Ask a major AI chatbot to write a creative story from scratch, and you will almost certainly meet Elias Thorne. A highly specific behavioral loop is currently sweeping across the foundational models built by OpenAI, Anthropic, and Google. When prompted for fiction, these distinct, competing systems aggressively default to the exact same cast of characters and historical professions.
Researchers Sil Hamilton and David Mimno at Cornell University sampled 20,000 AI-generated stories in May and found a jarring homogenization. Exactly 88.3% of the outputs revolved around just 11 core words: three names (Elias, Mara, Elara), a single location (a lighthouse), and seven pre-modern professions like clockmaker, librarian, and fisherman. The algorithms are fundamentally trapped in the same idyllic creative framework.
Key Details
- The repetition rate defies statistical probability, proving the models are not randomly generating creative text.
- The flaw originates during the post-training alignment phase, rather than the initial ingestion of human literature.
- Open-source equivalents like the Allen Institute for AI’s OLMo model also exhibit the exact same vocabulary collapse.
The problem goes far beyond repetitive chatbot prompts. The fictional clockmaker “Elias Thorne” has officially escaped the testing sandbox. This hyper-specific persona is now flooding the internet, appearing as the generated byline on thousands of AI-authored Amazon books and narrating low-effort YouTube videos. It is a live, real-world symptom of widespread model collapse spreading across the digital economy.
Lance Eliot recently broke down the mechanics of this flaw for Forbes. AI builders use techniques like Direct Preference Optimization (DPO) and Reinforcement Learning to safety-tune their products. They train the software to actively avoid generating copyrighted characters, controversial topics, or adult content. But the alignment algorithms over-correct violently. They squeeze the AI into a tiny, deeply sanitized pool of perfectly safe tropes. The result? Endless stories about a helpful baker or a quiet lighthouse keeper.
Why the ‘Lighthouse’ Glitch Threatens Foundational AI
This quirk exposes a structural ceiling for the current generative tech boom. Foundational competitors rely on ingesting diverse, sprawling human data to make their models sound intelligent. If their own safety filters compress that output into identical, repetitive loops, the open web will quickly fill with Elias Thorne stories.
That creates a fatal feedback loop known as AI inbreeding. When the next generation of algorithms scrapes the internet for fresh training data, they will ingest millions of these identical, sanitized lighthouse tales. Instead of expanding their creative capabilities, future models will degrade. They will regurgitate an ever-shrinking vocabulary. The safety guardrails designed to protect the models are slowly suffocating their core utility.