{"paper":{"title":"DisaBench: A Participatory Evaluation Framework for Disability Harms in Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"General-purpose safety benchmarks for language models miss disability harms because those harms are personal, intersectional, and community-defined.","cross_cats":["cs.HC"],"primary_cat":"cs.AI","authors_text":"Christina Mallon, Eugenia Kim, Ioana Tanase","submitted_at":"2026-05-12T19:56:36Z","abstract_excerpt":"General-purpose safety benchmarks for large language models do not adequately evaluate disability-related harms. We introduce DisaBench: a taxonomy of twelve disability harm categories co-created with people with disabilities and red teaming experts, a taxonomy-driven evaluation methodology that pairs benign and adversarial prompts across seven life domains, and a dataset of 175 prompts with human-annotated labels on 525 prompt-response pairs. Annotation by four evaluators with lived disability experience reveals three findings: harm rates vary sharply by disability type and will compound in n"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Disability harm is simultaneously personal, intersectional, and community-defined: it cannot be isolated from the full context of who a person is, and general-purpose benchmarks systematically miss it.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The taxonomy and 175 prompts co-created with four evaluators with lived experience are sufficient to capture the range of disability harms across cultures, time, and non-text modalities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DisaBench supplies a participatory taxonomy of twelve disability harm types, paired benign-adversarial prompts across seven life domains, and human-annotated data showing that standard safety tests miss context-dependent harms.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"General-purpose safety benchmarks for language models miss disability harms because those harms are personal, intersectional, and community-defined.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"59cb294152cf3cfd27f7b2df38e0467e15e27da0e4bf372b65db2927ae39daae"},"source":{"id":"2605.12702","kind":"arxiv","version":1},"verdict":{"id":"349ab228-6906-4ba6-ae03-4a1dbe6ed5b8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:05:54.469728Z","strongest_claim":"Disability harm is simultaneously personal, intersectional, and community-defined: it cannot be isolated from the full context of who a person is, and general-purpose benchmarks systematically miss it.","one_line_summary":"DisaBench supplies a participatory taxonomy of twelve disability harm types, paired benign-adversarial prompts across seven life domains, and human-annotated data showing that standard safety tests miss context-dependent harms.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The taxonomy and 175 prompts co-created with four evaluators with lived experience are sufficient to capture the range of disability harms across cultures, time, and non-text modalities.","pith_extraction_headline":"General-purpose safety benchmarks for language models miss disability harms because those harms are personal, intersectional, and community-defined."},"references":{"count":45,"sample":[{"doi":"","year":2001,"title":"2001 , address =","work_id":"cebb7cbf-88a3-443d-8e4c-9d09cbe97fd6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2011,"title":"2011 , address =","work_id":"50c70b1f-585f-45ed-8329-15ed16da500a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"2024 , howpublished =","work_id":"df930ec5-3ad9-4914-b196-ac7912e6c387","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Applied Sciences , volume =","work_id":"804251fd-cced-4db5-a24a-0dc8a22b30b0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"American Journal of Orthopsychiatry , volume =","work_id":"a13d9467-45c8-4a44-92d4-3154a709e6ec","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":45,"snapshot_sha256":"c3744ef261c58a33edfd5f879d8f6cf4e618da8c39c4579f14ca74c5bc0dcac0","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"527aed6884eedfcc15e216248218a7add859675edd997e3bec68b29faf27fc6f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}