Visual Content Considered Private by People Who are Blind
Abigale Stangl, Kristina Shiroma, Bo Xie, Kenneth R. Fleischmann, Danna Gurari · 2020 · Proceedings of the 22nd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2020) · doi:10.1145/3373625.3417014
Summary
This paper presents the first empirical study investigating what types of visual content people who are blind consider private, in the context of their use of Visual Interpreters or Description Services (VIDS) — services like Seeing AI, Aira, Be My Eyes, and TapTapSee that describe visual surroundings from camera feeds. People who are blind share images and video through these services to accomplish everyday tasks (shopping, reading mail, cooking, navigation), but in doing so they may knowingly or unknowingly expose private visual content (PVC) to human agents or AI systems. The researchers conducted two-stage interviews with 18 participants (11 female, 7 male; ages 22-73; mean 42; 10 born blind, 8 acquired blindness; all totally blind) who used both human-powered VIDS (H-VIDS like Aira, Be My Eyes) and AI-powered VIDS (AI-VIDS like Seeing AI). The first stage used semi-structured questions to elicit what participants naturally consider private, their understanding of how VIDS work, and factors affecting their privacy. The second stage asked participants to rank their level of concern (1-5 scale) for 21 pre-identified PVC types across five conditions: shared publicly, shared with H-VIDS knowingly, H-VIDS unknowingly, AI-VIDS knowingly, and AI-VIDS unknowingly. This generated 1,890 short-answer responses that were coded alongside the interview data.
Key findings
Financial account information was the most concerning PVC type (average 4.6/5 for public sharing), followed by medical information (4.2), naked body images (4.1), and pregnancy test results (3.8). The ranking data revealed a critical nuance: privacy concerns shift significantly based on who provides the description and whether sharing is intentional. For financial information, concern dropped from 4.6 (public) to 2.8 (H-VIDS knowingly) to 3.4 (H-VIDS unknowingly) — participants trusted trained human agents due to professionalism, accountability, and company policies. AI-VIDS concerns for financial data were 2.5 (knowingly) and 3.1 (unknowingly). Participants identified distinct benefits and risks for each VIDS type: H-VIDS benefits included professionalism, trained agents, contractual accountability, and trust in human decency; H-VIDS risks included identity theft, social judgment ("there's certain things I may not want a human actually reading to me"), and unknown data handling. AI-VIDS benefits centered on anonymity — "no human eyes on data" and elimination of social judgment; AI-VIDS risks included unknown data retention ("What happens to the picture after it runs through the database?") and system vulnerabilities. Several participants directly identified that their blindness increases their privacy risk: "I recognize blind people have less privacy because we stand out in a crowd." Self-identified PVC types included financial accounts, medical information, identification/location information, paperwork, computer/online access data, images of people, and — newly identified beyond prior taxonomies — Social Security information and educational institution records. Eight privacy-adjacent values emerged: trust in human decency, anonymity, accountability, control/ownership, consent, acceptability, care/protection of others, and rights/justice.
Relevance
This paper provides essential, actionable guidance for developers of visual assistance technologies and AI-powered image description services. The PVC taxonomy — organized into financial, medical, people, location, identification, computer/online access, and affiliation clusters — can directly inform privacy policies, data handling protocols, and content-aware privacy controls in VIDS. The finding that privacy concerns vary dramatically depending on whether the agent is human or AI, and whether sharing is intentional or inadvertent, means one-size-fits-all privacy policies are insufficient. For accessibility practitioners, the paper raises a fundamental tension: people who are blind depend on sharing visual information to access the world, yet this dependency inherently creates privacy risks that sighted people do not face. The values framework (trust, anonymity, accountability, consent, control) provides a foundation for Value-Sensitive Design of assistive vision technologies. Participants' widespread confusion about how VIDS handle their data ("unknown data handling" was coded 117 times) highlights an urgent need for transparent, accessible privacy policies. Limitations include the small sample (N=18), all-US participants, and the hypothetical nature of the ranking task.
Tags: visual accessibility · blindness and low vision · privacy · computer vision · artificial intelligence · ethics · assistive technology · data protection
Standards referenced: GDPR