Researchers Use Generative Models to Create Synthetic User Data for Inclusive Testing

AI · 6 min read

Inclusive testing often requires data from users with diverse assistive technologies and atypical interaction patterns, which can be scarce due to privacy and recruitment barriers. The research team trained generative models on consented, anonymized sessions to synthesize interactions such as switch-device navigation, screen-reader gestures, and dwell-based pointing.

The synthetic sessions preserve behavioral characteristics without containing personally identifiable information. Design system teams can inject these sessions into automated UI test suites to validate component behavior under real-world assistive interaction patterns, revealing failures like missing focus outlines or unreachable controls.

The paper reports that adding synthetic edge-case sessions increased detection of accessibility regressions by 38% in a sample of corporate repositories. Tooling includes a browser-based replay harness and a labelling assistant that maps synthetic events to accessibility compliance checks.

Researchers caution that synthetic data complements — but does not replace — real user testing. They encourage teams to use synthetic datasets for broader automated coverage, while maintaining partnerships with disability communities for qualitative validation.