Let's CHAT About Artificial Intelligence for Students With Disabilities: A Systematic Literature Review

Synthesis: The first meta-analysis of AI-based interventions for students with disabilities (SWDs), synthesizing 29 (quasi-)experimental studies conducted globally and analyzed through the lens of cultural-historical activity theory (CHAT). AI-based interventions produced a statistically significant medium overall effect on SWDs' learning outcomes (Hedges' g = 0.588) across robots, computer software (speech recognition, intelligent tutoring systems, expert systems), and intelligent VR systems — operating as social-emotional coaches, instructional/learning tools, and teachable agents. Notably, no participant-, AI-, interaction-, intervention-, or methodology-related moderator was statistically significant. The paper advances Inclusive Learning and Accessibility by documenting that AI works for SWDs, and calls for a shift from ensuring mere access toward positioning SWDs in agentic roles — contributing to Agentic AI and Learner Agency. As a medium effect on SWD learning outcomes, it is a key evidence base for the Learning Gains concept.

Core Finding

A medium, statistically significant overall effect of AI-based interventions on SWDs' learning outcomes: Hedges' g = 0.588 (95% CI [0.349, 0.826], p < .01), across 239 effect sizes from 41 independent samples (RVE random-effects model). Between-study heterogeneity was high (τ² = 0.52; I² = 76.17), and sensitivity analyses confirmed the estimate was robust. Effect sizes varied by subgroup — but none of the moderators reached statistical significance:

• By learning outcome: Academic performance (k = 80, g = 0.929) was highest, followed by daily life/other skills (g = 0.766) and social-emotional skills (k = 144, g = 0.382) — no significant difference.
• By disability: Students with SLD, IDD, or who are deaf (g = 0.952) showed a larger effect than students with ASD (k = 150, g = 0.368), though not significantly so.
• By AI type: Computer software (k = 74, g = 0.959) and robots (k = 144, g = 0.509) were effective; intelligent VR (g = 0.528) was not statistically significant.
• By AI role: Teachable agents (g = 1.100) and instructional/learning tools (g = 0.863) outperformed social-emotional coaches (g = 0.336), though not significantly.
• By interaction type: Dyadic AI-SWD interactions (g = 0.973) were more effective than triadic AI-SWD-community interactions (g = 0.385), though not significantly.

Publication bias was present (Egger's test β = 2.837, p < .001); trim-and-fill adjustment reduced the effect to g = 0.2694, which remained statistically significant.

The CHAT Theoretical Lens

The authors employ second-generation cultural-historical activity theory (CHAT, after Vygotsky and Engeström) to frame AI-SWD interaction as a dynamic activity system: SWDs (subject) use AI tools (mediating tools) to pursue learning tasks (object), embedded within a community, rules, and divisions of labor. This theoretical frame — rooted in Vygotsky's sociocultural view of disability as socially constructed rather than purely biological — guides the construction of moderators across participant, AI, AI-SWD-community interaction, intervention, and methodology characteristics.

Types and Roles of AI for SWDs

The meta-analysis identifies three AI types and three AI roles:

• AI types: robots (70.0% of studies, mostly humanoid robots for students with ASD, often human-operated via the Wizard of Oz method), computer software (speech recognition, expert systems, intelligent tutoring systems), and intelligent VR systems.
• AI roles: social-emotional coaches/companions (n = 19), instructional/learning tools (n = 8), and teachable agents (n = 2). Teachable agents showed the largest effect (g = 1.100) — the rare arrangement where SWDs take an active, even teacher-like, role.

From Access to Agency

A central argument distinguishes accessibility (ensuring SWDs can access learning content and functional capabilities) from agentic participation. Most AI tools studied augmented one or more functional abilities needed to access learning. But only two studies (Brainin et al., 2022; Wilson, 1997) positioned AI as teachable agents that enabled SWDs to take a relatively agentic role. The authors call for AI that "not only ensures accessibility but also promotes opportunities for SWDs to take an agentic role in participating in and contributing to AI-mediated learning activities" — a strengths-based, agentic reframing of SWD participation in learning.

