What Security and Privacy Transparency Users Need from Consumer-Facing Generative AI

Abstract: Users increasingly rely on consumer-facing generative AI (GenAI) for tasks ranging from everyday needs to sensitive use cases. Yet, it remains unclear whether and how existing security and privacy (S&P) communications in GenAI tools shape users' adoption decisions and subsequent experiences. Understanding how users seek, interpret, and evaluate S&P information is critical for designing usable transparency that users can trust and act on. We conducted semi-structured interviews and design sessions with 21 U.S. GenAI users. We find that available S&P information rarely drove initial adoption in practice, as participants often perceived it as incomplete, ineffective, or lacking credibility. Instead, they relied on rough proxies, such as popularity, to infer S&P practices. After adoption, uncertainty about S&P practices constrained participants' willingness to use GenAI tools, particularly in high-stakes contexts, and, in some cases, contributed to discontinued use. Participants therefore called for transparency that supports decision-making and use, including trustworthy information (e.g., independent evaluations) and usable interfaces (e.g., on-demand disclosure). We synthesize participants' desired design practices into five dimensions to facilitate systematic future investigation into best practices. We conclude with recommendations for researchers, designers, and policymakers to improve S&P transparency in consumer-facing GenAI.

• The paper reveals that current S&P transparency measures in GenAI fail to effectively inform user decisions, impacting adoption and high-stakes usage.
• It utilizes 21 semi-structured interviews and participatory design sessions to uncover usability gaps in official S&P communications and documentation.
• The study proposes a multidimensional design space that prioritizes stakeholder access, independent evaluations, and clear data practices for enhanced transparency.

Security and Privacy Transparency in Consumer-Facing Generative AI: User Demands and Design Implications

Introduction

The rapid consumerization of generative AI (GenAI) systems has outpaced the evolution of effective security and privacy (S&P) transparency measures. Although prior work and regulatory developments have outlined technical vulnerabilities in GenAI—such as adversarial attacks, data extraction, and data sharing with third parties—user-facing S&P transparency mechanisms remain largely unvalidated and ineffective in practical adoption contexts. The paper "What Security and Privacy Transparency Users Need from Consumer-Facing Generative AI" (2604.17270) conducts a qualitative, empirical investigation into the transparency requirements and mental models of active GenAI users, identifies the shortcomings of current S&P communications, and synthesizes a multidimensional design space for future transparency mechanisms.

Methodology

The study leverages 21 semi-structured interviews and participatory design sessions with U.S.-based GenAI users. Participants are diversified by tool usage, context, and demographic background, with attention paid to pre-adoption and post-adoption phases of GenAI engagement. Data was analyzed via thematic coding, with iterative refinement of codebooks and inter-rater agreement ($\kappa \geq 0.7$). The research protocol was IRB-approved, and all transcripts were de-identified.

Figure 1: Overview of the study procedure, from participant recruitment to thematic analysis and prototype derivation.

Key Findings: User Motivations and Barriers

Adoption and Post-Adoption Drivers

Popularity, perceived utility fit, and price are primary adoption drivers; notably, users often conflate brand reputation/popularity with improved S&P protections, despite acknowledging the lack of specific evidence.

After adoption, sustained use is primarily oriented around output quality metrics (accuracy, bias, efficiency). However, S&P uncertainty—especially in high-stakes contexts such as legal, proprietary, or workplace use—directly constrains disclosure, personalization willingness, and may trigger discontinuation.

Crucially, S&P information almost never directly determines GenAI tool adoption in practice, contrary to the assumptions of many transparency frameworks.

User Information-Seeking Practices

Participants generally find S&P information from official sources to be inaccessible due to length and complexity, sometimes resorting to GenAI tools to summarize privacy policies. Independent sources such as technical news, Reddit, and community discussions serve as proxies for lived experience and policy changes, highlighting the perceived inadequacy and lack of credibility in official documentation.

Transparency Credibility and Effectiveness

Participants widely perceive current S&P communications as:

• Ineffective: Complex legalese leads to misunderstanding or information abandonment.
• Lacking credibility: Skepticism prevails regarding compliance and the honoring of toggled controls.
• Surface-level reassuring at most: Superficial engagement with policies provides minimal assurance, especially without external vetting.

