Dark&Stormy: Modeling Humor in the Worst Sentences Ever Written
This study addresses the challenge of detecting deliberately crafted “bad humor” in English—a genre where state-of-the-art humor detection models exhibit significant performance degradation. To tackle this, we introduce the first bad-humor corpus derived from the Bulwer-Lytton Fiction Contest, systematically analyzing its structural patterns involving puns, irony, metaphor, and metafictional devices. We conduct the first human–LLM comparative study on bad-humor generation, revealing that LLMs over-rely on specific rhetorical devices and nonce collocations, exposing a rhetorical control bias. Integrating literary rhetorical analysis, controllable prompt engineering, and human–AI collaborative evaluation, we demonstrate that current models lack robust semantic–stylistic disentanglement capabilities for low-quality humor. Our contributions include: (1) a novel, manually annotated bad-humor benchmark; (2) empirical evidence of LLMs’ rhetorical limitations; and (3) open-sourced data and code to advance computational humor research.