Helping or Homogenizing? GenAI as a Design Partner to Pre-Service SLPs for Just-in-Time Programming of AAC
Cynthia Zastudil, Christine Holyfield, Christine Kapp, Kate Hamilton, Kriti Baru, Liam Newsam, June A. Smith, Stephen MacNeil · 2025 · ASSETS 2025: 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746384
Summary
This paper investigates the integration of generative AI into augmentative and alternative communication (AAC) devices, specifically examining how AI-suggested hotspots affect the creation of visual scene displays (VSDs). VSDs are image-based communication tools that embed interactive hotspots — tappable regions linked to words or phrases spoken aloud — within photographs of real-world scenes. They are particularly effective for minimally verbal autistic children and other emergent communicators because they ground language in familiar, meaningful contexts rather than using abstract symbol grids. However, VSDs are difficult to configure, requiring expertise in both AAC intervention and the specific communication needs of each user, which limits their adoption outside clinical settings. The researchers developed a prototype that uses GPT-4o to automatically analyze uploaded images and suggest contextually relevant hotspots, which users can then accept, edit, delete, or supplement with their own additions. They conducted a within-subjects study with 16 pre-service speech-language pathologists (SLPs) — students training to become clinically certified — comparing their prototype against Tobii Dynavox's Snap Scene, a widely used commercial VSD application. Participants configured VSDs for two contexts (playing and retelling a past activity) using both applications, guided by case studies describing hypothetical autistic children with specific communication goals and developmental levels. The study measured creation time, user confidence, VSD quality against established best practices, and the degree to which AI suggestions were modified or accepted as-is.
Key findings
The results revealed a complex tradeoff between efficiency and quality. On the positive side, participants configured VSDs 17% faster with the AI prototype (median 63.5 seconds vs. 77 seconds, p < 0.05) and reported significantly higher confidence in their creations (medium effect size, r = 0.496). The prototype also encouraged more developmentally appropriate single-word hotspots (91.5% vs. 78.7%). However, three significant negative impacts emerged. First, over-reliance: 61.8% of AI-generated hotspots were used without any modification, and generated hotspots comprised 77.12% of all hotspots in prototype VSDs, suggesting participants passively accepted suggestions rather than critically evaluating them. Second, quality divergence from best practices: prototype VSDs used too many hotspots (only 40.63% stayed within the recommended 2-4 range vs. 60.5% for Snap Scene) and included more irrelevant hotspots about background objects (only 84.97% scene-relevant vs. 95.74%). Third, and most critically, homogenization: VSDs created with the prototype were significantly more semantically similar to each other (cosine similarity 0.39 vs. 0.46, p < 0.05, d = 0.60), with fewer unique hotspots (36.6% vs. 54.26%). The prototype also nearly eliminated social closeness hotspots (0.7% vs. 5.3%), focusing almost exclusively on information transfer.
Relevance
This study provides important evidence about the risks of integrating generative AI into assistive technology — risks that extend well beyond AAC. The homogenization finding is particularly concerning: VSDs are effective precisely because they are personalized to individual users' interests, relationships, and communication goals. When AI generates similar suggestions for everyone, it undermines the fundamental value proposition of the technology. For accessibility practitioners, the over-reliance finding highlights the danger of AI suggestions being treated as authoritative rather than as starting points, especially by non-experts. The paper suggests concrete mitigations including validation checks against best practices, reflection prompts before finalizing configurations, and more sophisticated prompting techniques. These findings should inform any project integrating AI into assistive technology or accessibility tools — the efficiency gains are real but must be weighed against the risk of reducing personalization, quality, and diversity of outcomes for disabled users who need individualized solutions.
Tags: augmentative and alternative communication · AAC · visual scene displays · generative AI · autism · speech-language pathology · just-in-time programming · AI homogenization · assistive technology design