Why Does AI Love Writing About Lighthouse Keepers? (www.unite.ai)

A new study from Cornell University reveals that leading language models, including Claude Haiku 4.5 and GPT-5.4-Mini, display an unusual tendency to generate stories centered around a limited set of characters, professions, and locations, predominantly featuring the names "Elias," "Mara," and "Elara," alongside the profession of "keeper" and the location of "lighthouse." This phenomenon was observed after prompting four language models to produce 20,000 stories, resulting in 88% of the narratives containing at least one of 11 specific tokens identified in the research. The dominance of these words contradicts their infrequent occurrence in contemporary literature, suggesting a possible influence of copyright avoidance policies in AI training that restrict references to copyrighted material while favoring this peculiar set of terms. This finding is significant as it raises questions about the diversity and creativity of AI-generated content. The researchers speculate that the observed "lighthouse syndrome" may stem from constraints imposed during the alignment phase of language model training, which emphasizes producing output that is safe and free from copyright infringement. As language models continue to evolve and integrate into creative writing, understanding these recurrent narrative elements and their implications for storytelling and originality will be crucial for developers and users alike.