Humor as a Constraint-Failure Diagnostic in Large Language Models

This deposition contains a formal academic paper and associated primary transcripts examining humor-like behavior in large language models (LLMs) as a diagnostic phenomenon rather than a cognitive capability. The study evaluates four instruction-following language models—three cloud-hosted frontier systems and one local quantized model—using a fixed five-part prompt battery designed to stress semantic compression, timing, meta-awareness, and intentional failure. Identical prompts were administered without retries, clarifications, or temperature adjustments. The included logs preserve verbatim model outputs for all five prompts across all four systems, enabling independent verification and secondary analysis. Results indicate that humor does not arise from an internalized “sense of humor,” but instead emerges at points of constraint failure, optimization conflict, or reduced epistemic anxiety. Models optimized for safety and pedagogical clarity consistently suppress humor via explanation and self-correction, while lower-guardrail or lower-capacity systems occasionally produce higher-quality humor through risk-tolerant or accidental behaviors. Humor is thus proposed as a behavioral stress test revealing internal model priorities, constraint sensitivity, and optimization boundaries.