Authors: Yuanyuan Feng, Abhilasha Ravichander, Yaxing Yao, Shikun Zhang, Rex Chen, Shomir Wilson, and Norman Sadeh
Introduction
Data privacy is often described as a "sighted person’s game." Privacy policies are notoriously dense, filled with legal jargon, and frequently buried within interfaces that are not optimized for screen readers. For the approximately 2.4% of US adults with visual disabilities, these "walls of text" represent more than just a nuisance; they are a significant barrier to informed consent and digital autonomy.
While substantial research exists regarding how sighted users navigate privacy, there is a critical knowledge gap concerning the specific information-seeking behaviors and risk perceptions of the blind and low-vision (BLV) community. As researchers develop Large Language Model (LLM) based tools to answer privacy questions—such as "Does this app sell my location data?"—it is vital to ensure these tools are designed with the specific accessibility needs and trust models of BLV users in mind.
Methodology
To address this gap, researchers conducted a qualitative interview study with 21 BLV participants across the United States. To ensure the study was culturally responsive and methodologically sound, the team collaborated with two blind consultants throughout the process.
The study employed several key techniques:
- Critical Incident Technique: Participants recalled specific, recent moments where they felt surprised or uncomfortable regarding their data privacy.
- Hypothetical "Digital Assistant": To avoid priming bias, participants were asked to imagine an ideal "digital expert" before the concept of an NLP-based Q&A tool was introduced.
- Thematic Analysis: Over 22 hours of interview audio were synthesized using a mix of inductive and deductive coding to identify high-level meta-themes.
Unique Privacy Risks for BLV Users
The study revealed that BLV users face a unique set of privacy challenges often overlooked by mainstream privacy research:
The Visual Interpreter Paradox
BLV users frequently rely on agent-based services (such as Aira or Be My Eyes) where a sighted human assistant views the user's environment through a camera. This creates a "physical-digital" risk: a human agent might inadvertently see sensitive items like credit cards, social security numbers, or private medical information.
Voice Assistant Exposure
Participants noted that voice-based tools (e.g., Siri, Alexa) often read sensitive information aloud. This creates privacy risks in public spaces that sighted users—who can read a screen silently—do not encounter.
Privacy Resignation
Many participants expressed a sense of "privacy resignation," feeling that they must trade their personal data for the utility of assistive tools. As one participant noted, "Somebody has to know everything to help me out with something."
Key Results and Design Requirements
The study found a high demand for Privacy Q&A tools, with 19 out of 21 participants expressing specific expectations for such an assistant. Rather than a standard chatbot, users desired a "snapshot" of an app's data practices.
To describe the frequency of these sentiments, the researchers mapped participant counts to specific linguistic quantifiers, as shown below:
Key findings regarding the design of these assistants include:
- Trust in Automation: Interestingly, participants generally viewed AI-based visual aids (like Seeing AI) as lower risk than human-agent services because they lack human involvement, showing a high level of trust in automated algorithms over human intermediaries.
- Multi-modality: 15 participants (the "majority") explicitly requested that a privacy assistant support both text and voice. Some preferred text-only output to be read by their own screen reader for better control and privacy in public settings.
- Neutrality and Source: 16 participants stated they would trust non-profit organizations or research universities to develop these tools over "Big Tech" companies, fearing that first-party developers would have a vested interest in obscuring data practices.
Conclusion and Future Work
The study highlights that for BLV users, functionality often takes precedence over form, with the top requested feature being "easy access to trusted information." Future work should focus on developing verification mechanisms to "double-check" company claims against actual data flows and integrating these privacy tools directly into existing workflows, such as screen readers or browser extensions. Furthermore, refining NLP for accessibility is crucial to avoid the "fatigue" of long-winded voice responses when summarizing legal text.
Summary updated