AI as an Accessibility Tool: What Works and What Doesn't

A genuine turning point — with caveats

Artificial intelligence is changing the experience of disability in ways that range from genuinely transformative to overhyped. Real-time captioning, image description, text simplification, and voice-controlled interfaces have improved significantly in recent years, and for many people with disabilities, AI-powered tools have materially changed what’s possible in daily life.

But AI is also being applied to accessibility in ways that don’t work — or that create new exclusion while claiming to solve it. Accessibility overlays promise AI-powered instant compliance and deliver neither. Training data biases mean that some AI systems perform significantly worse for users with disabilities than they do for the broader population. And the cost of high-quality AI accessibility tools remains a barrier for the people who need them most.

An honest look at AI and disability means acknowledging both.

What AI genuinely enables

Real-time captioning

Automatic speech recognition (ASR) has improved dramatically. AI-powered real-time captioning — through systems like Google Live Captions, Microsoft Azure Cognitive Services, and purpose-built tools like CART (Communication Access Realtime Translation) augmented by AI — can now produce captions accurate enough to be genuinely useful in live settings: meetings, lectures, events.

For deaf and hard-of-hearing users, this represents a meaningful practical improvement. A decade ago, accurate real-time captions required a human stenographer. Today, AI systems can produce captions with low enough error rates to support effective communication in most scenarios — though human review remains important for high-stakes contexts.

Accuracy varies by accent, background noise, technical vocabulary, and speaking speed. Users with non-standard accents (which includes many non-native English speakers) often find current ASR systems significantly less accurate for them than for native speakers — a bias baked into training data that remains an active problem.

Image description and alt text generation

Computer vision systems can now generate descriptions of images with enough accuracy to be useful. Tools like Microsoft Seeing AI, Google Lookout, and the image description features in iOS and Android describe photographs, documents, and scenes in real time for blind and low-vision users.

For websites, AI-generated alt text represents a meaningful improvement over the common alternative (no alt text at all). Social media platforms have deployed AI alt text generation at scale, and while the descriptions are often generic, they are better than silence.

The limitation: AI-generated alt text cannot reliably convey the meaning of an image in context. It can say “a group of people standing in front of a building” but not “the QualiBooth team at the 2025 Web Summit.” Human-written alt text that communicates why the image matters remains the gold standard. See our guide on writing accessible social media content for practical guidance.

Text simplification

Large language models can simplify complex text — government documents, legal terms, medical information — into plain language that’s accessible to people with cognitive disabilities, learning disabilities, or lower literacy levels. This application has genuine potential, particularly for content that organizations are legally required to publish but that is inherently complex.

Challenges: AI simplification can introduce errors. Simplified output should be reviewed by humans. And the “correct” simplification varies significantly by user — what helps someone with dyslexia may differ from what helps someone with an intellectual disability.

Voice control and dictation

AI-powered voice control has become a practical alternative to keyboard and mouse for many users with motor disabilities. Dragon NaturallySpeaking, Windows Speech Recognition, and iOS/Android voice control use machine learning to improve accuracy for individual users over time.

The quality improvement in these systems over the past decade is substantial. For users with conditions like multiple sclerosis, ALS, or severe arthritis, voice control has enabled independence that was significantly harder to achieve with earlier technology.

The gap: voice control accuracy still varies significantly by microphone quality, background noise, and speech patterns. Users with dysarthria (which affects many people with conditions like cerebral palsy or stroke) may find current voice recognition systems less effective than they are for clear, typical speech — a persistent accessibility gap within an accessibility tool.

Where AI falls short

Bias in training data

AI systems learn from data. When the data used to train a system underrepresents people with disabilities — or represents them primarily in pathological or stereotypical contexts — the resulting system may perform worse for users with disabilities, misinterpret disability-related content, or reinforce harmful assumptions.

This is not a theoretical concern. Studies have shown that ASR systems are less accurate for speakers with dysarthria or atypical speech patterns. Facial recognition systems perform less accurately for users who don’t fit demographic norms in training data. Hiring screening algorithms trained on historical data may disadvantage disabled applicants.

Developing AI that works well for users with disabilities requires including people with disabilities in the design and testing process — not as an afterthought, but as participants from the earliest stages.

Accessibility overlays: AI washing

AI-powered accessibility overlays claim to automatically detect and fix accessibility issues on websites. They are promoted aggressively and often claimed to produce instant compliance with WCAG or ADA requirements.

They don’t work. Overlays cannot fix the underlying code barriers that cause accessibility failures. They frequently interfere with the assistive technology users already rely on — screen readers behave unexpectedly, keyboard navigation breaks, the overlay itself introduces new barriers. And they have appeared in a growing number of accessibility lawsuits as evidence that the defendant was aware of accessibility issues but chose a non-solution over real remediation. We cover this in detail in our article on true digital accessibility.

Affordability and access

The most effective AI accessibility tools are often expensive, hardware-dependent, or subscription-based. Users with disabilities who already face higher costs for equipment, healthcare, and assistive technology often cannot afford the premium tiers of AI services that offer the best accessibility support.

Universal design — building accessibility into mainstream products rather than relying on users to purchase specialized tools — remains the most equitable approach.

What this means for how we design

The most important implication of AI’s role in disability is not that AI tools will solve the accessibility problem for us. They won’t. The baseline of accessible design — semantic HTML, keyboard operability, proper labeling, sufficient contrast — remains the foundation that both human users and AI tools depend on.

What AI changes is the support available to users on top of that foundation. When the foundation is solid, AI-powered tools can extend and amplify what’s possible. When the foundation is broken, AI tools are trying to compensate for something that should have been built correctly.

Build accessible digital products. Then AI assistive technologies can actually help the people who use them.

If you want to understand where your current digital products stand on that foundation, run a free scan or talk to our team about a comprehensive accessibility review.