AI for accessibility requires evaluation methods that go beyond standard metrics to capture the complex, intersectional needs of real users—one-size-fits-all approaches will fail people who depend on these systems.
This paper examines how AI can improve augmentative and alternative communication (AAC) systems for people with speech disabilities, but argues that standard evaluation metrics miss important nuances about what users actually need. The authors study six real AAC challenges and propose evaluation methods that account for the diverse, intersecting needs of individual users.