GovAI Audit Network forms to build federated models for automated accessibility assessments

AI · 5 min read

GovAI's federated approach lets agencies train on local data—real web pages and app interactions—without centralizing sensitive information. Participating organizations contribute model weight updates rather than raw content, and privacy-preserving techniques like differential privacy and secure aggregation are used to prevent reconstruction of original pages.

The consortium is prioritizing models that flag semantic issues (missing landmark roles, form-label mismatches), interaction regressions (keyboard traps, unreachable controls), and probabilistic recommendations (likely contrast failures and insufficient focus styling). Because the models are federated, they better reflect the diversity of government services and assistive-technology usage patterns than a single centralized dataset could.

Open governance is a core promise: public documentation, reproducible evaluation datasets, and a transparent upgrade cadence. If successful, the network's models could be integrated into procurement tooling so agencies can automatically scan vendors' sites and apps for accessibility risks during procurement reviews.