"Before, I Asked My Mom, Now I Ask ChatGPT": Visual Privacy Management with Generative AI for Blind and Low-Vision People
Tanusree Sharma, Yu-Yun Tseng, Lotus Zhang, Ayae Ide, Kelly Avery Mack, Leah Findlater, Danna Gurari, Yang Wang · 2025 · Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2025) · doi:10.1145/3663547.3746335
Summary
This paper presents the first in-depth empirical study of how blind and low vision (BLV) individuals use generative AI (GenAI) tools to manage visual privacy across diverse everyday contexts. The researchers conducted semi-structured interviews with 21 BLV participants in the United States who actively use GenAI tools such as ChatGPT, Be My AI, Envision AI, Seeing AI, and Gemini. The study explored two research questions: how BLV people currently use GenAI for visual privacy management, and what design opportunities they envision for future tools. Using a scenario-driven approach, interviews covered six privacy contexts: self-appearance and impression management, indoor spatial privacy, sharing visual content on social media, visual content privacy when sharing with employers, visual privacy as BLV professionals handling others' content, and outdoor spatial privacy. The findings reveal that GenAI has become a significant replacement for human assistance in privacy-sensitive tasks—participants use these tools to read pregnancy tests, interpret mammograms, check bank statements, verify prescriptions, sort mail, and scan surroundings before taking photos. The study surfaces nuanced privacy judgments where participants balance independence, convenience, emotional agency, and trust when choosing between GenAI tools, human-mediated visual interpretation services like Aira, and family or friends. Privacy management extends beyond data protection to encompass impression management (controlling how others perceive appearance and environment) and accountability in handling others' private information in professional settings.
Key findings
Participants demonstrated sophisticated privacy reasoning, employing strategies such as compartmentalisation (splitting sensitive information across multiple uploads to prevent full exposure), routine deletion of content after processing, and context-dependent tool selection based on perceived sensitivity. For medical content like pregnancy tests and mammograms, many participants preferred GenAI over human assistance specifically for emotional privacy—the ability to process sensitive results independently without exposing vulnerability. For financial documents, trust levels varied by tool and context: some participants used GenAI freely for bills and receipts but avoided it for tax documents and credit cards, while others adopted GenAI as a last resort when human assistance was unavailable. Several participants raised concerns about AI policy decisions around race and people descriptions, noting that overly cautious content filters that avoid identifying race, gender, or skin tone made descriptions less useful for blind users who rely on such details. Participants also expressed concerns about AI perpetuating disability myths and stereotypes. For outdoor spatial privacy, participants described feeling socially awkward using GenAI tools in public spaces, with bystanders thinking they were being photographed. BLV professionals working in education, social services, and finance described a distinct tension between GenAI's accessibility benefits and institutional data protection requirements, often self-governing their use without formal organisational guidance.
Relevance
This research is highly relevant for GenAI developers, accessibility practitioners, and policymakers. The paper proposes five actionable design recommendations: (1) on-device processing or local encryption using modular sandbox architectures per data type to eliminate cloud exposure for sensitive content; (2) a federated compliance-aware secure toolkit for BLV professionals that integrates with enterprise platforms and enforces role-based access controls aligned with FERPA and HIPAA; (3) personalised privacy-aware appearance feedback systems using fine-tuned visual classifiers with customisable sensitivity profiles; (4) visual disambiguation for shared spaces, enabling users to identify and tag their belongings versus others' items using personalised object recognition; and (5) zero-retention guarantees with post-task accountability prompts. The study expands the notion of visual privacy beyond data protection to include emotional vulnerability, impression management, and institutional accountability—dimensions particularly critical for BLV users who cannot independently verify what visual information they are sharing. The finding that participants strategically avoid processing certain images through GenAI to protect disability benefits highlights how surveillance concerns intersect with disability rights.
Tags: blind and low vision · visual privacy · generative AI · privacy management · impression management · visual interpretation services · AI ethics
Standards referenced: ADA · HIPAA · FERPA · Assistive Technology Act of 2004