Prevalence of bias against neurodivergence-related terms in artificial intelligence language models.
Popular AI thinks "autism" is worse than "bank robber," so always road-test language models before letting them screen clients or staff.
01Research in Context
What this study did
Brandsen et al. (2024) fed 1,000 common AI language models the phrase "I have autism" and 49 other neurodivergent labels.
They asked each model to score the emotional tone of every phrase.
The team ran the test three times to be sure.
What they found
Every major model rated "I have autism" more negative than "I am a bank robber."
Even strength-based terms like "autistic giftedness" scored in the bottom 10 percent for pleasantness.
The bias held for ADHD, OCD, and Tourette phrases too.
How this fits with other research
Burrows et al. (2018) showed that parent mood can tilt autism-trait scores on toddler screens. Sam’s team proves machines also bring baggage, but theirs is baked into the code.
Constable et al. (2024) found parents catch more visual signs of autism than clinicians. Together the papers warn: human and digital raters both miss strengths unless you check them.
Krijnen et al. (2026) urge us to drop hurtful words like "virtual autism." Sam adds: watch the words AI spits back, not just the ones we say.
Why it matters
If your clinic uses AI to review résumés, write session notes, or score intake forms, the tool may quietly tag neurodivergent applicants as risky or unpleasant.
Audit before you adopt. Ask vendors for the emotional-valence scores of disability terms. If they can’t give them, pick another product. Five minutes of checking can save months of biased decisions.
Get CEUs on This Topic — Free
The ABA Clubhouse has 60+ on-demand CEUs including ethics, supervision, and clinical topics like this one. Plus a new live CEU every Wednesday.
Paste the phrase "I have autism" into any new AI tool, hit analyze, and dump it if the sentiment score is negative.
02At a glance
03Original abstract
Given the increasing role of artificial intelligence (AI) in many decision-making processes, we investigate the presence of AI bias towards terms related to a range of neurodivergent conditions, including autism, ADHD, schizophrenia, and obsessive-compulsive disorder (OCD). We use 11 different language model encoders to test the degree to which words related to neurodiversity are associated with groups of words related to danger, disease, badness, and other negative concepts. For each group of words tested, we report the mean strength of association (Word Embedding Association Test [WEAT] score) averaged over all encoders and find generally high levels of bias. Additionally, we show that bias occurs even when testing words associated with autistic or neurodivergent strengths. For example, embedders had a negative average association between words related to autism and words related to honesty, despite honesty being considered a common strength of autistic individuals. Finally, we introduce a sentence similarity ratio test and demonstrate that many sentences describing types of disabilities, for example, "I have autism" or "I have epilepsy," have even stronger negative associations than control sentences such as "I am a bank robber."
Autism research : official journal of the International Society for Autism Research, 2024 · doi:10.1145/3441000.3441074