AI-Generated Web Interfaces: New Benchmark for Accessibility

Imagine asking a large language model to create a dropdown menu for a website. It might generate code that looks perfect on screen. The visual result could seem flawless, with neat borders and smooth animations. But what happens when someone tries to navigate that menu using only their keyboard? Or when a screen reader used by a blind person attempts to describe its options? The code might be silent, creating a dead end for millions of users.

This gap between visual appearance and functional accessibility is a critical flaw in today’s rush toward AI-assisted coding. A new pilot study puts a precise number on the problem and offers a potential path forward. Published by software engineer Harsha Vardhan Reddy Nooka, the research introduces AXIS-Bench, a targeted tool for measuring the real-world usability of AI-generated front-end code.

AXIS-Bench is not about checking if a button looks like a button. It asks if a button acts like one. The benchmark defines 12 specific interactive components that are common yet complex, such as:

• Modal dialogs
• Tabbed interfaces
• Sortable data tables
• Dropdown menus
• Sliders
• Accordions

It then runs automated tests to see if the code produced by an AI meets core accessibility standards, primarily the Web Content Accessibility Guidelines (WCAG) and the WAI-ARIA specification for defining roles, states, and properties.

The pilot findings are stark. When researchers fed component specifications to a large language model without any special guidance, its generated code passed only six of the twelve interactivity and conformance tasks. A non-semantic baseline, essentially code that creates the right shapes but none of the underlying functionality, passed zero. The crucial insight came from a third test condition. When the same AI model was supplied with explicit guidance on accessibility patterns before generating code, it passed all twelve tasks.

This does not mean the problem is solved. The study authors are clear that these results are limited to their pilot benchmark. Real-world applications are messier, with more edge cases and unique requirements. However, the jump from zero to twelve passes with the right guidance is impossible to ignore. It points to a fundamental truth about current AI tools. They are powerful pattern matchers but they lack inherent understanding of human need. Without explicit, structured prompts that define functional outcomes like keyboard navigation, the models default to replicating visual patterns alone.

For developers, this research underscores a shift in mindset that is already underway. Prompt engineering for code generation is moving beyond describing a layout. It now must include describing the experience. You are not just asking for a sortable table. You are asking for a sortable table where column headers are focusable with the Tab key, announce their sortable state to screen readers, and update the ARIA aria-sort attribute on click. This level of specificity bridges the gap between AI capability and human requirement.

The broader implication is for the software development lifecycle itself. As AI becomes a common co-pilot, quality assurance must evolve. Automated checks for visual regression are standard. This benchmark suggests we need automated checks for accessibility and interactivity regression just as deeply integrated. Testing that a feature works is no longer sufficient. You must test how it works and for whom.

Harsha Vardhan Reddy Nooka’s professional background at Amazon, working on platforms like the AWS Management Console, informs this practical approach. The console is a complex tool used by developers globally where clarity and reliability are non-negotiable. This experience translates into a research focus on dependability, asking not just if code can be generated but if it can be trusted.

For the technology industry, AXIS-Bench is a call for better metrics. Celebrating the speed or volume of AI-generated code is a hollow victory if that code locks people out. The benchmark provides a tangible if preliminary way to measure something more important: inclusion. It moves the conversation from whether AI can build an interface to whether it can build one that works for everyone.

The path forward is clear. Developers and companies investing in AI coding tools must demand and build evaluation frameworks that test for real-world use. Investors should ask not just about lines of code generated per hour but about compliance with WCAG success criteria. The research shows that with the right guardrails, AI can be guided toward more equitable outcomes. The responsibility now lies with us to insist on those guardrails, making accessibility a core prompt in the genesis of every new digital interface.