Rules, Equity, and the Distribution of Responsibilities

Consistent with CHAT, the paper foregrounds community, rules, and division of labor. Most studies (96.6%) used step-based instruction in special education or clinical settings, and most were triadic (involving community members such as therapists, educators, families). The authors note the absence of research on AI supporting collective learning among SWDs and peers in inclusive settings, and call for human-centered design, addressing algorithmic bias, and strengths-based values in AI design and deployment for SWDs.

Relevance to the Knowledge Base

This is a landmark contribution to Special Education, Inclusive Learning, Accessibility, Robots in Education, and Learning Gains. As the first meta-analysis of AI for SWDs, it provides the strongest quantitative evidence that AI-based interventions yield a medium positive effect on SWD learning outcomes, directly supporting the Learning Gains concept's evidence base for Special Education. It is the clearest empirical anchor for the knowledge base's distinction between Accessibility (ensuring access) and Agentic AI/Learner Agency (agentic participation) — explicitly recommending AI that promotes SWDs' agentic roles. It also synthesizes evidence on Robots in Education (robot-assisted instruction for ASD), virtual reality (intelligent VR), Intelligent Tutoring (ITS), and Student Engagement, with implications for K-12 (all studies were PK-12) and Equity (cultural-historical, strengths-based, anti-deficit framing). Notably, most interventions relied on less-advanced rule-based AI and human-operated robots rather than modern Generative AI; the authors flag advanced AI techniques as underexplored for SWDs.

What this means for practice

• Instructors. Give students with disabilities direct, supported time with the AI tool: dyadic AI-SWD interactions showed the largest effect (g = 0.973) versus triadic arrangements involving community members (g = 0.385), so avoid routing every session through an adult intermediary.
• Instructors. Do not expect social-emotional coaching to carry the academic load: AI as a social-emotional coach produced the smallest effect (g = 0.336), while computer software (g = 0.959) and instructional/learning tools (g = 0.863) were stronger — most robot studies targeted social-emotional outcomes for students with ASD.
• Designers. Build teachable-agent activities where students with disabilities instruct, script, or correct the AI: teachable agents showed the largest effect (g = 1.100) and only two of the 29 included studies used them, so the arrangement is both promising and rare.
• Researchers. Treat Accessibility as a floor rather than the goal: only two studies (Brainin et al., 2022; Wilson, 1997) positioned SWDs in an agentic role, and none examined collective learning with peers in inclusive settings, so design studies that test equitable, strengths-based participation.

Limitations

• Publication bias was present (Egger's test β = 2.837, p < .001); trim-and-fill added 61 studies and reduced the pooled effect from g = 0.588 to g = 0.2694, so the headline medium effect is likely inflated even though the adjusted estimate stayed significant.
• Heterogeneity was high (τ² = 0.52; I² = 76.17) and none of the participant-, AI-, interaction-, intervention-, or methodology-related moderators reached significance; the authors note the null moderating effects may reflect small subcategory samples and that heterogeneity is often confounded with publication bias, so they urge caution in reading the overall positive effect.
• The corpus is 29 studies from 24 peer-reviewed articles and 4 dissertations published between January 1994 and January 2023, drawn from ERIC, PsycInfo, PubMed, IEEE Xplorer, and ProQuest and limited to English-language documents; single-case designs and conference proceedings were excluded, which the authors acknowledge may omit relevant evidence.
• The technology base is dated: robots made up 70.0% of studies and most were human-operated via the Wizard of Oz method, only two studies used teachable agents, and intelligent VR was not statistically significant (g = 0.528), so the findings do not speak to modern generative AI for SWDs.

Citation

Zhang, L., Carter, R. A., Jr., Liu, Y., & Peng, P. (2024). Let's CHAT About Artificial Intelligence for Students With Disabilities: A Systematic Literature Review and Meta-Analysis. Review of Educational Research, 96(1), 215-257.