Shared Privacy Concerns of the Visually Impaired and Sighted Bystanders with Camera-Based Assistive Technologies
Taslima Akter, Tousif Ahmed, Apu Kapadia, Manohar Swaminathan · 2022 · ACM Transactions on Accessible Computing · doi:10.1145/3506857
Summary
This paper investigates the privacy and ethical concerns surrounding camera-based assistive technologies (like smart glasses) from two perspectives: people with visual impairments (PVIs) as technology wearers and sighted people as bystanders who might be observed. The researchers conducted two online surveys—128 visually impaired participants and 136 sighted bystanders—examining comfort levels and perceived usefulness of receiving or sharing 11 types of "visually available" information about bystanders: age, height, weight, appearance, gender, ethnicity, activity, distance, attire, gaze, facial expression, and availability for conversation. The study tested two field-of-view conditions: "Front" (forward-facing camera only) and "Front-Back" (360-degree view including behind the wearer). Participants used common assistive technologies including Seeing AI (75%), BeMyEyes (68.7%), and TapTapSee (66.4%). The methodology carefully avoided social desirability bias through neutral question framing and included extensive bot detection for the MTurk bystander survey. The research addresses three key questions: what information PVIs find useful from different fields of view, what information both groups consider proper or improper, and what concerns are shared between PVIs and bystanders regarding these technologies.
Key findings
The overarching finding is a shared ethical concern between PVIs and bystanders about AI fallibility—specifically, that bystanders can be algorithmically misrepresented by assistive devices. Both groups worried about embarrassment from inaccurate information, with particular concern about AI inferring subjective social constructs like gender, emotion, and ethnicity. PVIs found behavioral information (gaze, activity, expression, availability for conversation) significantly more useful than visual attributes (weight, ethnicity, age), with a strong positive correlation (r=0.82) between perceived usefulness and comfort level. However, PVIs expressed "propriety" concerns about certain information—considering it improper or impolite to access someone's weight or ethnicity even if technically available, to avoid becoming judgmental or biased. Bystanders showed the opposite pattern: they were more comfortable sharing visual attributes than behavioral information, considering the latter a greater invasion of privacy. Both groups were less comfortable with the Front-Back (360-degree) field of view, though PVIs considered it "fair" since sighted people can turn their heads to access the same information. Critically, transgender participants and others from marginalized groups expressed heightened concerns about AI misgendering and mischaracterizing them, noting that automatic gender recognition systems reinforce binary assumptions and can be more harmful than human mischaracterizations.
Relevance
This research has significant implications for the design of AI-powered assistive technologies. The findings challenge the assumption that providing PVIs with "equal access" to visual information is inherently positive—both PVIs and bystanders share concerns about what information is appropriate to access and how AI errors could harm social interactions. For practitioners developing camera-based assistive tools, the paper offers six design recommendations: (1) make information needs context-dependent rather than universal; (2) set standards for AI-generated descriptions that use objective rather than subjective characterizations; (3) increase social acceptability through visible indicators that the device is assistive; (4) implement propriety policies allowing users to filter inappropriate information types; (5) convey confidence levels so users understand when AI inferences are uncertain; and (6) enforce algorithmic accountability and transparency. The work highlights tensions between accessibility goals and privacy/ethics concerns that cannot be resolved purely technically. Organizations developing assistive AI must involve people with disabilities in design processes, represent marginalized groups in training data, and acknowledge that providing "equal access to visual information" may perpetuate rather than challenge problematic social categorizations.
Tags: privacy · visual impairment · camera-based assistive technology · AI ethics · algorithmic bias · smart glasses · augmented reality · bystander privacy · facial recognition