From Technical Performance to Assistive Validity: A Critical Review of Evaluation Approaches for AI-Enabled Assistive Technology

Abstract

Background/Objectives: The increasing integration of artificial intelligence (AI) into assistive technologies challenges evaluation models traditionally centred on usability, satisfaction and device-related outcomes. Because AI-enabled systems are probabilistic, data-dependent and adaptive, their evaluation must also address algorithmic behaviour, user agency and consequences in everyday life. This critical review aimed to identify and compare standardised instruments, frameworks and structured procedures for evaluating AI-enabled assistive technologies and to determine the extent to which they connect AI-specific properties with the goals of persons with disabilities, activities, participation and environmental conditions. Methods: MEDLINE/PubMed, Scopus and IEEE Xplore were searched without publication-date restrictions, with final searches completed on 30 June 2026 and complemented by backward and forward citation searching. Data were extracted on approach type, purpose, stage of application, target populations and technologies, methodological evidence, evaluated dimensions and outcomes. Evidence was classified as documented, partially documented, not documented or not applicable, while dimensional coverage was coded as explicit, partial or absent. Results: Of the 2299 records screened by title and abstract, 478 publications underwent full-text assessment and 45 were retained in the documentary corpus. These publications supported 16 evaluation approaches: four frameworks, three structured procedures and nine measurement instruments published between 2001 and 2026. The person-related dimension was explicitly operationalised in 15 approaches and activity in nine, whereas participation was explicit in only one and environmental factors in six. AI-specific properties were explicitly evaluated in three approaches, and everyday-life outcomes over time in two. Among the approaches included in this review, none combined comprehensive coverage of person, activity, participation and environment with AI-specific evaluation and longitudinal monitoring. Conclusions: Evaluation remains fragmented across assistive technology, human–computer interaction, human–robot interaction and AI assessment traditions. The review proposes assistive validity as a higher-order criterion linking technical performance, agency and meaningful outcomes for persons with disabilities. It also identifies a modular and longitudinal evaluation architecture as a priority for future empirical development and validation.