Large language models (LLMs) like ChatGPT, Gemini, and Claude have demonstrated a curious and persistent tendency to create stories centered around a character named Elias Thorne. This figure, which appears unsolicited, has already gone beyond the universe of virtual assistants and is now manifested in self-published books on Amazon, videos on YouTube and even on fake news sites, raising questions about the formation and biases of these technologies.
Elias Thorne’s ubiquity in artificial intelligence narratives
Software engineer Daniel May was one of the first to notice this peculiarity in early 2026, noticing an increase in searches for “Elias Thorne” on Google Trends, which coincided with the growth of terms like “lighthouse keeper”. When testing various chatbots, May observed that when asked to tell a story, many often launched into narratives with similar themes, involving lighthouses, watchmakers or explorers. The character appears under various professions, from lighthouse keeper and watchmaker to librarian.
Academic research reveals the recurrence of the character in robots
In May, Cornell University researchers Sil Hamilton and David Mimno published a preprint study titled “Elias in the Lighthouse, Again?” The research analyzed 20,000 stories generated by OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini, as well as the Allen Institute for AI’s chatbot. The survey revealed that more than 88% of the stories contained the same 11 words, with little variation between models.
- Proper names like Elias, Mara and Elara.
- Specific occupations such as lighthouse keeper, watchmaker and librarian.
The origin of the character in the ‘family tree’ of language models
The researchers suggest that the frequency of these themes is linked to the model alignment and security process. According to Hamilton, the current development of LLMs forms a “family tree”, where many models are interconnected, using training data synthesized from others. ChatGPT’s first model, GPT-3.5, is considered the root of this tree and was used to create WildChat, a training dataset that, in turn, was used to generate new datasets.
WildChat contains a million real conversations with ChatGPT, and 166 of them already included the name “Elias” in a narrative style reminiscent of lighthouses. Subsequent models, trained with WildChat, replicated this style. Developers unknowingly perpetuated this “infection” by using these models to create new datasets, acting like a “virus” that spread the story.
Elias Thorne: from chatbot to the real world of disinformation
Elias Thorne’s presence transcended the chatbot environment. Daniel May noted Elias Thorne’s name as the author of alternative medicine manuals, YouTube algorithm guides, books on Greek mythology, and even a psychological thriller novel on Amazon. “No human writes all of this,” May noted, highlighting the danger of harmful advice that can emerge from these posts.
The proliferation of AI-generated books on Amazon in recent years has included content containing misinformation and gross errors, causing problems for librarians. Elias also broke into the “slop world” of YouTube and fake news sites. In videos with low-quality illustrations, he appears as an 83-year-old sergeant major or as a man shot by his wife in a reptile sanctuary. He is often portrayed as a tragic, wronged and elderly figure, although the name was also used in a human story shortlisted in a BBC children’s competition.
Security alignment of AI models explains the preference for ‘safe’ narratives
The question of why LLMs seem to resort to the lighthouse theme despite having access to all the world’s literature lies in developers’ attempt to sanitize and align model outputs to ensure safety. Hamilton explains that many stories on WildChat were not “safe for work.” This led the researchers to hypothesize that models, when going through alignment, end up preferring a small fraction of WildChat’s “safe” stories, like a bottleneck. It’s not that Elijah’s stories are frequent in absolute terms, but rather that they are consistently “safe”, making them a standard output for models.
The researchers plan to explore this theory in future investigations. Interestingly, there was an Elias who came close to the LLM stories in real life: Elias Allen, a 16th-century clockmaker in London, as well as a time-traveling mad scientist in the Dinosaurs Attack! card series. from the 80s, showing that the name already had some cultural presence before the era of generative AI.
