Ask eight different large language models to invent a fictional lighthouse keeper. Give them no guidance on the name. According to a widely-shared investigation, all eight return the same one: Elias Thorne.
That would be a curiosity worth filing away under "weird AI behavior" - except Elias Thorne now appears as the listed author on Amazon books selling cancer treatment advice.
How Models Converge on the Same "Default" Fiction
Large language models are trained on enormous amounts of text scraped from the internet. The models don't memorize specific passages outright - they learn statistical patterns about which words tend to follow which other words. When you ask a model to generate a plausible-sounding fictional name for a lighthouse keeper, it doesn't pull from a list. It generates whatever name has the highest statistical association with "lighthouse keeper" and "character" and "fiction" in its training data.
The problem is that the major commercial models - from different companies, built on different architectures - are all trained on largely overlapping internet text. They've all seen the same novels, the same writing forums, the same NaNoWriMo character name generators. When independent models converge on identical outputs, it's a signal that beneath the surface differences in tone and capability, these systems have absorbed the same patterns from the same sources.
Elias Thorne isn't a real person. He's what happens when multiple AI systems have collectively learned that a lighthouse keeper character should sound a certain way.
The Amazon Problem
Fictional name convergence would be a minor technical footnote if it stopped there. It doesn't.
Amazon's self-publishing platform (Kindle Direct Publishing) requires minimal verification of author identity. Someone - or something - used AI to generate books and attached "Elias Thorne" as the author. Those books are currently listed on Amazon and include titles offering cancer treatment guidance.
This is the direct line from abstract model behavior to concrete harm. AI tools have made it trivially cheap to generate plausible-sounding text and attach a plausible-sounding name. Amazon's content review has not kept pace with that volume. The result is that a name invented by committee across eight language models is now attached to medical claims that could genuinely hurt someone who acts on them.
Amazon has faced repeated criticism for AI-generated book spam in its catalog. The company added a disclosure requirement for AI-assisted content in 2023 but has not enforced it in any visible way. Searching for specific health conditions on Kindle Unlimited still returns dozens of books with no credible author history and suspiciously similar prose patterns.
What This Actually Means for AI Outputs
The Elias Thorne finding has a practical implication beyond the Amazon story: AI-generated text is less diverse than it appears. Two businesses both using AI to write product descriptions, customer personas, or training scenarios are producing more homogeneous output than either realizes. "Original" AI content may share structural fingerprints, default names, and stock phrasings with thousands of other AI outputs generated in the same week.
For anyone using AI tools to generate content at volume - marketing copy, fictional scenarios, persona research - it's worth running periodic audits to check how much your outputs actually differ from what competitors' AI tools would produce with similar prompts. The answer may be less than you'd want.