Special education has always been about designing instruction around the individual. Every IEP, every accommodation plan, every transition goal starts with the same premise: this student has specific needs, and the system needs to meet them. So when a technology comes along that can personalize at scale, adapt in real time, and generate content tailored to a student’s reading level or communication profile, you’d think special education would be the first field to embrace it.
The reality, as Harkins-Brown, Carling, and Peloff (2025) describe in a recent overview published in Encyclopedia, is more complicated. The potential is enormous, but the evidence base is thin and the ethical questions are real.
The paper comes from Johns Hopkins University’s Center for Technology in Education, and it draws on several recent literature reviews to map how AI is being used with students with disabilities. The numbers paint an interesting picture. Harkins-Brown et al. cite a large-scale review by Chen et al. (2022) covering all AI studies published in education between 2000 and 2019, which found that roughly 14% of that research was related to special education, mostly focused on tutoring and individualized learning in K-12 settings.
A separate systematic review by Hopcan et al. (2023), cited in the paper, looked at 29 studies from 2008 to 2020 and found that most AI research in special education focused on students with autism, followed by students with learning disabilities. Nearly half of those studies addressed the cognitive and affective domains.
What’s useful about Harkins-Brown et al.’s approach is that they don’t just catalogue the research. They organize it around concrete applications. And some of the results are genuinely encouraging. An experimental study by Rakap and Balikci (2024), cited in the paper, found that ChatGPT-assisted IEP goals were significantly better in quality than those written by teachers without the tool.
AI-powered interventions have improved phonics skills for students with dyslexia (Aravena et al., 2018, cited in Harkins-Brown et al., 2025). Web-based training programs have increased working memory capacity in students with autism spectrum disorders (Calub et al., 2022, cited in the paper). And an AI-driven matching database proved effective in reducing unemployment rates among adults with disabilities (Abid et al., 2024, cited in the paper). These are all studies with measurable outcomes.
The teacher workload angle is where the paper connects most directly to what I keep hearing from educators. Special education teachers are buried in documentation. IEP writing, progress monitoring, curriculum adaptation, parent communication, translating materials into multiple languages, adjusting reading levels of texts, generating fluency passages and comprehension questions.
Harkins-Brown et al. cite Goldman et al. (2024), who argued that the nationwide teacher shortage in special education, driven in large part by workload, could be mitigated through thoughtful AI integration. I’ve covered similar arguments about AI and teacher workload in the context of the OECD‘s review of AI in special education, and the pattern is consistent: AI isn’t going to solve the structural problems in special education staffing, but it can take some of the most repetitive, time-consuming tasks off teachers’ plates.
Harkins-Brown et al. also trace the evolution from traditional assistive technology to AI-powered tools, and the shift is worth understanding. Traditional AT focused on specific functional needs: mobility, communication, sensory support. AI-based systems expand those capabilities into predictive assistance and real-time adaptation. And there’s something worth noting about the shift toward mainstream devices.
Tablets, laptops, and smartphones now come with built-in accessibility features, which means students can access support without the stigma that specialized equipment sometimes carries. Harkins-Brown et al. cite this as an important development, and I agree. Accessibility tools that blend into the technology everyone uses are more likely to be adopted and less likely to be abandoned.
The paper also discusses how AI paired with Universal Design for Learning (UDL) principles can go further than either one alone. Harkins-Brown et al. describe this combination as supporting diverse methods of representation, engagement, and expression, which aligns with broader equity goals.
I think the UDL connection is underexplored here, though. The paper mentions it but doesn’t give it the space it deserves. UDL isn’t just a nice-to-have. It’s a framework that, when combined with AI’s adaptive capabilities, could reduce the need for individual accommodations by building flexibility into instruction from the start. That’s a fundamentally different approach from the retrofit model most schools currently use.
The ethical section of the paper is where I want to slow down. Harkins-Brown et al. raise familiar concerns, bias in training data, privacy of student disability information, the need for anonymization, but they also raise a question that doesn’t get enough attention: accountability. If an AI system produces an incorrect or harmful recommendation that affects a student’s educational trajectory, who’s responsible? The developers? The educators who used the system? The administrators who implemented it? The paper doesn’t resolve that, and neither has anyone else. That gap is going to get wider as AI tools become more embedded in educational decision-making.
There’s also a concept in the paper from Marino et al. (2023), cited by Harkins-Brown et al., that I think is worth highlighting: the idea of AI as a “cognitive prosthesis.” It’s a bold framing. It positions AI not as a convenience or an efficiency tool, but as something that extends cognitive capability in the same way a wheelchair extends physical mobility. I can see why that language is powerful, and I can also see why it makes some people uncomfortable.
Harkins-Brown et al. are careful to add that AI shouldn’t be used in ways that frame disability as something to fix: “When AI is used to remove the perceived burden of disability, this construct is inherently based on ableist assumptions, which operates from a deficits-based perspective” (p. 6). That’s an important caution, and it’s the kind of nuance that tends to get lost when the conversation focuses only on what AI can do.
Where I’d challenge the paper is on its evidence base, and the authors themselves acknowledge this. Many of the initial studies involving AI and students with disabilities haven’t been replicated. Replication remains rare in special education journals, and that limits the field’s ability to identify truly evidence-based practices.
Harkins-Brown et al. are clear about this: “practitioners and policymakers need scientific evidence to know when and how to adopt it for use with students who have disabilities” (p. 7). That’s exactly right. The promise is real, but the gap between promising and proven is still wide. Special education, of all fields, can’t afford to adopt tools based on enthusiasm alone. The students are too vulnerable and the stakes are too high.
References
- Chen, X., Zou, D., Xie, H., Cheng, G., & Liu, C. (2022). Two decades of artificial intelligence in education: Contributors, collaborations, research topics, challenges, and future directions. Educational Technology & Society, 25(1), 28–47.
- Harkins-Brown, A. R., Carling, L. Z., & Peloff, D. C. (2025). Artificial intelligence in special education. Encyclopedia, 5(1), 11. https://doi.org/10.3390/encyclopedia5010011
- Hopcan, S., Polat, E., Ozturk, M. E., & Ozturk, L. (2023). Artificial intelligence in special education: A systematic review. Interactive Learning Environments, 31(10), 7335–7353.