AI in Accessibility: Augmentation or Replacement?

You're facing a critical decision that will shape your accessibility program: how should AI fit into your conformance testing process? The question isn't whether to use AI, but how to integrate it without undermining the human expertise that keeps your program robust.

The stakes go beyond efficiency. A study in The Lancet Gastroenterology & Hepatology found that experienced doctors detected fewer precancerous growths after introducing AI-assisted colonoscopy. Detection rates without AI assistance fell from 28.4% to 22.4%. The tool didn't fail; the humans did because they stopped practicing the judgment task the tool Web Accessibility Specialist meant to support.

WCAG conformance testing is also a judgment task. Your auditors must interpret Success Criteria 1.3.1 (Info and Relationships), distinguish meaningful images from decorative ones under 1.1.1 (Non-text Content), and determine whether heading structures serve their intended purpose. If AI makes those calls first, your team stops practicing the skill. When the tool misses something, and it will, they won't catch it either.

The Decision You're Facing

You need to choose where AI sits in your conformance testing workflow:

Path A: AI conducts the initial scan, flags potential violations, human reviewers validate the findings.
Path B: Human auditors conduct the primary review, AI performs a secondary check for missed issues.
Path C: AI remediates code directly, human experts review and approve changes before deployment.

Each path affects skill retention differently and creates different liability exposures.

Key Factors That Affect Your Choice

Current team expertise level: If you're building a new accessibility function, Path A might seem efficient. However, you're outsourcing judgment to a model trained on web content that includes significant accessibility failures.

Regulatory exposure: Organizations subject to Section 508 of the Rehabilitation Act or the DOJ Final Rule (2024) need defensible conformance claims. "The AI said it passed" won't satisfy an Office for Civil Rights investigation. You need auditors who can articulate why a specific implementation meets or fails a Success Criterion.

Content complexity: AI-generated code can fail WCAG in novel ways. If your content involves dynamic interfaces, complex data visualizations, or custom components, AI lacks the context to judge whether semantic relationships are preserved.

Consensus on failure criteria: Your team must agree on what constitutes a violation before AI can reliably flag one. If your auditors debate whether a particular heading structure fails 1.3.1, the AI won't resolve that debate; it'll reflect the inconsistency in its training data.

Path A: AI-First Scanning

Choose this path when:

• You're conducting high-volume assessments of similar page templates.
• You have clear, documented standards for common violations.
• Your review team has strong WCAG expertise and actively challenges AI findings.
• You're testing static content with predictable patterns.

Governance requirements:

• Document which Success Criteria the AI is trained to evaluate.
• Maintain a human review sample of Assistive Technology least 20% of flagged items.
• Track disagreement rates between AI and human reviewers.
• Require auditors to manually test Assistive Technology least two full pages per assessment without AI assistance.

Risk: Your auditors may stop questioning the tool, validating findings instead of conducting independent analysis. When the AI misses a violation, particularly novel failures in dynamic content, they might miss it too.

Path B: Human-First Review

Choose this path when:

• You're building internal accessibility expertise.
• You're subject to federal accessibility requirements under Section 508.
• Your content includes custom components or complex interactions.
• You need to defend conformance claims in legal proceedings.

Implementation:

• Auditors complete manual testing against WCAG 2.1 Level AA (or 2.2, depending on your compliance target).
• AI performs a secondary scan after human review is documented.
• Discrepancies trigger team discussion and documentation.
• AI findings that humans missed become training cases.

Risk: Even with AI in a secondary role, humans might rely on it as a safety net, conducting less thorough initial reviews. The Lancet study found skill atrophy occurred even when AI served as a backup.

Mitigation: Rotate which team members receive AI backup. Some audits should proceed without AI review to maintain independent judgment.

Path C: AI Remediation with Expert Approval

Choose this path when:

• You're remediating legacy content Assistive Technology scale.
• You have clear patterns for fixes (color contrast, alt text for standard icons).
• You can afford to review every AI-generated change before deployment.
• You're willing to reject AI suggestions that don't meet your standards.

Critical control: The human reviewer must mark when expert review occurred. If your remediation tool applies fixes automatically without a documented approval step, you've lost the oversight that prevents deskilling.

Workflow:

1. AI proposes remediation for flagged violations.
2. Accessibility specialist reviews proposed code changes.
3. Specialist approves, modifies, or rejects each change.
4. System logs which changes received human review.
5. Approved patterns become templates for similar future violations.

Risk: Volume pressure. When remediating thousands of pages, the temptation to batch-approve AI suggestions without individual review becomes overwhelming. That's when novel failures slip through.

Summary Matrix

Factor	Path A (AI-First)	Path B (Human-First)	Path C (AI Remediation)
Best for	High-volume template scanning	Building defensible expertise	Legacy content remediation
Skill preservation	Low - auditors validate, don't analyze	High - AI challenges human work	Medium - depends on review rigor
Regulatory defense	Weak without strong human oversight	Strong - human judgment documented	Medium - requires approval documentation
Scalability	High	Low to medium	High if approval workflow is enforced
Novel failure detection	Poor - AI misses new patterns	Good - human pattern recognition	Poor - AI repeats learned fixes

The question isn't whether AI makes you faster. It does. The question is whether your team will retain the ability to question its output when the tool gets something wrong, and whether you can prove that ability when a regulator asks.

If you can't describe your human oversight controls in specific terms, you're on Path A whether you intended to be or not.