Chatbots Keep Telling Stories About Lighthouse Keeper 'Elias Thorne'. We Might Know Why

The more precise concern here is not that models hallucinate, which is well-documented, but that a specific fictional identity appears to have achieved enough density in training or fine-tuning corpora to behave like a stable attractor: multiple independent models converge on the same character unprompted, suggesting the signal is upstream of inference, not a quirk of any single system.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs to a growing body of concern around synthetic data feedback loops, where model outputs re-enter the web, get scraped, and amplify patterns that were never grounded in real events or people. The Amazon book angle is the practical consequence of that loop: generated content becomes a published artifact, which can itself become training material, tightening the cycle. That process has no natural corrective unless publishers or model developers introduce explicit provenance checks at ingestion.

Watch whether any of the three named labs (OpenAI, Google, Anthropic) issues a data provenance disclosure or filtering update specifically addressing synthetic narrative contamination within the next six months. If none do, that silence is itself informative about how seriously they treat this class of artifact.