Most of what I write about on this blog covers AI in general education settings. Classrooms full of neurotypical students, standard curricula, mainstream schools. But AI’s most meaningful impact might end up being somewhere else entirely: with students who have special education needs.
A new OECD paper by Linsenmayer (2025) takes a careful look at what AI tools are actually doing for students with learning disabilities, sensory impairments, and autism spectrum disorder. The findings are promising in spots, cautionary in others, and grounded in an evidence base that the author herself calls thin.
I wanted to cover this paper because it forces a conversation the AI-in-education field hasn’t had enough of. When we talk about AI and personalization, we usually mean adaptive quizzes or differentiated feedback for typical learners. Linsenmayer is talking about something fundamentally different: AI that reads handwriting pressure to detect dysgraphia, robots that learn how to interact with individual children on the autism spectrum, deep learning systems that translate sign language into text. The stakes are higher and the margin for error is smaller.
What the Tools Can Do
The strongest case for AI in special education comes from personalization, and Linsenmayer covers several tools that go well beyond generic adaptive learning. Dytective, she notes, adapts reading exercises to individual dyslexia profiles. Dynamilis analyzes the handwriting process itself, measuring pressure, speed, and tilt, with a machine learning algorithm that detected dysgraphia with approximately 96% accuracy in a 2018 study (Asselborn et al., 2019, cited in Linsenmayer, 2025). Linsenmayer also highlights Calcularis 2.0, which uses a dynamic Bayesian network to model each student’s mathematical knowledge in real time. These aren’t generic ChatGPT applications. They’re purpose-built tools designed around specific learning needs.
For students with sensory impairments, Linsenmayer’s review points to AI that builds on existing assistive technologies and makes them smarter. The EU-funded aiD project, for example, uses deep learning to translate between sign language and text. She also discusses UNICEF’s Accessible Digital Textbooks, which offer AI-powered features like text-to-speech, simplified text, and image descriptions through a Universal Design for Learning approach where students choose their own supports. That design philosophy is important as it puts control in the student’s hands.
Linsenmayer gives particular attention to socially assistive robots for students with ASD. The Kiwi robot, developed by USC’s Interaction Lab, used reinforcement learning to personalize interactions with 17 children over 30 days. The system detected engagement with 90% accuracy, and 92% of participants showed improvements in social skills.
The paper also describes the ECHOES project, which found something I think is even more interesting: AI-driven learning activities in a virtual environment actually sparked spontaneous social interactions between the children and human practitioners who were in the room. The technology didn’t replace human connection. It accidentally created more of it.
The Evidence Problem
And here’s where I need to be direct. The evidence behind most of these tools is thin. Linsenmayer says it plainly: “This review confirms a gap in robust research on the effects of AI-based tools on students with SEN” (p. 33). Most studies have small sample sizes. Cultural diversity in the research is limited. Randomized controlled trials and longitudinal designs are rare. Some tools claim effectiveness for students with special education needs without actually focusing on these students in their impact studies.
That’s a serious problem. Small pilot studies with 17 or 19 participants can show proof of concept. They can’t tell you whether a tool works across different schools, different countries, different presentations of the same condition. ASD alone varies so widely from person to person that any small-sample finding needs to be treated with extreme caution.
I’m not saying these tools don’t work. I’m saying we don’t have enough evidence to know how well they work, for whom, and under what conditions. And that should give anyone pause before rolling them out at scale, especially with a population that has already been underserved by mainstream educational technology.
Algorithmic Bias Hits Harder Here
Linsenmayer’s discussion of algorithmic bias carries more weight than it does in most AI-in-education research. As she points out, when AI systems are trained on biased datasets, the consequences compound for students with special education needs. These risks multiply when disability intersects with race, gender, or socioeconomic background.
She cites a study that evaluated 189 facial recognition algorithms from 99 developers and found the highest rates of false positives among Indigenous peoples. Now imagine that kind of bias embedded in a diagnostic tool meant to identify learning disabilities. The students who need the most accurate assessments would get the least reliable ones.
This connects to a larger pattern across the AI research I’ve been covering. Biased training data doesn’t just produce incorrect outputs. It produces confidently incorrect outputs, the kind that people accept without questioning because the system seems authoritative. For students with SEN, who depend on accurate identification and targeted support, the cost of that false confidence is measured in lost years of appropriate intervention.
The Environmental Question
One dimension of this paper that I haven’t seen in most education-focused AI research is the environmental cost argument. Linsenmayer notes that a single LLM query uses roughly 2.9 watt-hours of electricity, compared to 0.3 for a regular internet search. Training AI models requires massive amounts of energy and water for data center cooling. Projections suggest that by 2027, global AI water demand could surpass the United Kingdom’s total annual water usage.
The author frames this as a cost-benefit question, and I think that framing is exactly right:
More research is needed to evaluate whether the marginal benefits of using AI are justified given the risks of AI in terms of data misuse, financial and environmental costs, and biases. Education stakeholders need this type of comparative evidence to make informed decisions about which tools they invest in and integrate into their educational practice, particularly for those that involve students
with SEN.(p. 33)
For some use cases, like helping students with ADHD track tasks, a well-designed non-AI app might be equally effective and far more sustainable. The default assumption that AI is always the best tool needs questioning, even when the intention behind its use is genuinely good.
Techno-Ableism and Co-Design
Linsenmayer flags a concern I think deserves serious weight in any conversation about AI and disability: techno-ableism. Citing Shew (2020), she describes it as “the assumption that technology should render individuals able-bodied and neurotypical” (p. 32). The authors intentionally excluded tools from their review that framed students with SEN as needing to be “fixed.” That’s a methodological choice rooted in values, and I respect it. The alternative Linsenmayer proposes is participatory design, where people with disabilities, their families, and their teachers help shape the tools from the start.
What Educators Should Know
Linsenmayer closes with four policy priorities: ethical design guided by Universal Design for Learning principles, comparative research with long-term monitoring, strong data protection for vulnerable populations, and common accountability frameworks including AI “nutrition labels” modeled on the health sector’s Coalition for Health AI. On the accountability front, the author argues that “education systems could require AI-enabled tools to receive certification before allowing them into classrooms” (p. 42).
I agree with the direction. My concern is speed. Certification requirements, if poorly designed, could slow the availability of tools that students with SEN urgently need. The balance between protecting vulnerable students and getting them timely access to beneficial tools is one the paper acknowledges but doesn’t fully resolve.
Linsenmayer also calls for teacher professional development that goes beyond how to use AI tools. Teachers working with students with SEN need critical AI literacy: how AI techniques work, how to evaluate AI outputs, and how to weigh costs and benefits, including environmental ones, when deciding if and when to use AI. I’ve been making this argument across many posts on this blog, and it applies with special force here. The teachers supporting the most vulnerable students need the deepest understanding of what these tools can and cannot do.
References
- Asselborn, T. et al. (2018), “Automated human-level diagnosis of dysgraphia using a consumer tablet”, npj Digital Medicine, Vol. 1/1, https://doi.org/10.1038/s41746-018-0049-x.
- Linsenmayer, E. (2025). Leveraging artificial intelligence to support students with special education needs (OECD Artificial Intelligence Papers No. 46). OECD Publishing. https://www.oecd.org/en/publications/leveraging-artificial-intelligence-to-support-students-with-special-education-needs_1e3dffa9-en.html
- Shew, A. (2020), “Ableism, Technoableism, and Future AI”, IEEE Technology and Society Magazine, Vol. 39/1, pp. 40-85, https://doi.org/10.1109/mts.2020.2967492.