret.”
Chances are, a story titled “The Lighthouse Keeper’s Secret” isn’t one that you’re dying to read. (Or maybe you are; I do not pretend to know the workings of your heart.) Lighthouses and keepers — and Eliases, for that matter — don’t appear all that frequently in existing literature. The researchers surmise that such overrepresentation is the result of large language models trained on the same data sets. But it’s not as if lighthouses and keepers were such a dominant presence in those data sets either. Precisely why these cutting-edge robots “prefer” such quaint motifs is still a mystery. As one of the study’s authors put it, “The mood is always vaguely pre-contemporary.”
The study was submitted on May 26, 2026 by Sil Hamilton and David Mimno: Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories — HTML, PDF.
Figure 1: We prompt four models to write 20,000 stories — 88.3% of these stories contain at least one of 11 tokens at rates vastly higher than contemporary English literature (here measured in parts per million).