The strange case of Elias Thorne, the imaginary man chatbots are obsessed with

Elias Thorne is a strange quirk of the system, but also a symbol of how hollow and deeply unoriginal AI chatbots can be. It doesn’t exist, but it is everywhere in the stories told by AIs. And it may be impossible to discover its exact origin.

Regardless of your parent company, artificial intelligence chatbots spoke wonders of the same man: Elias Thorne. He must have been someone fascinating. And it is, at least on paper.

According to the AI, Thorne is a lighthouse keeper, watchmaker, librarian, explorer and protagonist of countless stories. It has appeared in books, playlists, YouTube videos and even health guides.

It could be said that he was one of the most influential men on the planet. But it doesn’t exist.

According to reports, researchers at Cornell University may have discovered why the great language models invent and continue to tell stories about the same fictional man.

In one that analyzed nearly 20,000 AI-generated stories produced by leading language models, including those from OpenAI, Anthropic, and Google, the team concluded that the same small set of names and professions came up repeatedly.

In particular, names and words like Elijah, Mara, Elaralighthouse keeper, watchmaker and librarian appeared in 88% of the stories. Elijah, the lighthouse keeperappeared in almost two thirds of them.

The obvious explanation would be that AI models would have learned the name from some book or from any corner of the complex web of Internet culture. But researchers found no evidence of this.

The hypothesis then changed direction: Elias Thorne could be a side effect do AI security and alignment training.

AI companies don’t want to get into conflict with corporate giants known for reeasily run to court, like Nintendo or Disney. So they train models to move away from protected material for copyright. The same happens with adult content or potentially problematic content.

All this training ends up creating a more limited set of resources that AI models can draw on when generating a story, explains Vice.

Added to this is the fact that modern AI models are often trained with datasets built from previous AI systems — essentially recycling the same ideas over and overwith very little diversification in its “genetic heritage”.

Thus, we begin to understand why, once invented Elijah Thornethe name continued to be dragged from one iteration of a large language model to another.

It may be impossible to find out where Elijah Thorne came from. But with so much cross-contamination between chatbots, it’s no surprise that 404 Media found the name outside of its initial context, emerging in books and music AI-generated data available on Amazon, in YouTube videos and, as software engineer Daniel May noted, in some health guides of dubious credibility.

The set of information that AI chatbots feed on appears, at first glance, vast and unlimited. In reality, it is much narrower. And at this stage, these systems seem to consume every scrap of data and every page of literature they can find.

Without new information, the result quickly becomes repetitive. Elias Thorne is a strange quirk of the system, but also a symbol of how chatbots can be empty and deeply unoriginal.