This conceptual working paper argues that large language models are emerging as psychological infrastructure. It distinguishes psychological understanding from operational readability, introduces the concept of interpretation transfer (AI-to-AI exchange of psychological meaning), and uses recognition versus capture as the normative frame for evaluating whether AI-mediated readability serves the person or the system. The paper examines how different psychological traditions may be affected by AI-mediated interpretation and outlines governance principles for psychological AI, including the principle that psychological data should not be allowed to age into authority.