Meihe Xu

Personalized privacy assistants (PPAs) are emerging AI assistants that hold the promise to manage privacy on users’ behalf. Prior to the design decisions and deployments of a PPA in different regions, it is important to understand prospective user acceptance and factors contributing to users’ behavioral intention to adopt the assistant. In this paper, we report a cross-regional survey study with participants from China, Germany and California analyzing a UTAUT2-informed acceptance model for PPA acceptance. The results suggest that cross-regional comparability of PPA acceptance cannot be assumed. While partial measurement invariance supports meaningful structural comparisons between California and Germany, compositional invariance was not established for several constructs when comparing the Chinese sample with Californian and German samples. Together, we raise three provocations based on these findings: 1) Cross-regional acceptance comparisons might often be invalid by design; 2) it is time to reconsider acceptance models for AI assistants for privacy management, and 3) some thoughts on the future of PPAs.

One-size-fits-all privacy notices and data decisions may no longer suffice for modern privacy management. Personalization, through personalized privacy assistants (PPAs), offers a potential solution. These assistants aim to provide tailored transparency disclosures and assist individuals in making decisions about their data. To actualize a future with PPAs, this paper presents findings from two expert focus groups that explore the challenges of developing these tools. Guided by the three dimensions of agency, accuracy, and accountability, and utilizing backcasting activities, we identified that hybrid PPAs are more desirable and welcomed by experts. However, the approaches to achieving a hybrid PPA that tackles concerns on accuracy, agency, and accountability were not agreed upon. Second, we highlight three main takeaways for successfully implementing PPAs: (1) user agency must be respected, (2) multiple stakeholders should be involved in developing, operating, and improving PPAs, and (3) interoperable standards are fundamental for successful deployment. We conclude that PPAs, and likely any technology concerning privacy protection, cannot be achieved by one stakeholder alone.

Individuals increasingly face an overwhelming number of tasks and decisions. To cope with the new reality, there is growing research interest in developing intelligent agents that can effectively assist people across various aspects of daily life in a tailored manner, with privacy emerging as a particular area of application. Artificial intelligence (AI) assistants for privacy, such as personalized privacy assistants (PPAs), have the potential to automatically execute privacy decisions based on users' pre-defined privacy preferences, sparing them the mental effort and time usually spent on each privacy decision. This helps ensure that, even when users feel overwhelmed or resigned about privacy, the decisions made by PPAs still align with their true preferences and best interests. While research has explored possible designs of such agents, user and expert perspectives on the acceptability of such AI-driven solutions remain largely unexplored. In this study, we conducted five focus groups with domain experts (n = 11) and potential users (n = 26) to uncover key themes shaping the acceptance of PPAs. Factors influencing the acceptability of AI assistants for privacy include design elements (such as information sources used by the agent), external conditions (such as regulation and literacy education), and systemic conditions (e.g., public or market providers and the need to avoid monopoly) to PPAs. These findings provide theoretical extensions to technology acceptance models measuring PPAs, insights on design, and policy implications for PPAs, as well as broader implications for the design of AI assistants.

Transparency is not a universal concept in its formulation or application. For example, transparency can be prospective or retrospective. Prospective transparency gives information about a process pre-emptively—that is, prior to the envisaged actions—while retrospective transparency gives information retroactively after the process is completed and thus serves as more of an explanation. ¹ Although each comes with its own conceptions of the information to be provided, prospective and retrospective transparency requirements can also complement one another.