Fluent dependence: A validity framework for human capability in human-AI interaction
Jeffrey E. Anderson, Sam Zaza · Computers in Human Behavior Artificial Humans · 2026
Fluent dependence describes a state in which an individual produces output that meets evaluative criteria through interaction with an artificial agent without having developed the cognitive capability that the output appears to demonstrate. Recent empirical work has documented patterns consistent with this state across multiple contexts of human-AI interaction (i.e., high school mathematics, university scientific inquiry, knowledge worker decision-making, consumer category competence) though these have been discussed in fragmented terms. Drawing on cognitive load theory, the paper proposes a mechanism by which sustained interaction with artificial agents may produce fluent dependence: when intrinsic load is transferred from the human to the artificial agent, the germane processing that produces durable schema construction may not occur. One plausible mechanism is proposed here, while recognizing that other mechanisms may also contribute. Drawing on construct validity theory, the phenomenon is framed as a validity problem: work produced through human-AI interaction supports inferences about joint human-agent performance but not about human capability alone. Paired-task assessment is proposed as a detection method, in which performance with and without the artificial agent on construct-equivalent tasks is compared to produce a validity signal that, with appropriate aggregation, may support inferences at the individual level. Withdrawal is developed as a structural design feature for both assessment and learning systems. Domain applications in education, professional development, and consumer behavior demonstrate the construct’s generality and specify a research agenda for empirical validation and intervention work.