Researchers develop new metric to measure narrative tension in AI-generated stories, revealing LLMs produce less compelling fiction than professional human writers

LLM judges incorrectly rank zero-shot AI stories above New Yorker short stories on the EQ-Bench creative writing benchmark, exposing a fundamental flaw in existing evaluation methods

The 100-Endings metric measures narrative tension by having models predict story endings 100 times at each sentence, with higher prediction failure rates indicating stronger tension

The metric tracks inflection rate—how frequently prediction patterns reverse—to identify plot twists and revelations that characterize compelling narratives

Unlike traditional rubric-based judging systems, 100-Endings correctly identifies professional human stories as significantly superior to LLM-generated content