Study finds 11 words dominate 88% of LLM stories, revealing low diversity

Arxiv CS.CL (NLP/language models)

Researchers sampled 20,000 stories from ChatGPT, Claude, Gemini, and AI2's chatbot, finding 11 words (including 'Elias', 'lighthouse', 'clockmaker') appear in 88.3% of outputs. The paper posits safety and alignment tuning may cause this low diversity.

LLMChatGPTClaude

See 5 more sources

His name is Elias. He lives in a lighthouse. He is probably a clockmaker. He is about to learn...5 hours agoNav Toor

Constrained Semantic Decompression in LLMs through Persian Proverb-Conditioned Story Generation5 days agoZahra Habibzadeh, Paria Khoshtab, Amir Mesbah, Yadollah Yaghoobzadeh

Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories21 days agoSil Hamilton, David Mimno

Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories16 hours agoannodominiDiscuss

Chatbots Keep Telling Stories About Lighthouse Keeper 'Elias Thorne'. We Might Know Why6 days ago404mediacoDiscuss

Arxiv CS.CL (NLP/language models)·Ramya NamuduriManya WadhwaAnshun Asher ZhengGreg DurrettJunyi Jessy Li

5 days ago