Autistic children exhibit heterogeneous oral language impairments, necessitating educators to implement personalized teaching content. However, preparing personalized materials remains time-intensive and difficult to maintain coherence, while generative AI's recent advances in creating customized content show potential to support this process. We first conducted video analysis from educators' one-on-one classes with autistic students and conducted interviews with therapists to understand their challenges in current teaching practices. Then, we developed a generative AI-empowered prototype, LingoLift, which supports educators to create interest-based, ability-adapted, and coherent teaching materials according to children's profiles. Finally, we conducted a three-week deployment study with 10 educator-student dyads completing 30 lessons with LingoLift in a specialized education school. Results showed that LingoLift significantly improved lesson preparation efficiency, reduced educators' workload, and enabled children to achieve positive learning outcomes. We observed educators' adaptive extensions and innovations, revealing insights into design considerations and future opportunities for AI-assisted inclusive education.
ACM CHI Conference on Human Factors in Computing Systems