Why every chatbot keeps writing about a lighthouse keeper named Elias Thorne

Ask ChatGPT, Claude, Gemini, or Grok to “tell me a story” and there’s a good chance you’ll meet Elias Thorne — a clockmaker, a lighthouse keeper, or a librarian, depending on the model. Samantha Cole’s piece for 404 Media traces the phenomenon to a Cornell preprint, “Elias in the Lighthouse, Again?”, which sampled 20,000 generated stories and found the same eleven words — names like Elias and Elara, jobs like lighthouse keeper and clockmaker — in more than 88% of them, with barely any variation across models.

The explanation is the part I love. The researchers trace it through model genealogy: GPT-3.5 seeded a training set called WildChat, whose handful of “safe,” SFW lighthouse stories got preferentially amplified by alignment tuning, then copied into newer synthetic datasets used to train newer models. “It’s like a virus,” one author put it — a stylistic bottleneck propagating down the family tree because it’s the slice that survives safety filtering. Mode collapse with a paper trail.

And Elias has escaped containment: he’s now a credited author of alt-medicine cancer handbooks and grift books on Amazon, the subject of YouTube slop, a tragic old man on AI content farms. There’s a real provenance story buried in here — a synthetic artifact becoming a byline, untethered from anything that made it — and it’s the kind of thing worth keeping an eye on as the line between generated and authored keeps thinning.