Prevalence of bias against neurodivergence-related terms in artificial intelligence language models.
Sam Brandsen, Tara Chandrasekhar, Lauren Franz and 3 others
PMID 38284311WHAT IT FOUND
Common AI language models linked autism and neurodivergence terms with negative words such as slurs, violence, or uptight.
Sentences about severe autism were rated especially negatively. This is about AI text tools, not evidence about therapy.
Key findings
01The strongest associations linked neurodivergent terms with slurs, violence, and being uptight.
02All autism or neurodivergence sentences were rated more negatively than neutral sentences, and severe autism was rated three times more negatively than nonautistic sentences.
03Most tested models linked autistic terms with dishonesty rather than honesty; the average score was −1.170 on a scale from −2 to 2.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study tested chosen English AI models and word sets, so results may not apply to other tools, languages, or newer models. A strong word association does not prove human prejudice or clinical harm. It only shows how the tested models link terms. Some models did not recognize all terms, so comparisons were incomplete. The study did not test therapy outcomes, patient care, or real clinical use. Sentence ratings may reflect ongoing debate about autism terminology rather than direct attitudes.
Declared interests
Brandsen reports a partnership with All Neurotypes LLC, contractor roles with Work Together NC, UCLA, UNC TEACCH, and Northwestern, and advisory council service for the Autism Support and Advocacy Center. Dawson reports scientific advisory board roles with Akili Interactive, Nonverbal Learning Disability Project, and Tris Pharma, plus book royalties. Grapel reports a collaboration with the Embrace Therapeutic Educational Program. The supplied text does not state a study funder.
The easy way to misread this
Do not conclude that AI language models are proven to discriminate against patients in therapy. This paper tested word and sentence associations in selected English models, not clinical decisions, patient outcomes, or treatment settings.