Trying to Piece It Together: Exploring Accessible Error Detection in Emerging Privacy Techniques With Blind People
Rahaf Alharbi, Angela D Cheong, Jaylin Herskovitz, Robin N. Brewer, Sarita Schoenebeck · 2025 · ASSETS 2025: 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746376
Summary
This paper investigates how blind people might detect errors in AI-enabled privacy techniques (obfuscation) used within visual assistance technologies (VAT). Blind people routinely use VAT apps like Seeing AI, Be My Eyes, Aira, and TapTapSee to access visual information, but these tools can inadvertently capture and transmit sensitive private content. Researchers have developed obfuscation techniques that automatically detect and remove private content by applying filters such as blurring or masking. However, these AI systems are imperfect and may misrecognize objects, creating errors that are particularly difficult for blind users to identify. The study explores whether "assessment descriptors" — brief visual attributes of objects such as color, size, dimensions, and distance — could help blind people find obfuscation errors. The researchers conducted a two-part qualitative study: first, interviews with 26 blind participants to introduce obfuscation concepts (focus mode, which spotlights one object while hiding others, and background mode, which hides one object while preserving everything else), and second, seven focus groups with 16 blind participants using fictional audio probes to ground discussions about assessment descriptors. The study was conducted in collaboration with the National Federation of the Blind for participant recruitment, and used reflexive thematic analysis to identify patterns across the data.
Key findings
The majority of participants did not find the presented assessment descriptors (color, dimensions, distance) useful for detecting obfuscation errors. Participants identified these descriptors as vague, sighted-centric, and misaligned with how blind people actually identify objects and navigate spaces. Color was the least desirable descriptor, particularly for people born blind who have no color reference. Dimensions without context were seen as meaningless since different objects can share identical measurements. Distance was relative and camera-dependent, making it unreliable. Instead, participants preferred assessment descriptors that name objects directly, include unique identifying features like text or logos, and describe multiple objects in the surrounding environment to enable cross-referencing with personal knowledge of the space. Participants also suggested multimodal alternatives including haptic feedback and audio tones rather than purely verbal descriptors. Beyond assessment descriptors, participants called for greater system-level transparency about how obfuscation works, what data is collected and processed, and where failures might occur. They emphasized the need for co-created training materials developed with blind communities, and preferred human support over AI-powered chatbots for error resolution. Participants strongly rejected automated nudges to obfuscate content as potentially coercive and autonomy-reducing.
Relevance
This research has significant implications for anyone designing AI-enabled privacy or verification tools for blind users. The central finding — that commonly used visual descriptors reflect sighted assumptions and fail to serve blind users — challenges designers to rethink how they communicate AI output non-visually. The concept of sighted bias in assessment descriptors connects to broader accessibility concerns about centering sighted norms in assistive technology design. For practitioners, the study highlights that accessible error detection requires more than adding verbal descriptions; it demands fundamentally different information architectures built around how blind people actually make sense of their environments. The emphasis on transparency, training materials co-created with blind communities, and human support pathways provides a practical roadmap for organizations developing privacy features in VAT. The findings also underscore that trust in AI privacy tools cannot be achieved through technical features alone but requires organizational accountability and community engagement.
Tags: visual assistance technologies · privacy · obfuscation · AI error detection · assessment descriptors · blind users · sighted bias · transparency · trust · qualitative research · focus groups