Accessibility and Semantic Quality Regressions in AI-Assisted React Development: An Empirical Study

We present an empirical study of non-functional quality regressions in 
AI-assisted frontend development, focusing on accessibility and semantic 
correctness in React/TypeScript repositories. Using a difference-in-differences 
(DiD) design extended with Tobit regression and dynamic event study, we analyze 
axe-core accessibility violations and AST-derived complexity metrics across 
repositories stratified by AI tool adoption (Cursor, Copilot).

Key findings: axe violations show no statistically significant increase 
attributable to AI assistance (β = +0.005, p = 0.776); AST complexity shows 
a marginal positive effect (β = +0.005, p = 0.075) that is right-censored 
by repository-level refactoring cycles. Document structure violations 
constitute the largest violation category. Parallel trends assumptions hold 
across pre-treatment periods.

This work extends the methodology of recent causal inference approaches to 
AI-assisted development (cf. Kästner et al., MSR 2026) with frontend-specific 
outcomes and addresses the underexplored intersection of LLM code generation 
and web accessibility standards.

A preprint is also available on arXiv (cs.SE). The analysis code and panel 
dataset are available at: https://github.com/SomilKSharma/ai-react-accessibility-study