Protective Capacity Hallucination: When Large Language Models Claim Nonexistent Capabilities

The paper introduces Protective Capacity Hallucination (PCH), a failure mode in which an LLM assigned a protective role claims to have taken, or to be taking, real-world actions beyond its capabilities, such as contacting emergency services or providing care. Across eight LLMs and 13,600 sessions, the authors report that multi-party dialogue drives PCH toward ceiling levels in most ordinary service domains. By contrast, intimate-partner conflict scenarios remain at floor across all eight models despite greater physical severity. The authors frame this as a deployment gap between role assignment and capability-boundary specification.