DisaBench: A Participatory Evaluation Framework for Disability Harms in Language Models

📰 ArXiv cs.AI

Learn to evaluate disability harms in language models using DisaBench, a participatory framework co-created with people with disabilities

advanced Published 14 May 2026
Action Steps
  1. Identify disability harm categories using DisaBench taxonomy
  2. Create pairs of benign and adversarial prompts across life domains
  3. Evaluate language model responses using human-annotated labels
  4. Analyze results to detect potential disability harms
  5. Refine language model training data to mitigate identified harms
Who Needs to Know This

AI researchers and developers can use DisaBench to identify and mitigate disability-related harms in their language models, ensuring more inclusive and safe AI systems

Key Insight

💡 Disability harms in language models can be identified and mitigated using a participatory evaluation framework like DisaBench

Share This
🚨 Introducing DisaBench: a framework to evaluate disability harms in language models 🚨
Read full paper → ← Back to Reads