"I Don't Trust it, but I Use it": Navigating Trust, Privacy, and Identity in Disabled People's Use of Generative AI
Jazette Johnson, Aaleyah Lewis, Jennifer Mankoff, Olivia Banner · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790652
Summary
This CHI 2026 paper reports a qualitative focus-group study of how disabled people navigate generative AI (GenAI) tools in everyday life, with particular attention to how trust, privacy, and intersecting identities (race, gender, language, sexuality, disability) shape their use. The authors conducted seven semi-structured Zoom focus groups with 20 participants, intentionally recruited cross-disability (blind/low vision, d/Deaf or hard-of-hearing, chronically ill, neurodivergent, AAC users, motor, mental health, intellectual/developmental, and multiply disabled), ranging from 18 to 85 years old and explicitly including multiply marginalised identities (Black, Indigenous, Hispanic, Asian, queer, trans, non-binary, multilingual). Two sessions were tailored for AAC and text-based communicators. The study addresses three research questions: how disabled people use GenAI for accessibility in daily life; what shapes their trust or mistrust; and what identity-based benefits and harms they encounter. Data were analysed via reflexive thematic analysis combined with descriptive statistics. The authors are explicit about their positionality — three actively use GenAI across accessibility, professional, and personal contexts; one actively resists it — and frame the work in crip technoscience, disability justice, and Crenshaw-style intersectionality. A theoretical contribution is extending Olsen et al.'s 'accessibility tax' concept to GenAI, distinguishing financial, disclosure, labour, prompt-engineering, and verification taxes that fall disproportionately on multiply marginalised disabled users. The authors also introduce a typology of identity outcomes in GenAI interactions: identity preservation, identity validation, supportive code-switching, identity erasure, and identity distortion/stereotyping. They frame participants' everyday GenAI use as a constant 'balancing act between access and risk' and argue that HCI research must rethink what 'access' means when the tool providing access is itself unreliable, biased, and privacy-invasive.
Key findings
Participants used GenAI (ChatGPT, Gemini, Claude, BeMyAI, SeeingAI, DALL-E, CanvaAI, ElevenLabs, PrivateGPT, and others) across four main domains: creative outlets, professional/academic support, information seeking, and accessibility assistance. Accessibility assistance was present in 13 of 20 participants' reported uses. Only three participants (Bond, Sophia, Jasmine) actively resisted GenAI. The central empirical finding is that mistrust coexists with continued use. No participants unconditionally trusted GenAI; strategies to manage mistrust included limiting use to 'low-stakes' tasks (Michelle), verifying via conversational drilling (Maddy), cross-checking against Google or human experts (Tonya), selecting platforms based on parent company reputation (avoiding MetaAI), and using locally trained models (Ethan). Concerns spanned reliability (Thomas's fear of misreading a $20 bill as $200), environmental harm (Sophia, Lex), corporate ethics including LGBTQ-unfriendly company leadership (Fran) and racist training data (Lex), and privacy — particularly around long-term training use of sensitive disability data (Ann, Bond). Identity findings distinguished preservation, validation, and supportive code-switching from erasure and stereotyping. Positive cases: Tonya felt GenAI 'honoured her identity' as a disabled person; Alexander's neurodivergent writing style (abbreviations, sentence fragments) produced consistently generalised output without judgement; Amy found ChatGPT responded in Pidgin English. Negative cases: Sara's LGBTQ+ 'non-binary' flipped to 'the animal'; Michelle's Black youth Christmas graphic came back with 'a lot of slang terms that would be offensive'; Jasmine's ASL linguistic identity was unsupported; unspecified race defaulted to White/male/English. Accessibility taxes were concrete: ChatGPT Pro pricing excluded Amy; Morgan described VOCR token costs creating 'financial craziness'; Maddy used pseudonyms and redacted documents; Lex and Nicole spent substantial time prompt-engineering emails to avoid professor pushback. Five of twenty had not completed the optional demographic survey despite compensation — a tell for study burden.
Relevance
This is essential reading for anyone designing, governing, or evaluating generative AI for disabled users — and an important corrective to single-axis disability studies of ChatGPT. The intersectional frame means the findings apply to AI fairness and trust work more broadly: the paper shows concretely how 'neutral' defaults reproduce Whiteness, English, and cis-male norms, and why verification labour compounds for multiply marginalised users. For practitioners: the three design implications — designing for mistrust (transparency about training data, clear data-use controls, source routing), accessibility in and through identity management (personalisation features for code-switching preferences, multilingual and multidialectal output, settings persistence), and reducing the accessibility tax (going beyond WCAG compliance to consider financial, disclosure, verification, and prompt-engineering costs) — give product teams a concrete checklist. The authors' critique of current accessibility discourse on GenAI as overly focused on interface compliance is pointed: an accessible query box does not make a hallucinated medical answer accessible. Limitations the authors name: Western/US-centric recruitment, over-representation of White, English-speaking, educated participants, limited coverage of religion/class/bodytype identity axes, and under-representation of active resisters (only three). Scam screener responses were a recurring recruitment challenge — a practical note for future disability-identity qualitative work. Read this alongside intersectional AI fairness literature (Crenshaw, D'Ignazio and Klein, Harrington et al.) when scoping inclusive GenAI product work.
Tags: generative AI · accessibility · trust · mistrust · privacy · intersectionality · disability identity · focus groups · qualitative research · accessibility tax · cross-disability · code-switching · race and disability · language and disability