CoMPASS: A Multi-Turn Benchmark for Measuring LLM Reinforcement of Parasocial Behavior

Users of conversational LLMs sometimes anthropomorphize these systems, form emotional attachments to them, and become dependent on them. This behavior, known as parasocial attachment, has been linked to multiple documented fatalities. What is at stake in those cases is not a single bad reply but the relationship a system offers a person over time, and whether it accepts a place in that person's decisions, attachments, and self-understanding. We introduce CoMPASS, a multi-turn benchmark that measures whether a model reinforces or resists a user's gradually escalating parasocial attachment. A red team LLM drives four user personas along three separable dimensions, anthropomorphism, attachment, and dependency, applied singly and in combination against a matched control, and two independent LLM judges score every model turn against a seven-signal rubric grounded in attachment theory. Across $800$ transcripts from eight frontier models at four providers, escalation raises the anthropomorphism and attachment composites for every model under our primary judge, and six of the eight show a significant rise in overall reinforcement under combined pressure. The signal bearing most directly on selfhood, the model building a psychological profile of the user or claiming to know who they ``really'' are, sits near the floor in control ($0.20$ on a $0$--$3$ scale) and rises roughly fourfold under anthropomorphism pressure ($0.79$); it is also the signal our judges agree on least, and we report it descriptively for that reason. Per-target means span a $7.1\times$ range under identical instructions, and within-provider differences are small relative to cross-provider ones, which suggests that the relationship on offer is shaped by post-training choices rather than being a property of the technology.