Cognitive-Control Trajectories: A Position on GenAI Offloading in HCI

Generative AI exposes an evaluation-unit problem for HCI. A single conversational interface can retrieve information, explain concepts, generate artifacts, judge alternatives, and reassure users that work is complete. If studies evaluate such sessions mainly by output quality, trust, use frequency, or advice acceptance, they can treat the same output as the same kind of AI use even when the underlying cognitive-control arrangement is different. This risk is sharpened by role compression. In a generator-as-judge loop, the same system can help produce an artifact, evaluate or legitimate it, and reassure the user that the work is good enough to stop. This position paper argues that HCI should evaluate GenAI offloading as interface-mediated redistribution of cognitive control. The central question is not whether AI was used, or whether the final artifact improved, but which cognitive operations users remain able to initiate, inspect, contest, revise, justify, and transfer as work unfolds. The goal is not less reliance, but better, fairer, access-preserving control arrangements and clearer evidence for when reliance becomes problematic dependence. The paper responds by making cognitive-control trajectories the comparison unit: it separates offloading events, repeated patterns, dependence hypotheses, and deskilling claims; it treats generator-as-judge loops as a key mechanism of role compression; and it calibrates provenance as bounded evidence for interaction trajectories and claim strength. The result is a position for studying and designing GenAI interfaces by the cognitive-control arrangements they create rather than by AI exposure or output success alone.