Generative Artificial Intelligence and Large Language Models in Rehabilitation and Disability Services: A Scoping Review

Abstract

The rapid emergence of generative artificial intelligence (GenAI) tools and large language models (LLMs) since late 2022 has introduced new possibilities for rehabilitation and disability services. However, no review has systematically mapped the scope and nature of evidence on GenAI applications specifically in these contexts. This scoping review aimed to map the evidence on GenAI and LLM applications in rehabilitation and disability services, examining the tools applied and their purposes (SQ1), populations and service contexts represented (SQ2), reported outcomes, benefits, and concerns (SQ3), and evidence gaps with implications for future research, practice, and policy (SQ4). Sources published from late 2022 to May 2026 in English, addressing the application of GenAI or LLMs within rehabilitation, disability, or special education service contexts, involving persons with disabilities, rehabilitation professionals, special educators, or related stakeholders. Scopus, Web of Science, and Lens were searched in May 2026. Screening reliability was assessed computationally (Cohen’s κ = 0.938). Data were extracted using Elicit across 10 variables and reviewed by the principal investigator. Synthesis followed a narrative and descriptive approach organised by the four sub-questions. Ninety-six studies published between 2023 and 2026 were included. ChatGPT was referenced in 85.4% of studies. The dominant purposes were assistive and adaptive technology (40.6%), professional training (21.9%), and content generation (21.9%). Higher education rehabilitation programmes (51.0%) and school-level special education (28.1%) were the most common service contexts. Learning disabilities (17.7%), autism spectrum disorder (14.6%), and visual impairment (11.5%) were the most specified disability types, though 56.2% of studies did not specify a disability diagnosis. Reported benefits included measurable improvements in learning outcomes under structured, educator-guided conditions, significant time savings for special educators and rehabilitation professionals, and enhanced autonomy for students with disabilities. Concerns included clinical reasoning accuracy limitations, AI-generated accessibility failures, readability exceeding health literacy thresholds, and equity barriers in low-resource settings. Critical gaps included the near-total absence of longitudinal outcome data, very limited direct patient-facing clinical evidence, sparse representation of complex disability profiles, and no governance or institutional policy research. The evidence supports cautious, educator-guided integration of GenAI in rehabilitation and special education training, with structured pedagogical design and human oversight as critical conditions for benefit. The evidence does not yet support conclusions about patient outcomes, long-term competency development, or population-level equity effects. Future research should prioritise longitudinal designs, direct patient-facing clinical studies, under-represented disability populations, and disability-inclusive AI governance frameworks.