Privacy Concerns for Visual Assistance Technologies
Abigale Stangl, Kristina Shiroma, Nathan Davis, Bo Xie, Kenneth R. Fleischmann, Leah Findlater, Danna Gurari · 2022 · ACM Transactions on Accessible Computing · doi:10.1145/3517384
Summary
This comprehensive study examines privacy concerns of blind users when sharing images and videos with visual assistance technologies (VATs) like Aira, Be My Eyes, and Seeing AI. The research addresses a fundamental tension: blind people must share visual data to receive assistance, but cannot independently verify what private content their images contain before sharing. Through two complementary studies, the researchers interviewed 18 blind participants about their privacy perceptions, had them rate concern levels for 21 types of private visual content (PVC) across different sharing conditions, and analyzed the privacy policies of 13 major VAT companies. The study distinguishes between human-powered VATs (using remote sighted assistants or volunteers) and AI-powered VATs (using automated image recognition), finding that privacy concerns differ substantially between these service types. Participants valued the professionalism and trained agents of human-powered services but worried about identity theft and social judgment; they appreciated the anonymity of AI systems but were concerned about unknown data handling practices.
Key findings
Financial account information was rated most concerning (4.6/5 if made public), followed by medical information, naked body images, and location data. "Unknown Data Handling" emerged as the dominant risk factor, appearing in 12.8% of 901 coded risk responses—and was disproportionately associated with AI-powered VATs. Participants with prior visual experience (acquired blindness) showed consistently higher privacy concerns than those born blind, suggesting visual memory affects privacy perception. The privacy policy analysis revealed alarming gaps: 7 of 13 companies do not mention collecting images or video; none disclose retention periods for visual data; only 2 mention using visual data to train AI; only 2 disclose selling data to third parties; and none allow users to opt out of visual data collection specifically. Sharing data "unknowingly" consistently generated higher concern than "knowingly" sharing across all PVC types and VAT types.
Relevance
This research has immediate implications for VAT developers and the broader camera-based technology industry. The findings expose a critical misalignment between what users want to know about their visual data and what companies disclose. The study recommends designing VATs with "non-visual consent" mechanisms, providing accessible tools to help users assess whether their images contain private content before sharing, and adopting a "Privacy by Design" approach. For practitioners, the 21-type PVC taxonomy and privacy concern ratings provide concrete guidance on which content types require extra safeguards. The finding that prior visual experience affects privacy perceptions suggests VAT companies should offer personalized privacy features. This work also extends to other camera-based services—smart home devices, autonomous vehicles, life-logging cameras—where users similarly share visual data without full awareness of its contents.
Tags: visual impairment · privacy · visual assistance technology · artificial intelligence · remote sighted assistance · image description · privacy policy · data protection
Standards referenced: GDPR · HIPAA · Fair Information Practice Principles