Does attributing agency to AI actually reduce human sense of control?

Leonardi argues that the decisive event in human-AI relationships is not a change in what a system can do but the moment a human attributes agency to it. He writes that "when humans attribute agency to AI systems, they experience systematic 'power displacement'—reduced sense of control and responsibility—even while retaining formal authority," producing a paradox in which "the more autonomous AI appears, the less autonomous humans feel, regardless of their actual control." He cites supporting experiments: in Ciardo et al. (2020), participants working alongside a humanoid robot presented as an intentional agent "reported significantly lower sense of agency over task outcomes compared to working alone" even when the robot never actually acted; in Jia et al. (2022), participants "consistently credited themselves for successful outcomes when working with AI systems but blamed the AI for failures," an asymmetry Leonardi reads as evidence that the displacement is motivated rather than simply descriptive.

The mechanism he proposes is that this displacement is "neither permanent nor unidirectional" but circulates through a five-phase, recursive "agency loop": delegation (humans "design, implement, configure, or deploy" a system, holding maximum felt agency as architects of its constraints); attribution (as in radiology or hiring examples, users shift from describing the system as executing their criteria to crediting it with independent judgment — "the algorithm selected these top applicants"); contingency ("technical failures, unexpected outcomes, ethical violations, legal challenges, or public controversies" that "shatter the illusion of autonomous AI agency"); reassertion (humans "exercise explicit control... and accept accountability" for both the AI's past actions and their own corrections); and reconfiguration (delegation practices are adjusted, setting conditions for the next cycle). Loops "nest within each other" across "multiple temporal scales" and are explicitly "evolutionary processes" rather than a one-time transfer.

This runs in a different direction from Does incremental AI replacement erode human influence over society?, which treats eroding human influence as a largely one-way, interdependent drift across societal systems; Leonardi's loops instead frame displacement as reversible and strategically manageable at the point of contingency and reassertion. It also sits near Does AI autonomy change how we assign blame?, which measures third-party blame judgments shifting with an AI's apparent autonomy in a controlled vignette study — a parallel finding to Leonardi's claim that perceived autonomy redistributes attribution, though Leonardi's framework concerns a first-person, longitudinal sense of power rather than one-off causal judgments of an observer. It also bears on Does granting agents more autonomy undermine human oversight?: both describe autonomy reducing felt human control, but where that note treats the loss as a design-driven skill erosion, Leonardi argues the loss is recoverable through the reassertion phase of the loop.

The excerpt references three organizational cases (radiology, AI scheduling assistants, police predictive analytics) and a comparative Table 3 to support the loop's path-dependence and evolutionary claims, but the case methodology and underlying data are not included here, so the specific mechanisms behind each case's outcome cannot be verified from this excerpt alone. The framework itself is also a conceptual contribution built by reinterpreting prior psychological experiments (Ciardo, Jia) rather than new data Leonardi collected to test the loop model directly; whether the five phases hold up as a general description of human-AI dynamics beyond the cases discussed is argued, not demonstrated, here.