Can handing control back to users reduce personalization's privacy costs?

In a 3×3 between-subjects experiment (N = 450), participants used an LLM agent that acted on their behalf in an interpersonal discussion task, under one of three autonomy levels (No Autonomy, Full Autonomy, Risk-Contingent Autonomy) crossed with one of three personalization conditions (Full Personalization, No Personalization, Privacy-Aware Personalization). The authors report that "risk-contingent autonomy, where the agent delegates control back to users upon detecting potential privacy leakage, improves users' perceived control," and that "this in turn attenuates personalization's adverse effects: privacy concerns rise less and trust declines less." Full personalization without regard to privacy preferences raised privacy concern and lowered trust, consistent with prior work the authors cite; risk-contingent autonomy flattened both effects.

The mechanism the paper gives is perceived control, one of two dimensions of human autonomy it draws on (the other being oversight effectiveness). The finding is explicitly counterintuitive on this point: "users interacting with risk-contingent autonomy agents reported greater perceived control than those with no-autonomy agents, despite technically having less direct oversight" — the No Autonomy condition required approval before every message sent. The authors attribute this to decision fatigue and responsibility offloading under constant-approval designs, versus a design that "delegates control when necessary," concentrating human review at detected risk points instead of spreading it evenly across every action.

This gives a design lever to the trade-off described in Does chatbot personalization build trust or expose privacy risks?: where that longitudinal study treats rising trust and rising privacy concern as two sides of the same personalization process, this experiment shows the pairing is not fixed — autonomy design, not just personalization technique, can move both legs of the trade-off at once. It also complicates a design intuition drawn from What makes people distrust AI agents they delegate to?: that study ties trust withdrawal to the absence of a pre-send approval step, but here the condition that always required approval (No Autonomy) scored lower on perceived control than the condition that required approval only at detected privacy risk — approval everywhere is not the same design choice as approval where it matters, and the latter tested better on this measure.

The study was run in a simulated Qualtrics environment on two scripted scenarios (a professional meeting, a personal trip discussion), with personal information bounded to what participants disclosed in a pre-survey; no genuine privacy leak occurred, and the authors state the setup "does not fully reproduce the risk perception of a genuine leak." The reported means describe this task set and this sample, not agent users generally, and the paper names real-world deployment studies as necessary follow-up. Within that limit, the result supports a narrower design implication than "give users more control": it supports triggering control hand-back specifically at moments of detected privacy risk, rather than requiring constant approval or none at all.