Exploring Collaboration to Center the Deaf Community in Sign Language AI
Rie Kamikubo, Abraham Glasser, Alex X Lu, Hal Daumé III, Hernisa Kacorri, Danielle Bragg · 2025 · ASSETS 2025: 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746390
Summary
This paper investigates the critical disconnect between machine learning (ML) practitioners and Deaf American Sign Language (ASL) signers in the development of sign language AI technologies. Despite sign language processing's promise for societal inclusivity, the field frequently excludes meaningful Deaf community participation, raising ethical concerns about whether resulting technologies actually serve Deaf people's needs. The research comprises two interrelated studies. First, parallel surveys of 35 ML practitioners and 10 ASL signers identify misaligned priorities, misconceptions, and divergent expectations between the two groups. Second, eight paired co-design sessions — each pairing one ASL expert with one ML expert — use a guided inquiry process called "starbursting" to generate questions (rather than answers) that could structure future collaborative sign language AI projects. The surveys revealed that while both groups share motivation for real-world impact, ML practitioners are primarily driven by dataset availability and technical advancement, while ASL signers prioritize societal impact, Deaf advocate involvement, and ML practitioners' ASL proficiency. A Deaf awareness assessment found significant knowledge gaps among ML practitioners: only 31% answered all terminology questions correctly, 26% answered people-related questions correctly (with 60% incorrectly believing most deaf people who sign are bilingual), and misconceptions persisted even among those with sign language processing experience. The co-design sessions generated 150 guiding questions across 10 themes, revealing that ASL experts emphasize project motivation, methodology rationale, stakeholder analysis, and ethical implications, while ML experts focus on objectives, use cases, scope constraints, and deliverables.
Key findings
The research surfaces critical friction points that reflect deeper systemic and epistemic barriers to collaboration. ASL signers identified communication access as their biggest barrier to collaboration with hearing people — limited interpretation services, inconsistent interpretation quality, and reliance on text-based communication that loses nuance. ML practitioners' Deaf awareness gaps are concerning: 51% considered "hearing impaired" appropriate terminology (it is rejected by the Deaf community), 74% incorrectly identified ASL as universal, and many with sign language processing experience still rated their ASL proficiency at Level 0 (unable to function in the language). There is a stark asymmetry in topic priorities: 90% of ASL signers prioritize signed-to-spoken real-time translation, while ML practitioners spread interest across many topics with sign language understanding being most popular (57%) — a topic ASL signers ranked lowest (10%). On timelines, 30% of ASL signers believe single sign recognition already works or will within 2 years, while ML practitioners estimate 3-10 years. In the co-design sessions, ASL experts consistently foregrounded community accountability, ethical risks, and long-term impact — with one participant powerfully warning "This could be a form of genocide" regarding AI's potential to flatten linguistic diversity. ML experts focused on technical feasibility, scope constraints, and deliverables. Both groups shared concerns about data collection ethics and representativeness, but framed them differently — ASL experts asked "Who should we recruit?" while ML experts asked "What are our resources?"
Relevance
This paper provides essential guidance for anyone developing AI technologies intended to serve the Deaf community or other marginalized language communities. The findings challenge the common assumption that including disabled people in research is sufficient — meaningful collaboration requires addressing power imbalances, epistemic differences, and communication access barriers throughout the entire project lifecycle. The guiding questions generated in the co-design sessions (publicly available) serve as practical resources for structuring interdisciplinary teams and surfacing hidden assumptions early. For accessibility practitioners more broadly, this research illustrates how technology development for disability communities can perpetuate harm when led by people without lived experience — the "disability dongle" problem of well-intentioned but misaligned solutions. The Deaf awareness findings are a wake-up call: even researchers with sign language processing experience hold fundamental misconceptions about Deaf culture, language, and community preferences. The paper advocates for shifting from "inclusion" to "agency and control," where Deaf people lead rather than merely participate in shaping technologies that affect their lives and language.
Tags: Deaf community · sign language · American Sign Language · machine learning · participatory AI · co-design · interdisciplinary collaboration · AI ethics · community-centered design · sign language recognition