LLM-powered browser tool fixes web accessibility issues in real time while blocking harmful changes

A new research paper describes a browser-based system that deploys large language models to automatically repair web accessibility failures as users browse, while safeguarding against unintended harm.

The Problem#

Website accessibility gaps persist because site developers often leave WCAG violations unfixed. Traditional approaches rely on authors to implement improvements, leaving disabled users without recourse. The researchers saw an opportunity: if AI agents could modify pages locally on a user's device, they could sidestep this bottleneck and address failures in real time.

How It Works#

The team built a Chrome extension that operates entirely on the user's machine, preserving privacy. The agent:

• Extracts and condenses a page's style sheets to fit within a language model's context window
• Queries the model for additive CSS rules targeting 18 WCAG metrics and W3C cognitive accessibility standards
• Injects the changes reversibly into the live page

They tested six open-weight models (ranging from 7B to 14B parameters) across twenty real websites—ten with significant accessibility violations, ten already accessible.

The Surprise: Unverified Changes Break as Often as They Fix#

Early results proved sobering. Without verification, the models improved pages and introduced regressions at roughly equal rates. Across 100 trials, five models that produced valid CSS achieved 24 improvements but also caused 20 regressions. The AI excelled at typography tweaks but often broke perception-dependent styles, trading one problem for another.

The Solution: Verified Repair Protocol#

To solve this, the researchers introduced a verified repair instrument using an audit-inject-verify loop:

• The system proposes a CSS change
• It applies the change and runs automated accessibility checks
• It accepts the modification only if violations strictly decrease
• By design, regression is impossible—the system rejects any change that doesn't measurably improve the page

Testing on a seeded-violation benchmark with real browser execution, the instrument detected 57 of 57 introduced violations with zero false positives and rejected all 126 adversarially harmful candidates.