DisaBench: A Participatory Framework for Evaluating Disability Harms
๐กStandard benchmarks miss disability bias; use DisaBench to improve your model's safety and inclusivity.
โก 30-Second TL;DR
What Changed
Introduces a taxonomy of twelve disability harm categories co-created with domain experts and people with disabilities.
Why It Matters
This framework addresses a critical gap in AI safety by moving beyond overt failure detection to capture nuanced, intersectional harms. It enables developers to build more inclusive models by incorporating lived-experience data into their evaluation pipelines.
What To Do Next
Integrate the DisaBench red teaming framework into your model's safety evaluation pipeline via Hugging Face to identify previously overlooked disability-related biases.
Key Points
- โขIntroduces a taxonomy of twelve disability harm categories co-created with domain experts and people with disabilities.
- โขProvides a dataset of 175 prompts with 525 human-annotated prompt-response pairs for safety testing.
- โขReveals that standard safety benchmarks often miss subtle, context-dependent harms that require lived experience to identify.
- โขOffers an open-source red teaming framework for integration into existing AI safety pipelines.
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI โ