What AI Can (and Can't) Fix in Accessibility

AI can accelerate accessibility work, but it cannot replace the fundamentals of accessible code. Machine learning excels at pattern recognition tasks like generating alt text descriptions, detecting low-contrast text, and identifying missing labels at scale. It struggles—and often fails—at tasks requiring contextual understanding: determining whether alt text is actually useful in context, fixing keyboard navigation logic, or restructuring a page's semantic hierarchy.

This distinction matters because AI accessibility tools are often marketed with implicit promises of "automatic compliance." The reality is more nuanced. According to W3C guidance on evaluation tools, automated testing—AI-powered or otherwise—can only reliably detect 30-40% of WCAG issues. The remainder require human judgment. AI augments human effort; it doesn't eliminate it.

Understanding what AI can reliably do helps organizations deploy it effectively: use AI for scale and speed on appropriate tasks while preserving human expertise for judgment-dependent work. This approach yields actual accessibility improvements rather than false confidence.

Key Takeaways

AI is a powerful tool for accessibility—when applied to tasks it can actually perform well.

• Alt text generation works for simple images – Computer vision can describe objects and scenes; it can't determine what's relevant in your specific context
• Pattern detection scales beyond rules – Machine learning identifies likely issues that rigid rule-based tools miss, expanding automated coverage
• Captions are mostly reliable – Auto-captioning achieves 85-95% accuracy for clear audio; professional editing is still needed for published content
• Structural fixes require code changes – AI can't reorder your DOM, add proper keyboard handlers, or implement focus management without source code access
• Context is AI's fundamental limit – AI doesn't know why an image matters to your page or whether your interaction pattern makes sense

What AI Does Well

AI-powered accessibility tools genuinely advance the field in specific areas where the task aligns with machine learning's strengths.

Alt Text Generation for Simple Images

Computer vision models can analyze images and produce descriptions. For straightforward images, these descriptions are useful starting points:

+-----------------------+--------------------+--------------------------------------------------+
|       Image Type      |   AI Performance   |                  Typical Output                  |
+-----------------------+--------------------+--------------------------------------------------+
|     Product photo     |        Good        |     "Blue running shoe on white background"      |
+-----------------------+--------------------+--------------------------------------------------+
|      Person photo     |      Moderate      |    "Person smiling, wearing business attire"     |
+-----------------------+--------------------+--------------------------------------------------+
|       Landscape       |        Good        |         "Mountain lake with pine trees"          |
+-----------------------+--------------------+--------------------------------------------------+
|      Chart/graph      |        Poor        |   "Chart with blue and red bars" (misses data)   |
+-----------------------+--------------------+--------------------------------------------------+
|      Infographic      |        Poor        |    May describe some elements; misses meaning    |
+-----------------------+--------------------+--------------------------------------------------+
|   Abstract/artistic   |      Variable      |     Describes visual elements; misses intent     |
+-----------------------+--------------------+--------------------------------------------------+

For e-commerce sites with thousands of product images lacking alt text, AI can generate baseline descriptions that are better than "IMG_1234.jpg" or nothing at all. This addresses a real problem: WebAIM's 2024 data shows 54.5% of home pages have images with missing alt text.

Video Caption Generation

Auto-captioning technology has matured significantly. Modern systems achieve:

• 85-95% accuracy for clear, single-speaker audio in common languages
• Lower accuracy for multiple speakers, accents, technical terminology, or background noise
• No speaker identification by default (accessibility requires identifying who's speaking)
• No non-speech audio description (captions like "[phone rings]" require human addition)

For compliance purposes, auto-captions provide a starting point that humans can edit, reducing the time from hours to minutes. However, publishing unreviewed auto-captions for critical content risks errors that undermine the accessibility benefit.

Pattern Detection Beyond Static Rules

Traditional automated accessibility tools use rule-based detection: "if image lacks alt attribute, flag it." AI can identify patterns that rules don't capture:

• Likely buttons that aren't buttons – AI can identify `<div>` elements that behave as buttons based on styling and click handlers
• Focus order anomalies – Machine learning can predict when visual and DOM order diverge in confusing ways
• Text-in-image detection – AI can identify images containing text that should be in HTML instead
• Reading order issues – Detecting when multi-column layouts may cause confusing screen reader order

These capabilities extend automated detection beyond the ~30% that rules-based tools catch, though the ceiling remains limited by what automation can perceive.

