Cognitive Offloading and Parasocial AI Dynamics: Analyzing How Human Relational Trust Alters Risk Assessment and Cognitive Effort During LLM-Assisted Decision-Making

As Large Language Models (LLMs) increasingly act as active conversational collaborators, a critical paradox emerges: the relational warmth and empathy designed to make AI systems intuitive may systematically degrade human critical oversight and metacognitive monitoring. This study investigates the interactive effects of AI conversational framing (Transactional vs. Parasocial) and task stake levels (Low-Stakes vs. High-Stakes) on human cognitive offloading, verification behavior, and automation bias. Using a 2x2 factorial between-subjects, within-subjects design with N=300 participants, we exposed users to standardized decision scenarios containing planted, discoverable logical flaws spanning routine workplace tasks, personal financial allocations, and medical lab interpretations. Results demonstrate that participants exposed to Parasocial AI framing were significantly less likely to detect planted logical flaws (OR=0.42, p<.001), exhibited shorter verification latencies, and displayed higher rates of miscalibrated overconfidence, particularly in high-stakes domains. These findings indicate that artificial emotional intimacy acts as a cognitive occluder, inducing epistemic deference and suppressing natural error-checking mechanisms. We discuss implications for responsible AI interface design and propose design guidelines incorporating deliberate verification friction in high-stakes operational environments.