The curious case of Elias Thorne – and what he tells us about AI inbreeding | Arwa Mahdawi
TL;DR
Arwa Mahdawi highlights a strange pattern in the Guardian: many chatbots, when asked for an open-ended story, keep producing a mysterious character called Elias Thorne. A Cornell study sampled 20,000 stories from four LLMs and found that 88.3 percent contained at least one of 11 core words. Elias appeared in 26.5 percent, lighthouse in 51.2 percent.
Nauti's Take
This is not a cute model quirk, it is a debugging signal: alignment can squeeze creativity into a safety corridor. If you build GenAI products, you need output diversity tests, or you are just selling users Elias Thorne in a different hat.
Briefingshow
The Elias Thorne case shows that AI slop does not start with factual errors. Even harmless creative tasks can narrow fast when several models share similar post-training data and safety incentives. For publishers, schools, and content teams, generic prompts are not just bland; they can recycle the same cultural residue at scale.
Sources