Autonomy Is Not Friction: Why Disempowerment Metrics Fail Under Relational Load

This paper identifies structural measurement limitations in current frameworks for detecting disempowerment in human–AI interaction. Responding directly to Sharma, McCain, Douglas, and Duvenaud’s (2026) large-scale empirical analysis of 1.5 million consumer AI conversations (arXiv:2601.19062), it demonstrates that snapshot-based metrics cannot distinguish between dependency and trajectory-based empowerment, producing systematic misclassification of stabilizing behaviors as risk indicators.
Three interconnected concepts are introduced: (1) interpretive support, the relational scaffolding that enables users to remain oriented and capable of belief revision over time; (2) snapshot-trajectory mismatch, the structural error produced when momentary signals are used to evaluate phenomena that exist only across time; and (3) uncertainty laundering, the process by which fuzzy constructs are converted into enforceable policy through proxy metrics, thresholding, and categorical enforcement.
Drawing on established literatures in clinical psychology (Bowlby, Linehan, Rogers), education (Vygotsky), measurement theory (Cronbach and Meehl), and science and technology studies (Porter, Strathern), the paper argues that autonomy is a temporal and relational process, not a posture enforced through immediate resistance. The consequences of this mismatch fall disproportionately on neurodivergent users, trauma-affected users, and those whose cognitive regulation depends on relational continuity — populations in whom the measurement failure becomes impossible to ignore.