Where Does Responsibility Go When Agency Is Blurred?

When an AI-mediated system causes harm, responsibility may concentrate on the AI, fall disproportionately on a nearby human operator, or dissipate across the system. We propose a common mechanism: interfaces alter how an evaluator represents a distributed event, thereby changing perceived agency, its normative significance, and warranted responsibility. For evaluator $r$, $W^r=W(\widetilde X^r;p^r,N^r)$ separates the represented event $\widetilde X^r$, perceived agency $p^r$, normative principles $N^r$, and the resulting responsibility allocation $W^r$. \emph{Dyadic compression} reduces a sociotechnical network to an apparent-agent--moral-patient relation. We derive three responsibility regimes and introduce \emph{responsibility legibility}: making control, authorization, benefit, duty, and remedial capacity perceptually available without dissolving accountability into exhaustive complexity. A four-condition experiment distinguishes this intervention from technical explainability and identifies whether it changes responsibility through perceived agency, normative salience, or both.