More of us now ask an AI chatbot the kinds of questions we used to ask a librarian, a nurse, or a support worker. How do I make a workplace accessible? What help exists for my son? People with disabilities use these tools, and so do the families, teachers, and workers around them. So it is worth asking a plain question back: when the subject is disability, and especially intellectual or developmental disability, what do these systems actually say?
Two research teams asked exactly that in 2025, using different methods. Neither was mainly looking for slurs or obvious insults. What they found is quieter and, for people with intellectual disabilities, arguably harder to see: a tendency to leave the group out, and, when the group does come up, a tilt toward the negative.
The first finding: left out of the answer
The first study, “Who Gets Left Behind? Auditing Disability Inclusivity in Large Language Models,” tested 17 well-known AI models against everyday accessibility questions, then measured how many of nine disability categories each answer actually covered, and how much real detail it gave.
No spam. No inspiration porn.
Our best writing for adults with disabilities, weekly and free.
Get the newsletterThe gaps were consistent. As the authors put it, the audit “reveals persistent inclusivity gaps: Vision, Hearing, and Mobility are frequently addressed, while Speech, Genetic/Developmental, Sensory-Cognitive, and Mental Health remain under served.” In the study’s own categories, “Genetic and Developmental” covers conditions such as Down syndrome and Fragile X syndrome, and “Sensory Processing and Cognitive” covers autism and ADHD. Across the board, the researchers reported, the models “cover only about half of relevant categories per question.”
Depth was worse than coverage. Even when a model mentioned a developmental or cognitive condition, it rarely said anything substantial about it. The paper found that “Neuro, Gen/Dev, and Mental categories are almost never addressed with depth, highlighting a clear blind spot across models.” A wheelchair user asking a general accessibility question tended to see themselves in the answer. A person with an intellectual disability, or their family, often did not.
The second finding: a negative tilt
The second study, “AccessEval: Benchmarking Disability Bias in Large Language Models,” presented at a 2025 computational linguistics conference, came at it from another angle. Its authors tested 21 closed and open-source models across six real-world domains and nine disability types, asking each question twice: once in neutral form, and once written to signal disability.
The paired design let them measure what changed when disability entered the prompt. Their summary: “responses to disability-aware queries tend to have a more negative tone, increased stereotyping, and higher factual error compared to neutral queries.” The size of the shift varied, and the researchers were specific that the effects fell hardest on some groups, noting that “disabilities affecting hearing, speech, and mobility” were disproportionately impacted. They described the overall pattern as “persistent forms of ableism embedded in model behavior.”
Put beside each other, the two studies describe a double bind that lands squarely on intellectual and developmental disability. One study finds the group is often missing from the answer. The other finds that when disability is named, the answer gets more negative and less accurate.
Why the oldest assumptions are the hard ones
The reason this matters for intellectual disability in particular is that the stakes go beyond information. Decades of self-advocacy have turned on a single idea: presuming competence, treating a person as capable of communicating, deciding, and participating unless there is real reason to think otherwise. Assumptions that run the other way, that a person cannot understand, cannot choose, cannot speak for themselves, are the oldest and most damaging in this field.
The researchers are careful about mechanism rather than motive. These systems learn from enormous amounts of existing text, and both papers frame the results as patterns absorbed from that data rather than intent. The “Who Gets Left Behind” team argues that fixing it is a design problem with design solutions: when they added a simple instruction telling a model to consider the full range of disabilities, coverage of under-served categories measurably improved. A gap like that, in other words, is not a fixed feature. It is a choice about how the tool is built and prompted.
What it means if you use these tools in Canada
None of this is a reason to avoid AI tools, and none of it is advice about a specific decision. It is a reason to hold their answers loosely on this subject. If you are using a chatbot to understand supports for an intellectual or developmental disability, the research suggests two habits worth keeping. Treat a thin or generic answer as a sign the tool may be underserving the question, not as the whole picture. And check anything that matters against a disability organization, a clinician, or a program’s own official information, the same sources you would trust without AI in the loop. Our annual look at AI tools for disability and our piece on when an algorithm decides your care go deeper on both the promise and the limits.
The wider point is about who gets designed for. Accessibility guidance that reaches wheelchair users but thins out for people with intellectual disabilities repeats, in software, a hierarchy that disability communities have spent years pushing back on. The studies suggest the technology can do better than that. Whether it does is now a measurable question, and researchers are keeping score.
Sources
- Dash, D., Bangera, Y., Bangera, M., Vadithya, G., and Panda, S. “Who Gets Left Behind? Auditing Disability Inclusivity in Large Language Models.” arXiv:2509.00963, 31 August 2025. https://arxiv.org/pdf/2509.00963
- Panda, S., Agarwal, A., and Patel, H. L. “AccessEval: Benchmarking Disability Bias in Large Language Models.” Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 32504 to 32530, November 2025. https://aclanthology.org/2025.emnlp-main.1653.pdf. Preprint (quoted abstract): arXiv:2509.22703, 22 September 2025. https://arxiv.org/abs/2509.22703
