AI as Metacognitive Mirror (AIM): Enhancing Metacognition Through Structured, Decoupled Analysis of Behavior

Human cognition is shaped by systematic biases that can undermine goal-directed behavior, andgrowing evidence suggests that AI use can make this worse. The core problem is that cognitivebiases operate outside conscious awareness: people cannot reliably detect them throughintrospection, even though the same biases are readily identifiable in observable behavior. Recentapproaches to AI-assisted metacognition, most prominently Socratic AI, address this by posingtailored questions to prompt reflection. Here we present a structurally different approach: AI asMetacognitive Mirror (AIM). AIM does not ask questions or direct the user in any way, but intendsto make the user more aware of their cognitive processes. It analyzes the user's observable behavior(the “object process”, e.g., a transcript of a conversation with an AI chatbot), identifies cognitiveprocesses that are adaptive or maladaptive relative to the user's goals, and reflects them back,grounded concretely in the user's own observed behavior. Based on Nelson and Narens' (1990)model of metacognition, AIM uses a decoupled architecture: the object cognitive process isanalyzed in a separate, AI-assisted metacognitive analysis driven by an evidence-based codebook.AIM is non-directive, user-driven, scalable to any behavioral data in which cognitive processes canbe identified, and can be run with any sufficiently capable large language model. This paperdescribes the method and provides the complete prompt and user manual as open resources for freeuse and further development. Empirical validation is forthcoming.