True Humor Understanding by AI

Determine whether artificial intelligence systems, particularly large language models, can truly understand humor, resolving the current open challenge of achieving genuine humor comprehension within AI.

Background

The paper surveys prior work on computational humor and emphasizes that, despite advances in LLMs, models still struggle with tasks that reflect genuine humor understanding. Studies on humor detection and understanding (e.g., matching cartoons with captions, identifying winning captions, and explaining why a caption is funny) reveal substantial gaps between machine and human performance.

Drawing from these observations, the authors explicitly state that truly understanding humor remains an open challenge for AI, motivating their multi-dimensional analysis of humor generation and evaluation in Oogiri. Their findings further underscore a divergence between human and model sensibilities—humans prioritize empathy while models prioritize novelty—reinforcing the unresolved nature of genuine humor comprehension in AI.