DisaBench: New Framework Addresses Disability Harms in AI Models

Eugenia Kim and her team have developed DisaBench, a groundbreaking participatory evaluation framework that focuses on assessing disability-related harms in large language models. Recognizing that existing safety benchmarks fall short, DisaBench introduces a taxonomy of twelve specific disability harm categories co-created with individuals with disabilities and red teaming experts. The framework includes a dataset of 175 prompts, which allows for a nuanced evaluation across seven life domains, ensuring that evaluations consider the context of users’ experiences.

Key findings from the evaluation process reveal that harm rates differ significantly based on disability types, with terminology-driven harm being culturally and temporally specific. Standard safety evaluations typically capture overt failures but often miss subtle, nuanced harms that require domain expertise to identify. By releasing the dataset, taxonomy, and methodology via Hugging Face, DisaBench aims to integrate seamlessly into existing safety pipelines, fostering a more inclusive approach to AI development and deployment.