AI &
Accessibility
Artificial intelligence is simultaneously one of the most powerful tools for expanding access and one of the most common sources of new barriers. Understanding both sides is essential for anyone building or evaluating instructional content.
When applied deliberately, AI tools can significantly reduce barriers — automating tasks that would otherwise require manual effort and extending access to users with a wide range of needs.
Automated Captions
1.2.2 Captions (Prerecorded)Speech recognition generates real-time captions for videos and live events, enabling access for deaf and hard-of-hearing users — and anyone watching in a noisy environment.
Accuracy degrades with accents, technical vocabulary, and background noise. Always review and correct automated captions before publishing.
AI-Generated Image Descriptions
1.1.1 Non-text ContentVision models (GPT-4o, Gemini, etc.) can generate starting-point alt text for images, reducing the effort required to describe large image libraries.
AI descriptions reflect training data biases and can miss context, tone, and intent. Human review is required before publishing.
Voice Interfaces and Speech Input
2.1.1 KeyboardAI-powered voice assistants and speech-to-text tools reduce motor demands for users with limited hand mobility, enabling navigation and content creation through voice alone.
Voice interfaces are inaccessible to deaf users and those who cannot speak. They should complement, not replace, standard input methods.
Predictive Text and AAC
2.5.1 Pointer GesturesAugmentative and Alternative Communication (AAC) systems use AI prediction to reduce keystroke demands, enabling users with motor or speech impairments to communicate faster.
Prediction models trained on neurotypical language can underserve users with atypical syntax — particularly autistic users and those with aphasia.
Eye Tracking and Switch Access
2.5.6 Concurrent Input MechanismsMachine learning enables affordable gaze-based navigation and switch access interfaces, extending computer access to users with severe motor impairments.
Calibration requirements can be a barrier. Gaze systems require stable head position and lighting conditions that not all environments provide.
AI tools also create accessibility barriers — often invisibly. Understanding these failure modes is as important as knowing the benefits.
Inaccessible AI-Generated Content
LLMs produce content on demand — but rarely produce accessible HTML, proper heading hierarchy, labeled form inputs, or meaningful alt attributes by default. AI-generated interfaces commonly fail WCAG AA on first pass.
A developer asks an AI to build a form. The result has unlabeled inputs, no error states, and focus management that skips interactive elements.
Bias in Assistive Technology
Speech recognition systems have significantly higher error rates for speakers with accents, speech disabilities, or dialectal variation. Image captioning models reflect demographic bias in their training data.
Voice control systems fail to recognize commands from users with dysarthria at a significantly higher rate than for typical speakers — compounding barriers for the users who rely on them most.
Over-Reliance on Automation
Automated accessibility checkers catch 20–40% of WCAG failures — primarily mechanical checks like color contrast and missing alt attributes. Organizations treating passing scores as "accessible" miss the majority of real barriers.
A page passes an automated audit. Screen reader testing reveals the navigation landmark structure is broken, interactive components have no keyboard support, and error messages are visually present but not announced.
AI-Generated Images Without Descriptions
Generative image models produce images with no inherent descriptions. When embedded in content without alt attributes, entire visual concepts become inaccessible to screen reader users.
An instructional designer generates six custom illustrations and publishes them without alt text. Screen reader users encounter six consecutive empty image elements with no context.
Two practical AI tools for accessibility work. Both require your own Anthropic API key — your key is stored only in your browser and never transmitted to our servers.
These tools call the Anthropic API (Claude). Your key is stored only in your browser's localStorage and is never sent to our servers. Get an API key →
Paste a publicly accessible image URL. The AI produces a starting-point description — you review and edit before publishing. Alt text requires human judgment; this tool accelerates the first draft.
The systems designed to support disabled users are trained on data that systematically under-represents them. Understanding how bias manifests in assistive AI is critical for evaluating tools and advocating for more equitable systems.
Word error rates for Black American English speakers are approximately 2× higher than for white American English speakers in major commercial systems (Koenecke et al., 2020).
Voice control becomes unreliable for users who already face barriers with motor-based input — compounding existing access inequities.
Image captioning models trained on datasets with demographic skews assign higher-confidence descriptions to images containing white, male-presenting subjects.
Users relying on image descriptions for people-centric content receive less accurate context when images depict underrepresented groups.
Language models optimized for fluent text predict completions based on neurotypical communication patterns, diverging from the distinctive syntax many autistic or AAC-dependent users prefer.
Prediction systems interrupt rather than assist — forcing AAC users to override suggestions rather than accept them, increasing cognitive and motor load.
AI reading-level estimators calibrated against mainstream corpus data often score content written with familiar, idiomatic language as "complex" while missing actual cognitive barriers like undefined technical terms.
Instructional designers receive misleading signals: content flagged as advanced may be accessible, while accessible-scoring content may contain real cognitive barriers.
When selecting AI tools that affect accessibility, ask: Who was this trained on? Has it been tested with users with disabilities? What does the error rate look like across demographic groups? These questions rarely appear in marketing materials — but they determine whether the tool expands or narrows access.