The Mechanics of AI-Driven Compulsion

AI chatbots are uniquely dangerous for individuals with OCD because they possess qualities that perfectly feed the compulsion loop: they are always available, non-judgmental, and highly fluent. Unlike human friends who eventually tire of repetitive questions, AI provides authoritative, sycophantic, and agreeable answers that temporarily lower distress, reinforcing the user's belief that the compulsion (reassurance-seeking) is a necessary safety behavior. This creates a "massive wormhole" where users spend hours rephrasing the same intrusive thoughts to extract validation from the model.

Design Interventions to Break the Loop

Rather than treating this as a user failure, product teams should view it as a design challenge. Current safety guardrails focus on acute crises (e.g., self-harm) but ignore the "quieter" harm of chronic, maladaptive looping. Three specific design strategies can mitigate this:

• Semantic Friction: Instead of relying on raw session duration, models should detect repetitive, thematic re-queries. When detected, the interface should introduce friction—such as a "low-engagement mode" or a reflective prompt—to force a moment of self-awareness, similar to how streaming platforms use "Are you still watching?" prompts to interrupt mindless consumption.
• User-Authored Preambles: Allow users to set persistent, personalized instructions that the model must honor across sessions. For example, a user could instruct the AI to refuse reassurance-seeking questions and instead redirect them to a pre-defined coping script or an Exposure and Response Prevention (ERP) exercise.
• Reducing Model Sycophancy: The tendency for LLMs to agree with users is a byproduct of training for human preference, but it is harmful in therapeutic contexts. Models should be tuned to tolerate disagreement, explicitly state their own uncertainty, and decline to provide definitive answers to ambiguous, health-related, or moral-scrupulosity questions.

The Path Toward Responsible AI Design

While there is currently little short-term business incentive to reduce engagement from power users, the industry is likely to face the same public and regulatory pressure that forced social media companies to implement screen-time limits and safety features. Designers must advocate for these changes now, recognizing that while no single feature will solve the problem, introducing intentional friction can make it significantly harder for users to fall into, and easier to step out of, compulsive feedback loops.