The Ghost in the Machine: How "Elias Thorne" Haunts AI Storytelling
A single fictional name—"Elias Thorne"—has appeared in over 88% of AI-generated stories across the world's leading chatbots, revealing a strange and unintended side effect of how these models are trained.
The Discovery
Software engineer Daniel May first noticed something peculiar in early 2025. Across multiple chatbot-generated stories, the same name kept appearing: "Elias Thorne." What began as a curiosity soon spiraled into a widespread phenomenon.
By early 2026, Google Trends data showed a sharp spike in searches for "Elias Thorne," alongside related queries like "lighthouse keeper."
The Research
A team of researchers—Sil Hamilton, David Mimno, and Rebecca M. M. Hicke at Cornell University—decided to investigate. In late May, they published a preprint paper on arXiv titled "Elias in the Lighthouse, Again?"
Their methodology was systematic:
The study sampled 20,000 stories generated by ChatGPT, Claude, Gemini, and the Allen Institute for AI's chatbot using five different prompts.
The results were startling. An identical set of 11 words—names like Elias, Mara, and Elara; occupations like lighthouse keeper, clockmaker, and librarian—appeared in over 88% of all generated stories. There was little difference between models.
The Root Cause
The researchers hypothesize that the answer lies in safety and alignment tuning. Hamilton noted that modern model development resembles a family tree, with OpenAI's GPT-3.5 as a common ancestor.
The trail leads to a training dataset called WildChat, derived from 1 million real ChatGPT conversations. This dataset contained 166 instances of the name "Elias" used in a "lighthouse" style. Models trained on WildChat replicated this style, and as subsequent models were trained on those outputs, the pattern propagated further.
"During alignment, models prefer a small subset of safe stories from WildChat, focusing on these bland, repeated themes to avoid generating unsafe content," the researchers suggest.
Beyond the Chatbot
The name "Elias Thorne" has escaped the confines of AI training data:
On Amazon, an author named "Elias Thorne" now publishes across wildly different genres:
- Alternative medicine cancer handbooks
- YouTube algorithm guides
- Books on Greek mythology
- Psychological thrillers
In fiction, Elias Thorne appears as a protagonist in fantasy books and as a musical artist creating ambient albums.
On YouTube and AI-generated slop sites, he has become a recurring tragic figure—a "83-year-old Sergeant Major Elias Thorne" or a wealthy man who died with just twelve dollars.
Even the BBC got involved, publishing a short story with a character named Elias in its 2024/2025 children's writing competition—though this may have been human-written.
A Curious History
The name "Elias" appears to have a pre-generative AI legacy.
In the 1980s, the trading card series "Dinosaurs Attack!" featured Elias as a time-traveling mad scientist. And in real life, Elias Allen was a 16th-century clockmaker in London—a profession that echoes the "clockmaker" occupation now so common in AI stories.
What Comes Next
The researchers plan to further explore their hypothesis that alignment training inadvertently creates these recurring, safe themes. For now, the ghost of Elias Thorne continues to haunt the world's most advanced AI systems—a digital phantom born from the very safeguards designed to keep these models in check.