Abstract
<p>We propose and examine a “Humorous Hill” pattern. The pattern suggests that the output of learning systems will show a U-shaped relationship between humor and quality. We tested the Humorous Hill by using language models to create novel examples of open-ended categories. The language modelsvaried in quality and architecture. Human participants (N = 450) rated the output of the language models for quality (category membership, goodness of fit), and humor. Across models and categories, we found an inverted U-shaped relationship between humor and accuracy. We evaluate the pattern against existing theories of humor and suggest that many do not obviously account for it without modification. We propose that the Humorous Hill applies generally to learning dynamics, including those found in children.</p>