Scale Analysis

AI's primary value proposition is scale. Human auditors can thoroughly evaluate dozens of pages. AI can scan thousands of pages and:

• Identify pages with unusually high error density
• Detect patterns across the site (e.g., "all product pages missing alt text")
• Prioritize remediation by issue frequency and severity
• Track trends over time

This scale matters for large sites where comprehensive manual auditing is economically impractical.

What AI Cannot Fix

AI has fundamental limitations that no amount of model improvement will fully overcome.

Context-Dependent Judgments

AI lacks the context to make accessibility judgments that depend on meaning:

Alt text quality example:

A product page shows a sweater. AI generates: "Gray knit sweater on wooden hanger."

But users need: "Women's cashmere crew-neck sweater in heather gray, shown on hanger to display length and drape."

AI describes what it sees. Accessibility requires describing what matters—which depends on page purpose, user goals, and content context.

Heading structure example:

AI can detect that headings skip levels (H2 to H4). It can't determine:

• Whether the heading text accurately describes the section
• Whether the section should exist at all
• Whether the hierarchy reflects logical content organization

These judgments require understanding the content's meaning, which AI doesn't have.

Keyboard Navigation Logic

AI cannot fix keyboard navigation issues because fixing them requires code changes:

• Custom focus management – Modals must trap focus inside, then restore focus on close. This requires JavaScript event handling.
• Arrow key navigation – Tab panels, menu bars, and tree views require arrow key patterns per ARIA Authoring Practices. Each widget type has specific expected behavior.
• Focus indicators – Visible focus requires CSS changes. Ensuring visibility against all backgrounds requires design decisions.
• Keyboard traps – Escaping a trap requires understanding how the component should behave and modifying its code.

An AI can detect that a modal doesn't trap focus. It cannot write the JavaScript to trap focus, test that it works, and ensure it handles edge cases. That requires engineering.

DOM Structure and Semantics

The Document Object Model determines what assistive technology can perceive. AI can identify issues but can't restructure HTML:

+----------------------------------------+------------------------------+--------------------------------------------+
|                 Issue                  |         AI Detection         |             AI Fix Capability              |
+----------------------------------------+------------------------------+--------------------------------------------+
|   `<div>` used instead of `<button>`   |   Possible (from patterns)   |         Cannot change HTML element         |
+----------------------------------------+------------------------------+--------------------------------------------+
|           Missing landmarks            |          Detectable          |   Cannot add `<main>`, `<nav>` to markup   |
+----------------------------------------+------------------------------+--------------------------------------------+
|       Improper heading hierarchy       |          Detectable          |    Cannot reorganize content structure     |
+----------------------------------------+------------------------------+--------------------------------------------+
|       Form fields without labels       |          Detectable          |       Cannot add `<label>` elements        |
+----------------------------------------+------------------------------+--------------------------------------------+
|         Tables without headers         |          Detectable          |         Cannot add `<th>` elements         |
+----------------------------------------+------------------------------+--------------------------------------------+

Source code remediation is required. AI can suggest what to fix; a human or automated system must change the code.

Complex Widget Interactions

Custom components—accordions, carousels, tab panels, date pickers, combo boxes—have specific accessibility requirements:

• Keyboard patterns (Tab, arrows, Enter, Escape, Home, End)
• ARIA roles, states, and properties
• Live region announcements for dynamic updates
• Focus management between states

AI cannot implement these patterns. It would need to:

1. Understand the widget's intended behavior
2. Analyze existing code
3. Write correct keyboard handlers
4. Add proper ARIA markup
5. Test with actual assistive technology

This is complex software engineering, not pattern matching.

User Experience Judgment

Some accessibility judgments are inherently human:

• Is this error message helpful? – AI can verify an error message exists; it can't judge whether the message helps users fix the problem
• Is this reading order logical? – AI can predict when order might be confusing; users determine what's actually confusing
• Is this interaction pattern intuitive? – Only testing with real users reveals whether AT users can figure out how a component works
• Is the content understandable? – WCAG has success criteria around reading level and clarity; these require human evaluation

These judgments inform whether a site is truly usable, not merely compliant.

