"Dump it, Destroy it, Send it to Data Heaven": Blind People's Expectations for Visual Privacy in Visual Assistance Technologies
Abigale Stangl, Emma Sadjo, Pardis Emami-Naeini, Yang Wang, Danna Gurari, Leah Findlater · 2023 · Proceedings of the 20th International Web for All Conference (W4A) · doi:10.1145/3587281.3587296
Summary
This paper investigates the visual privacy expectations of 16 totally blind individuals who use visual assistance technologies (VATs) such as Aira, Be My Eyes, Seeing AI, and Envision AI. VATs provide blind users access to visual information by connecting them to remote human assistants or AI systems that describe what their camera captures. However, approximately 10% of the roughly 40,000 photos shared through VATs contain private visual content including medical, financial, proprietary, and biometric information. The researchers conducted 1.5-hour semi-structured interviews guided by Rao's conceptual framework on privacy expectations, which incorporates desires, predictions, tolerances, and rights. Participants ranged from 19 to 72 years old, with nine identifying as female and seven as male. All used VATs weekly for everyday tasks like navigating, shopping, washing clothes, and reading mail. The study departs from prior work that focused on privacy concerns by instead examining expectations — what blind users believe VATs should do to preserve their visual privacy throughout the data life-cycle. The analysis was both inductive and deductive, using MAXQDA for qualitative coding, organized around parent codes of privacy values (control, trust, accountability/transparency) and data life-cycle stages (communication of practices, collection, storage, processing, security management).
Key findings
Three overarching user-centered expectations emerged. First, participants expected VATs to never collect images or videos containing private visual content — preferring on-device processing with immediate deletion over cloud storage. Participants expressed this forcefully: "Dump it. Destroy it. Send it to data heaven." Second, participants wanted interactive features providing real-time logs of who handles their visual data and why, along with non-disclosure agreements from software developers and third parties. Third, participants expected caution and ethical consideration during AI development, including blind co-developers in the process. Participants held distinctly different expectations for human remote-sighted assistants versus AI-based VATs: they wanted human assistants to proactively notify them about private content in their camera feed, while they expected AI systems to automatically recognize and obfuscate private content. Payment for service significantly increased trust expectations — paid services like Aira were held to higher confidentiality standards than volunteer services like Be My Eyes. When presented with AI-based obfuscation (automatically blurring or blacking out private content), participants set an average accuracy threshold of 88.75% before they would trust the feature. Most preferred blacking out over blurring as a security measure since "blurring can be undone." Participants wanted granular controls over which content types to obfuscate and the ability to toggle features on and off.
Relevance
This research is critical for developers of camera-based assistive technologies, which are proliferating rapidly with advances in AI vision models. The findings provide actionable design guidance: VATs should implement data minimization (collect only what is necessary), offer on-device processing options, provide transparent real-time data logs, and include blind people as co-developers in privacy feature design. The distinction between expectations for human versus AI assistants is particularly valuable — users have different mental models for how each should handle private content. For organizations deploying visual assistance services, the study underscores that privacy policies alone are insufficient; blind users want interactive, accessible mechanisms for understanding and controlling their data. The work also highlights an equity concern: tiered privacy models where paying users get better protection could further disadvantage blind people who are already systematically underemployed.
Tags: visual privacy · blind users · visual assistance technology · privacy expectations · data protection · AI ethics · camera-based assistive technology · obfuscation
Standards referenced: HIPAA