The Invisible 30% - ETEC 565T

Uncovering AI’s Neurodiversity Blind Spots

GenAI and LLMs have come on the scene and rapidly changed many aspects of our lives, work and school. As more of the tools we use on a regular basis begin integrating LLMs, it’s important that we know whose voices are shaping the outputs and whose voices are fading into the background.

Many of us know about Autism Spectrum Disorder (ASD), whether through a personal connection or relationship, or from representation in popular media (such as shows like Atypical, or The Good Doctor). Dr. Stephen Shore famously said “if you’ve met one person with autism, you’ve met one person with autism” to emphasize how vast the needs, behaviours and supports are between individuals with autism. This has certainly mirrored my own professional experience; working with the autistic community constantly requires adapting my methods for communication, assessment, and teaching from one individual to the next.

Because LLMs are trained primarily on vast amounts of text-based internet data, I was curious to see how individuals with high needs, or those who have minimal language would be represented within these systems. To explore this, I gave 3 popular LLMs the same prompt and conducted a qualitative analysis of the results.

What I discovered was troubling. The outputs essentially flattened a highly fluid and complex spectrum into a handful of digestible tropes and stereotypes. This phenomenon perfectly illustrates what scholar Beth Coleman describes as supervised machine learning functioning primarily as a "reproduction of a status quo" (2021, p. 6). The algorithmic erasure of individuals with higher support needs isn't necessarily due to malicious programming, but is rather caused by the "empirical patterns of 'big data' culture on which AI relies" (2021, p.6). With research indicating that about 25% to 30% of autistic children remain minimally verbal or non-speaking by school age (Molko 2026), we are effectively leaving a massive portion of this community entirely unrepresented in this pervasive new technology.

In the presentation below, I analyze how these biases might manifest in real-world educational and professional settings, and offer practical strategies for users to identify and address these blind spots.

References

Coleman, B. (2021). Technology of The Surround. Catalyst: Feminism, Theory, Technoscience, 7(2), 1–21.

Eubanks, V. (2017;2018;). Automating inequality: How high-tech tools profile, police, and punish the poor (First ed.). St. Martin's Press.

Molko, R. (2026, June 25). What is nonverbal autism? Dispelling the myths around non speaking autism. LEARN Behavioral. https://learnbehavioral.com/blog/myth-nonverbal-or-nonspeaking-people-with-autism-are-intellectually-disabled

Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism (1st ed.). New York University Press. https://doi.org/10.18574/9781479833641

Treffert D. A. (2009). The savant syndrome: an extraordinary condition. A synopsis: past, present, future. Philosophical transactions of the Royal Society of London. Series B, Biological sciences, 364(1522), 1351–1357. https://doi.org/10.1098/rstb.2008.0326

Understanding autism data in the Canadian Chronic Disease Surveillance System. Government of Canada. (2025, December 23). https://health-infobase.canada.ca/autism/ 08.0326