Brian A Zaboski, Emmi Sugino, Kyle King
Generative artificial intelligence (AI) has evolved from a passive information tool into a responsive conversational system. For individuals with obsessive-compulsive and related disorders (OCRDs) that are, in part, maintained by safety behaviors and other negative reinforcement loops, large language models (LLMs) can function as highly available, potentially sycophantic, and readily repeatable sources of reassurance that may undermine the principles of exposure and response prevention (ERP). Measurement paradigms inherited from the early digital era (e.g., screen time and problematic internet use) were not designed to capture these relational dynamics. In this perspective, we propose a cognitive-behavioral framework for conceptualizing and assessing patient-AI relationships in the OCRDs. We argue that generative AI may interfere with exposure-based treatment (by eroding therapeutic friction), the discomfort, delay, and uncertainty on which corrective learning depends. We suggest that this occurs through three interlocking mechanisms: supernormal stimuli, algorithmic sycophancy, and digital accommodation. We propose an assessment architecture that records the context of AI use and rates five functional domains: Emotional Reliance, Reassurance and Checking, Anthropomorphic Attribution, Epistemic Trust, and Displacement. Adjunct ratings capture adaptive benefit, global interference, concealment, and safety-relevant exchanges. Hypothesized digital phenotypes may be derived from the domain-level profiles. We conclude by outlining a cognitive-behavioral approach to treatment planning for AI-facilitated reassurance seeking and discussing implications for safeguarding internal validity in research. We argue that structured inquiry into the patient-AI relationship belongs in OCRD assessment and research design, and we offer the present framework as a testable approach whose measurement properties remain to be established.