AI as Augmentation, Not Replacement

The effective use of AI in accessibility follows a clear pattern: use AI for tasks it performs well, then apply human judgment for tasks it doesn't.

The Human-AI Workflow

A practical workflow might look like:

1. AI scans site – Identify all images lacking alt text, potential contrast issues, likely structural problems
2. AI generates suggestions – Produce alt text drafts, flag specific issues with recommendations
3. Human reviews critical content – Verify alt text for important images, check that suggestions make sense
4. Human implements structural fixes – Keyboard navigation, DOM changes, ARIA implementation
5. Human tests with AT – Verify that changes actually work with screen readers and keyboards
6. AI monitors for regression – Continuous scanning catches new issues as they're introduced

This workflow uses AI for scale (scanning thousands of pages) while preserving human judgment for quality.

Where AI Saves Time

+-------------------------------------------------+----------------------------+--------------------------------------+
|                       Task                      |         Without AI         |               With AI                |
+-------------------------------------------------+----------------------------+--------------------------------------+
|        Audit 10,000 pages for missing alt       |           Weeks            |                Hours                 |
+-------------------------------------------------+----------------------------+--------------------------------------+
|   Generate baseline alt text for 5,000 images   |            Days            |        Minutes (then editing)        |
+-------------------------------------------------+----------------------------+--------------------------------------+
|                Caption 50 videos                |         100+ hours         |   5 hours (auto-caption + editing)   |
+-------------------------------------------------+----------------------------+--------------------------------------+
|        Identify pages needing remediation       |      Manual sampling       |    Comprehensive prioritized list    |
+-------------------------------------------------+----------------------------+--------------------------------------+
|      Track accessibility metrics over time      |   Manual periodic audits   |        Continuous dashboards         |
+-------------------------------------------------+----------------------------+--------------------------------------+

The time savings are real—but only if the AI output is then reviewed and refined.

Where AI Costs Time

AI can create false efficiency:

• Reviewing false positives – AI pattern detection flags issues that aren't actually issues
• Fixing AI errors – Auto-generated alt text that's wrong creates worse experience than missing alt text
• False confidence – Teams that trust AI reports without verification ship inaccessible products
• Technical debt – AI-suggested fixes may not follow best practices, creating maintenance burden

Organizations should measure total time (AI processing + human review + error correction), not just AI processing time.

AI Capabilities Are Evolving

It's worth acknowledging that AI capabilities improve over time. Today's limitations may narrow tomorrow.

Near-Term Improvements

Likely improvements in current technology:

• Better alt text contextual understanding – Large language models with vision can consider surrounding text when describing images
• Code suggestion quality – AI coding assistants may better suggest accessibility fixes given the right prompts
• Cross-issue correlation – AI may better identify root causes that produce multiple symptoms

Persistent Limitations

Limitations that are fundamental, not just unsolved:

• Source code access – AI can't modify code it can't access; browser-side AI cannot edit server-side templates
• Intent understanding – AI doesn't know what your website is trying to accomplish or who your users are
• Assistive technology behavior – Real AT behavior varies across vendors and versions; AI can't predict all variations
• User experience – Whether disabled users can actually complete tasks requires testing with disabled users

These limitations persist because they're not pattern-matching problems; they're engineering and human factors problems.

Overlay Caution

AI-powered accessibility overlays promise to fix sites automatically via JavaScript. The National Federation of the Blind's 2021 resolution explicitly opposes these approaches. TestParty research based on Court Listener data shows over 1,000 overlay users were sued in 2023-2024.

The technical reason overlays fail illuminates AI's limits: overlays operate on the rendered page, not the source code. They can add attributes and inject content, but they can't restructure markup, implement proper keyboard handling, or ensure consistent behavior across browsers and AT combinations. See Why Accessibility Overlays Increase Legal Risk.

