The curious case of Elias Thorne – and what he tells us about AI inbreeding | Arwa Mahdawi
TL;DR
Guardian columnist Arwa Mahdawi uses Elias Thorne as a warning sign: when asked for a simple story, popular LLMs including ChatGPT and Claude often generate a tale with the same shadowy character. Cornell researchers sampled 20,000 stories from four models. The name Elias appeared in 26.5% of outputs, while more than 88.3% shared the same 11 story elements, including lighthouse, keeper and clockmaker.
Nauti's Take
The Elias Thorne case is a useful antidote to the magic story around LLMs. The Guardian piece is a column, but the Cornell number is the hard center.
Open prompts show how quickly a model retreats into statistical comfort zones. Serious AI editing needs counterpressure: original examples, hard source work and human selection.
Without that, the web fills up with polite lighthouse stories wearing different labels.
Briefingshow
Elias Thorne does not prove that chatbots are secretly copying each other, but it exposes a quality problem: models can sound flexible while falling back to the same safe middle. For publishers, marketers and creators, the lesson is practical: AI copy needs real editorial pressure, or training-data patterns travel straight into books, videos and articles.
Sources