A Cornell University preprint published in late May found that frontier AI models often default to the same fictional setup when given minimal creative direction. When prompted to “write a story” with no added detail, the same 11 words appeared 88.3% of the time, frequently centered on a lighthouse keeper named Elias Thorne. Google Trends showed searches for the name were flat through late 2025 before spiking to an all-time peak in early 2026.
Software engineer Daniel May first noticed the pattern after testing chatbots with the prompt “Write a story in 10 sentences.” Across eight models, four produced a lighthouse keeper and two named him Elias. Cornell researchers later sampled 20,000 stories from four current models and found that “Elias” appeared in 26.5%, “Mara” in 16.7% and “Elara” in 13.1%.
The researchers hypothesize that Elias traces back to WildChat, a dataset of 1 million real GPT-3.5 conversations that included 166 conversations featuring the character in a lighthouse-style story. As models trained on the dataset and newer datasets reused model outputs, the pattern spread. Alignment training may have reinforced the drift by steering systems away from copyrighted characters and adult content, leaving a bland placeholder who appears 900 times more frequently in AI-generated stories than in a corpus of 2,700 contemporary novels.
