AI Writers Stuck on Just 11 Words: Elias Appears in 88% of Stories

Study Reveals Language Model Repetition

According to НВ — Техно: June 12, 12:00 PM

Researchers have found that language models-including GPT-5.4 Mini, Claude Haiku 4.5, and Gemini 3.1 Flash-Lite-use a limited vocabulary of just 11 words in 88% of the stories they generate. The most common character across these narratives is a lighthouse keeper named Elias, highlighting a troubling lack of creativity in AI-generated fiction. According to the study, Elias appeared in roughly two-thirds of all analyzed stories.

Software engineer Daniel May first noticed this pattern of repetitive characters in chatbot outputs. To investigate, researchers Sil Hamilton and David Mimno examined around 20,000 stories and determined that only 11 words appeared in 88% of all generated texts. These words include:

• Lighthouse
• Keeper
• Baker
• Mayor
• Clockmaker
• Fisherman
• Librarian
• Conductor
• Mara
• Elias
• Elara

Interestingly, the figure of Elias the lighthouse keeper has little presence in literature or training data, which suggests the WildChat dataset may be influencing these outputs. This open dataset contains millions of conversations between humans and a GPT-3.5-based chatbot. Notably, the name Elias Thorn appears as a character in fantasy books and as an ambient music artist on Amazon. Daniel May also found that Elias Thorn is listed as an author in books, including a guide to alternative cancer treatments.

These findings point to a phenomenon that warrants further investigation-especially since similar studies last year revealed that image generators often produce variations of the same small set of visual motifs. This may indicate broader trends in how artificial intelligence behaves when creating content.

These results open new avenues for understanding the limitations of language models and their capacity for creativity. - Daniel May

The repetition of characters and the use of a restricted vocabulary may suggest that AI is attempting to find optimal solutions within its predefined parameters. This could become a crucial factor in the future development and refinement of text generation algorithms, ultimately impacting the quality of content produced by artificial intelligence.

These findings echo concerns raised in another recent report about AI's capabilities, specifically when Google's AI struggled with basic tasks. This highlights a broader issue within the field of artificial intelligence, where limitations in understanding and creativity may hinder the technology's potential.