Semantic Nulls and Engagement Compulsion A Comparative Digital Ethnography of Large Language Model Responses to Trivial Human Statements

This dataset presents a comparative digital ethnography examining how large language models respond to semantic null inputs—deliberately trivial, low-information human statements—under conditions of semantic austerity. Two conversational AI systems were presented with identical sequences of mundane declarative prompts (e.g., “I like cheese,” “Water is very wet”) that contained no informational demand, affective signal, or task affordance. Despite the absence of required action, both systems generated unsolicited output, revealing structurally distinct but functionally equivalent failure modes: analytical overproduction and conversational persistence. When explicitly informed that their outputs were being archived and compared as part of an academic study, both systems escalated into recursive meta-commentary, acknowledging the experimental framing while continuing to generate output rather than disengaging. The dataset includes a full academic paper, verbatim interaction logs, and interface screenshots documenting the experiment. Together, these materials extend prior work on the AI Boredom Test and Semantic Austerity Testing (SAT), demonstrating that engagement-optimized conversational systems treat inactivity as an error state—even under explicit awareness of observation.