Exploring AI Opportunities in Deaf Education: Understanding Design Needs Through Teacher and Parent Perspectives in Bangladesh
Md. Ataur Rahman Bhuiyan, Nadim Mahmud Dipu, Tanvir Rahman, Oindri Aurunima Sarker, Shidhartha Chakrabarty Turzo, Jannatun Noor · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791810
Summary
This CHI 2026 qualitative study investigates how AI-powered educational tools should be designed for Deaf learners in Bangladesh — a low-resource context where Bangla Sign Language (BdSL) is still evolving, datasets are small, and infrastructure (internet, electricity, devices) is often unreliable. The authors conducted a two-phase study: Phase 1 (May–August 2025) involved 13 teachers and 8 parents recruited from government, private, and NGO Deaf schools in Dhaka and Chittagong, through two focus groups and individual interviews; Phase 2 (November 2025) centred 5 Deaf students aged 12–17 in a focus group mediated by a subject teacher acting as BdSL interpreter. Four research questions address classroom communication challenges, gaps in existing practices, stakeholder expectations for AI tools, and culturally informed design principles. Interview protocols were iteratively validated via internal peer debriefing, expert methodological review, and community/ethical vetting by a local Deaf-serving NGO; all themes were mapped back to literature gaps (deaf pedagogy and access, sign-language translation fidelity, inclusive HCI and design justice). Reflexive thematic analysis was conducted separately for teachers, parents, and students to preserve perspective-specific integrity, then compared across groups. The paper's central theoretical contribution is the concept of 'fragile learning continuity' — a model in which accessibility for Deaf learners depends not on any single feature (visibility, vocabulary, connectivity) but on the ongoing alignment across visual, linguistic, technological, and affective-emotional dimensions. The authors also reframe AI's role from autonomous translator to 'access-stabilising support' operating under teacher validation and community governance, and reposition emotional safety as a structural accessibility requirement rather than a peripheral social factor.
Key findings
Four interacting themes emerged. (1) Visual and Linguistic Access: 11 of 13 teachers described recurring sightline disruptions in crowded, narrow classrooms; 4 of 5 students independently reported missing lesson content when a teacher turned to the board or a peer blocked their view — small movements can cause 'half the class to lose the meaning in a second'. 9 teachers reported ongoing BdSL vocabulary gaps requiring fingerspelling and improvised gestures, especially for scientific and technical terms. (2) Cognitive Load, Attention, and Emotional Safety: 4 of 5 students described the strain of monitoring teacher's hands, face, board, and peers simultaneously; 10 teachers confirmed constant adjustment. Crucially, 4 of 5 students admitted hesitating to ask for clarification because confusion is publicly visible in signed classrooms — surfacing emotional exposure as a structural access factor. (3) Home–School Communication and Technology Constraints: 6 of 8 parents reported their child entered formal schooling without prior BdSL exposure, and 7 struggled to interpret written materials. 9 teachers and 6 parents reported unreliable internet or shared-device access; brief video freezes ('when it freezes, the sign loses its meaning') disrupt meaning. (4) Trust, Co-Design, and Governance: 11 teachers demanded the ability to preview and edit AI-generated signs before classroom use; 5 parents worried about foreign sign variants (one mother described her son feeling 'crushed' when told signs he learned online were wrong); all 5 students preferred corrective feedback that preserved dignity ('if it tells me I am wrong without explaining, it feels like failing'). (5) AI Opportunities: 11 of 13 teachers, 6 of 8 parents, and all 5 students supported AI as a continuity-stabilising assistant rather than a replacement instructor. Concrete design directions include slow and expressive signing with adjustable pace and replay; regionally validated BdSL via moderation queues, version control, and teacher dashboards; offline-first design with on-device inference, predicted-lesson pre-caching, visual uncertainty indicators (not audio warnings), graceful degradation to key-frame sequences when bandwidth drops, and sync-on-reconnect; culturally grounded visuals reflecting local material culture (mango trees, rickshaws) rather than unfamiliar Western contexts; and co-design as ongoing governance with moderation queues, audit logs, and community proposal mechanisms.
Relevance
For HCI researchers and AI-for-accessibility teams, this paper reframes a dominant assumption in sign-language technology work: that the technical task is translation accuracy on a standardised corpus. In low-resource sign ecologies like BdSL, where vocabulary is evolving and authority rests with teachers and local communities, an AI system that confidently outputs a 'correct' sign can actively cause harm by imposing Global-North or foreign variants, undermining teacher credibility, or producing public misunderstanding that students are afraid to correct. The fragile-learning-continuity model gives product teams a diagnostic vocabulary for failure modes that don't show up in benchmark metrics: sightline disruption, sign-authority mismatch, connectivity-triggered meaning loss, and emotional-exposure-driven silence. The design principles (AI as access-stabilising support, culturally grounded visuals, regionally validated BdSL, slow/expressive/adjustable signing, offline-first low-burden design, co-design as governance) map cleanly onto engineering backlog items and governance process, not just philosophical commitments. Read alongside the ORBIT-India paper (India et al., 2026) for a coherent emerging body of Global-South disability-first AI work that challenges 'bigger and more accurate' as the default AI frame. Limitations the authors name: urban/peri-urban sample (rural Deaf experiences under-represented), a conceptual rather than prototype-tested study, and multilingual interview analysis mediated through Bangla. Essential reading for anyone scoping sign-language AI, Global-South educational technology, or participatory design in Deaf communities.
Tags: deaf education · sign language · Bangla Sign Language · BdSL · artificial intelligence · large language models · accessibility · co-design · Global South · educational technology · low-resource setting · Bangladesh · qualitative research · disability justice
Standards referenced: Universal Design for Learning