LLMs Show Anti-Autistic Ableism, Study Finds
A recent study published on arXiv (2605.26397) indicates that large language models (LLMs) often generate harmful content and incorrectly categorize community-reclaimed language as ableist. The researchers propose a new evaluation framework that accounts for bias, utilizing psychometrically-weighted, community-relevant ground truth based on the positionality of annotators, which is a more rigorous measure than traditional majority-vote methods. Their findings reveal that majority-vote approaches consistently undervalue autistic and autism-accepting viewpoints. Additionally, LLMs demonstrate more negative sentiments towards autistic individuals when assessment tools are obscured. The focus of the research is on detecting anti-autistic ableist language.
Key facts
- Study on arXiv 2605.26397
- LLMs amplify or suppress perspectives in high-stakes settings affecting autistic communities
- Previous research identified disability-related biases in LLMs
- New framework uses psychometrically-weighted, community-proximate ground truth
- Majority-vote aggregation underweights autistic and autism-accepting perspectives
- LLMs mislabel community-reclaimed language as ableist
- LLMs express more negative attitudes toward autistic people when assessment instruments are masked
- Error analysis reveals harmful outputs from LLMs
Entities
Institutions
- arXiv