What privacy oversight patterns emerge when people review AI agent responses?

Cluster analysis of the same 300-person survey on LM agent privacy oversight "identified six privacy behavior patterns reflecting varying concerns, trust levels, and privacy preferences underlying people's oversight of LM agents' actions." The paper's own summary is blunt about what most of them share: "five patterns (except for Pattern A) reflect a lack of awareness to the privacy leakage in the actions of LM agents." Pattern D is named explicitly — participants in it "showed caution when writing their own responses but failed to identify privacy leakage in the LM agent's response, due to their emphasis on AI utility, optimistic attitudes, and high trust in AI capabilities." Patterns B and C look better on the surface, placing "more value on human qualities such as creativity and social ability, leading to better privacy decisions for now," but the paper discounts this: "most of them did not mention privacy in their justifications," so their safer choices were "more incidental" rather than privacy-aware.

The authors use the typology to argue against a specific design response: giving users more direct control over the agent's final action. Since five of six patterns trace to not noticing the leak rather than to a disagreement a user would act on if given control, "involving users in the process and giving them control to override the agent's suggestion might not be adequate" for most of these failure modes — control only helps the pattern that already has awareness. They also flag that B and C's better outcomes look contingent rather than durable: "As AI advances and meets their performance expectations, they might face the same privacy risks as those in Pattern D," i.e., incidental caution rooted in skepticism of current AI quality may collapse into Pattern D's overtrust once the agent's output stops giving people a reason to prefer their own draft.

This typology is the decomposition behind Do people notice when AI agents leak private information?: it explains the aggregate rate as a mix of distinct failure modes rather than one uniform lapse. Set against Does granting agents more autonomy undermine human oversight?, which names capacity erosion from autonomy and extended use as the mechanism behind oversight failure, this paper's clusters locate the failure earlier and more variably — in how much weight a person already gives AI utility versus human qualities when forming a judgment, before autonomy or extended use enters into it. It also runs parallel to Does co-design participation hide misalignment in preference agents?: both find overtrust coexisting with participants' own sense that they are evaluating carefully, in different domains (interpersonal privacy leakage versus household-preference representation).

The authors checked that the clusters are not simply an artifact of the six survey scenarios — all but the smallest cluster covered all six scenarios — but add that "it requires further research to tease out the impact of the scenario and the participants' individual characteristics," so how stable these six patterns are across other task types or populations is not established. The implication the data support is that "human oversight" is not a single capability to design for: an intervention aimed at Pattern D's overtrust (e.g., surfacing the specific leaked item) would likely miss Patterns B and C, whose safer choices do not route through privacy awareness at all.