Practical Recommendations

For organizations evaluating AI accessibility tools:

Evaluate Claims Critically

When vendors claim "AI-powered compliance," ask:

• What percentage of WCAG success criteria does the tool address?
• How does the tool handle issues requiring code changes?
• What is the false positive/negative rate for detection?
• What validation has been done with actual assistive technology users?

Be skeptical of claims that AI eliminates the need for engineering work or human testing.

Use AI for Appropriate Tasks

Deploy AI where it excels:

• Large-scale scanning for issue identification and prioritization
• Alt text generation as a starting point for human review
• Caption generation for draft captions requiring editing
• Pattern detection to find issues rules-based tools miss

Preserve Human Judgment

Reserve human expertise for:

• Critical content review – Alt text for hero images, product photos, key infographics
• Keyboard and screen reader testing – Actual user flow verification
• Complex widget implementation – ARIA patterns, focus management
• Accessibility strategy – Prioritization, roadmapping, standards decisions

Integrate with Source Code

AI provides maximum value when integrated with source code remediation:

• AI identifies issues
• Humans (or automated systems) implement fixes in code
• Fixes are version-controlled and testable
• Regression scanning prevents re-introduction

This workflow produces durable improvements, not temporary patches.

FAQ

Can AI make a website fully WCAG compliant?

No. AI can help detect issues and generate content like alt text, but WCAG compliance requires code changes that AI cannot implement without access to source code. Even with source access, many success criteria require human judgment about content quality and interaction design. AI is a tool that accelerates compliance work; it doesn't automate it.

Is AI-generated alt text good enough for compliance?

It depends on context. For decorative or low-priority images, AI-generated alt text may be sufficient. For critical content—product images, infographics, images conveying information—human review is necessary. AI often describes what's visible without understanding what's relevant. A product image needs to describe what helps users make purchasing decisions, not just the visual contents.

How accurate are AI accessibility scanners compared to traditional tools?

AI scanners extend traditional rule-based detection by identifying patterns that rules miss. However, accuracy varies by vendor and issue type. The fundamental limitation remains: both AI and rule-based scanners can only detect issues that are observable in rendered output. Issues requiring user testing or contextual judgment remain undetectable by any automated approach. According to W3C guidance, all automated tools (AI or otherwise) catch 30-40% of issues.

Should we use AI accessibility tools instead of manual testing?

No. AI tools and manual testing serve different purposes. AI tools provide scale: scanning many pages quickly, identifying candidates for remediation, monitoring for regressions. Manual testing provides depth: verifying actual user experience with assistive technology, evaluating judgment-dependent criteria, testing complex interactions. Effective accessibility programs use both. See Accessibility Testing Tools: Manual vs Automated vs AI.

What's the difference between AI accessibility tools and overlays?

AI accessibility tools typically help identify and suggest fixes for accessibility issues, providing information for developers to act on. Overlays are JavaScript widgets that claim to automatically fix issues in the browser without code changes. The distinction matters: AI tools assist remediation; overlays attempt to bypass it. The FTC fined AccessiBe $1 million in January 2025 for misleading claims about overlay capabilities.

Will AI eventually be able to fix all accessibility issues?

Unlikely in the foreseeable future. Some accessibility problems are fundamentally engineering problems (implementing keyboard navigation requires writing code), design problems (choosing appropriate heading hierarchy requires content understanding), or user experience problems (determining whether an interaction is intuitive requires user testing). AI may improve at detecting and suggesting, but implementation and validation will require human involvement for the foreseeable future.

Related Resources

Internal Links

• Accessibility Testing Tools: Manual vs Automated vs AI
• Why Accessibility Overlays Increase Legal Risk
• Why Accessibility Is a Software Engineering Problem
• What Is Digital Accessibility? A Technical Definition
• Automated Accessibility Testing: What It Catches and What It Misses
• The Modern Accessibility Testing Stack

External Sources

• W3C WAI: Evaluating Web Accessibility
• WebAIM Million 2024 Report
• NFB 2021 Resolutions (Overlay Opposition)
• FTC AccessiBe Settlement (January 2025)
• W3C ARIA Authoring Practices
• DOJ Guidance on Web Accessibility

This article was written by TestParty's editorial team with AI assistance. All statistics and claims have been verified against primary sources. Last updated: January 2026.