This culminates in a context where risk mitigation is largely guesswork, with tangible trade-offs between privacy and the utility of personalization features.

User-Centric Transparency Wishlist

Content Requirements

Participants prioritize S&P transparency content in the following order:

1. Stakeholder access: Who—external parties, models, other users—can access user data, especially third-party plug-ins and model vendors.
2. Independent evaluations: Verified, comparative, digestible audits by credible third parties—preferably a hybrid of government and reputable private actors.
3. Data storage/training/inference practices: Visibility into what data is stored or used for training, including retention, reuse, and potential exposure in generated outputs.

Desired Design Dimensions

The study identifies five key dimensions for scalability and contextual fit in transparency mechanisms:

1. Modality: Both interactive (AI-driven FAQs, progressive disclosure) and static (visual badges, labels) elements, with the former supporting on-demand contextualization and the latter enabling rapid comparison.
2. Granularity: Low granularity summaries suffice pre-adoption; high post-adoption granularity (fine-grained, time-bounded permissions) is critical for context-sensitive control.
3. Explainability/Reflectiveness: Every opt-in/opt-out action must provide just-in-time explanations and pre-commit confirmation to prevent habituation and support informed reflection.
4. Findability: S&P options must be visible and directly accessible, avoiding hidden or buried menus.
5. Timing: Repeated, persistent, and just-in-time disclosures (e.g., every app launch, action-triggered reminders, always-visible banners) are essential for dynamic risk environments.

   Figure 2: Prototype interface synthesizing user-sketched features: pre-adoption audit badges, concise explanations, granular permissions, AI FAQ, post-adoption privacy modes, persistent reminders, and session/global control management.

   

   Figure 3: Traceability of original user sketch concepts to implemented prototype features—demonstrating alignment between articulated needs and functional transparency affordances.

Design and Policy Reflections

Emerging Challenges

The study highlights a critical mismatch between the evolving, multi-contextual privacy risks of GenAI and the static, coarse-grained S&P control paradigms enforced by current tools. This issue is compounded by frequent, silent policy modifications and poor notification strategies.

A "credibility gap" is identified: users lack trusted, independently-validated signals to distinguish actual S&P qualities from marketing-driven assurances, mirroring decades-old issues with "notice-and-choice" frameworks that have plagued traditional digital services and IoT.

Evaluating Existing Approaches

Analysis of leading S&P transparency implementations finds the landscape dominated by expert-facing, persistent, high-granularity static documentation (e.g., model cards, fact sheets), with little support for consumer legibility, interactive engagement, or independent comparison. Interactive modalities and repeated, action-coupled reminders remain underdeveloped.

Recommendations

• Empirical evaluation: Systematic measurement of proposed design dimensions on user understanding, behavior, and disclosure under varying risk contexts.
• Scalable defaults: Secure, expectation-aligned default S&P settings with optional granularity to minimize user burden and prevent false confidence.
• Policy intervention: Mandating standardized, independently-certified, consumer-facing S&P disclosures, leveraging lessons from IoT labeling programs, while safeguarding against transparency-washing and "dark patterns" that obfuscate risk.

Implications and Future Directions

This study establishes a taxonomy of user-informed design levers for S&P transparency in GenAI, emphasizing the need for adaptable, context-sensitive mechanisms that explicitly acknowledge dynamic risk profiles. It empirically demonstrates the gap between expert-designed documentation and the needs of general users, advocating for a paradigm shift toward interactive, explainable, and externally-audited transparency.

Future research should focus on quantifying the effect of these dimensions on longitudinal trust, disclosure practices, and the interplay with evolving regulatory environments. Practically, scaling independent S&P certification and integrating privacy-supportive interaction patterns into GenAI UX remains an open design and engineering challenge.

Conclusion

The analysis reveals that S&P transparency in consumer-facing GenAI fails to substantively influence adoption, yet S&P uncertainty directly constrains high-stakes use. Users operate in a climate of inaccessible, low-credibility S&P communication and are forced to rely on indirect proxies or abandon personalization benefits as a mitigation strategy. The paper provides actionable design guidelines and a robust empirical foundation for developing the next generation of user-empowering transparency practices in generative AI (2604.17270).