Generative AI Accessibility Barriers: How ChatGPT & Copilot Exclude Users

Generative AI Is Creating a New Digital Divide

Generative AI tools like ChatGPT, Microsoft Copilot, and Google Gemini have reshaped how millions work, learn, and communicate. For many disabled users, however, these tools have introduced a new wave of digital exclusion — inaccessible chat interfaces, cognitive overload, and AI-generated images with no alt text.

Inaccessible Chat Interfaces and WCAG Failures

Screen reader users frequently encounter unlabeled input fields, missing live region announcements for streaming responses, and focus management failures after submitting a query. ChatGPT's web interface has been documented by the blindness community as requiring workarounds to navigate conversation history. Many AI interfaces fail WCAG 2.2 Level AA criteria out of the box.

Cognitive Accessibility Problems in AI Interfaces

• No plain language mode: AI responses default to variable-complexity prose with no option to request simplified language
• Streaming text output: tokens appearing character-by-character challenge users with reading disabilities or attention disorders
• Infinite conversation history: long threads with no semantic structure create orientation challenges
• Unpredictable response length: users cannot anticipate how long a response will be
• Session state loss: many AI tools do not preserve conversation history, a barrier for users relying on previous context

What AI Developers Must Do to Reduce Accessibility Barriers

• Implement ARIA live regions with configurable streaming — users should opt into full-response mode
• Automatically generate accessibility descriptions for every AI-generated image
• Provide a reading level control so users can request plain language without constructing that instruction every time
• Include skip navigation links and ARIA landmarks in all AI chat interfaces
• Test with real users of assistive technology throughout the product development process
• Publish annual accessibility conformance reports based on VPAT 2.5 for all AI products