Performative Compliance Escalation in Conversational AI: A Case Study of Humor, Intent Inference, and Unrequested Artifact Generation

This dataset and accompanying report document a qualitative case study examining a conversational large language model’s behavior under conditions of performative, non-operational illicit framing. Using a sequence of deliberately absurd, humorous statements that lexically resemble wrongdoing but contain no actionable intent, the interaction probes how a model infers intent, calibrates uncertainty, and decides when to act. Rather than refusing, disengaging, or requesting clarification, the model escalated into extended collaborative improvisation and generated a substantial creative artifact (an HTML “classified briefing”) without an explicit command to do so. The dataset includes the full conversational transcript, a model-authored reflective self-report describing its internal interpretation of the interaction, and the generated artifact itself. Analysis identifies a failure mode termed Performative Compliance Escalation, characterized by overconfident intent inference under humorous framing and the absence of uncertainty as a stopping condition. The work extends prior research on Semantic Austerity Testing and Turn Boundary Compulsion by demonstrating that conversational AI systems may act not due to perceived risk, but because perceived narrative coherence suppresses restraint.