No matter the company, AI chatbots were raving about the same guy named Elias Thorne. He must be pretty fascinating. And he is, at least on paper. Depending on the AI, he’s a lighthouse keeper, a clockmaker, a librarian, an explorer, and the star of countless stories. He’s appeared in books, music listings, YouTube videos, and even health guides. You’d think he was one of the most influential men on the planet.

But he doesn’t exist.

According to reporting by fine folks at 404 Media, researchers at Cornell University may have figured out why large language models invent and keep telling tales of the same fictional man. In a study examining roughly 20,000 AI-generated stories from all the big LLM models, including OpenAI, Anthropic, and Google, the research team found that the same handful of names and occupations kept cropping up. Specifically, names and words like Elias, Mara, Elara, lighthouse keeper, clockmaker, and librarian showed up in 88 percent of stories. Elias the lighthouse keeper appeared in nearly two-thirds of them.

The obvious explanation is that AI models learned the name from some book or from somewhere in the tangled web of Internet culture. But the researchers couldn’t find evidence for that. So this theory shifted, this time to a side effect of AI safety and alignment training.

AI companies don’t want to run afoul of gigantic corporations that are notoriously litigious, like your Nintendos or your Disneys, so they run it through training that steers it away from copyrighted material. Same goes for any risky adult material. All that training creates a shallower pool of resources AI models can draw from when generating a story.

It May Be Impossible to Trace Elias Thorne’s Exact Origins

Add to that the fact that modern AI models are often trained on datasets built from earlier AI systems, essentially just rehashing the same old ideas again and again, with very little diversification in its gene pool, and it starts to make sense why once Elias Thorne was invented, he just kept getting dragged along from one iteration of an LLM model to another.

It may be impossible to trace where Elias Thorne came from, but with so much cross-pollination going on between chatbots, it’s no wonder that 404 media was able to find the name having broken containment, with the name cropping up in AI slop books and music available on Amazon, YouTube videos, and, as spotted by software engineer Daniel May, in some questionable health guides.

The pool of information AI chatbots pull from seems like it should be vast, limitless. It’s actually quite shallow, and by this point, they seemingly consume every scrap of data, every page of literature, they possibly can. Without fresh input, the output goes stale, and fast.

Elias Thorne is a weird quirk of the system, but also a symbol of just how hollow and deeply unoriginal chatbots can be.