Privacy Considerations of the Visually Impaired with Camera Based Assistive Technologies: Misrepresentation, Impropriety, and Fairness
Taslima Akter, Tousif Ahmed, Apu Kapadia, Swami Manohar Swaminathan · 2020 · Proceedings of the 22nd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2020) · doi:10.1145/3373625.3417003
Summary
This paper investigates the privacy concerns of both visually impaired people (PVIs) and sighted bystanders regarding camera-based assistive technologies like smart glasses (Orcam, Aira, eSight) that can identify people and provide demographic and behavioral information about those nearby. Through two online surveys — one with 128 visually impaired participants (65.6% totally blind, 34.4% partially sighted; recruited via NFB and ACB) and one with 136 sighted bystanders (Amazon Mechanical Turk) — the researchers examined comfort levels and usefulness perceptions across 11 types of "visually available" information (activity, distance, attire, whether someone is alone or available for conversation, emotion, gaze, gender, age, ethnicity, height, and weight) under two experimental field-of-view conditions: "sighted" FoV (front-facing camera only) and "extended" FoV (all directions including behind). The study addresses three research questions: what information PVIs consider fair and useful from each FoV, what information PVIs consider proper or improper to receive about bystanders, and what concerns PVIs and bystanders share about these technologies.
Key findings
Three major themes emerged. First, regarding fairness: PVIs considered it equally fair and useful to receive information from both sighted and extended FoVs, reasoning that sighted people can easily obtain the same information "with a simple turn of the head." PVIs showed no significant difference in comfort between the two FoV conditions. However, bystanders were consistently less comfortable sharing information in the extended FoV, especially behavioral information (looking, activity). Second, regarding impropriety: PVIs were much less interested in visual attributes (age, height, weight, gender, ethnicity — usefulness mean 2.67) than behavioral information (expression, gaze, activity, available for conversation — usefulness mean 3.78) and distance (3.64). Critically, PVIs reported discomfort receiving certain visually available information about bystanders because they considered it "impolite" or "improper" — particularly weight ("I feel it is an invasion of privacy to know approximately how much someone weighs"), gender ("It's already considered negative to assume a person's gender"), and ethnicity. Some PVIs explicitly preferred not to receive such information to avoid developing implicit biases: "I might fall into traps of judging people by their appearance." One participant characterized blindness as a "blessing" in this regard. Bystanders shared these impropriety concerns, worrying that PVIs might become judgmental after receiving appearance information. Third, regarding misrepresentation: both groups (N=24 PVIs, multiple bystanders) expressed deep concerns about AI fallibility — incorrect information could lead to embarrassment, misgendering ("I'm trans... the glasses would get it wrong at some point, which would give people more reasons to misgender me"), or unsafe physical interactions (moving too close to someone misidentified as smaller). One participant with visual impairment reported AI had estimated their age as anywhere from 16 to 45. Low-vision participants were slightly less comfortable receiving bystander information than totally blind participants (p=0.01).
Relevance
This paper makes a crucial contribution by showing that the privacy debate around camera-based assistive technologies is not simply PVI needs versus bystander discomfort — both groups share significant concerns about misrepresentation, impropriety, and the quality of AI-generated information. For developers of assistive vision technologies, the findings challenge the assumption that providing "equal" visual information to PVIs is always desirable. PVIs themselves want to filter out information they consider socially inappropriate (weight, ethnicity, gender) and prioritize dynamic behavioral cues (expressions, gaze, activity, availability for conversation) that are more useful for social interaction. The three design implications are: (1) address fairness by supporting extended FoV but finding ways to increase bystander comfort (e.g., visible indicators of what information is being gathered); (2) address propriety by allowing PVIs to specify what information they consider inappropriate, similar to privacy policies; (3) address misrepresentation by conveying AI confidence levels and using "negative framing" for uncertain inferences. Limitations include the US-only sample, online survey format (no prototype interaction), gender imbalance, and the social gathering scenario that may not generalize to workplaces or other contexts.
Tags: visual accessibility · blindness and low vision · privacy · AI bias · AI fairness · computer vision · wearable technology · ethics · social accessibility