AI and Accessibility: Promise, Limitations, and the Importance of Human Expertise

However, a significant risk remains: organizations may over-rely on AI-generated outputs, bypassing critical human verification.

Ultimately, AI serves as a valuable starting point rather than an authoritative source of accessibility expertise.

The Danger of Overtrusting AI

Modern AI systems are designed for responsiveness, often generating confident answers even when underlying data is uncertain or incomplete. This creates a false sense of security, particularly for those new to the field.

However, accessibility is a complex field that requires understanding user needs, assistive technologies, standards, human behavior, and context. These are areas where AI frequently struggles because accessibility cannot always be reduced to simple rules or patterns.

AI Learns from an Imperfect Web

AI models are constrained by their training data. Much of that training data comes from publicly available websites, documentation, code repositories, blogs, and tutorials. Unfortunately, the web contains a vast amount of inaccessible content, meaning AI often learns from and reproduces poor accessibility practices.

This raises an important question: if AI is learning from inaccessible examples, how can it consistently produce accessible solutions?

Chatbots frequently generate accessibility recommendations based on patterns they have observed. If those patterns are flawed, incomplete, or outdated, the resulting guidance may also be flawed. AI may recommend inaccessible components, misuse ARIA attributes, suggest keyboard interactions that do not work correctly, or implement patterns that fail when tested with assistive technologies.

Without specialized training, AI cannot distinguish between effective accessibility and flawed workarounds. This limitation is equally evident in automated testing.

Accessibility Testing: Automation Is Not Enough

Automated accessibility testing tools provide another example of both the strengths and limitations of AI and automation. Industry experience has consistently shown that automated testing can identify only a portion of accessibility issues. Many barriers, particularly those involving usability, cognitive load, content clarity, and assistive technology interactions, require manual assessment.

While scanners effectively identify binary issues like missing alt-text or labels, they lack the nuance required for complex evaluation. In one accessibility review I participated in, an automated testing tool reported approximately 150 accessibility issues. After manual verification, only 12 were found to be genuine accessibility problems, and only 6 had been reported accurately.

This experience highlights a recurring reality in automation: scanners can generate a high volume of noise, often producing false positives or misleading reports. Consequently, teams may spend hours debugging non-existent issues while missing genuine barriers. Key takeaways include:

• Automated tools can produce a large number of false positives.
• Reported issues may be technically correct but inaccurately described.
• Some findings may be duplicates or low-priority concerns.
• Human expertise is required to determine whether an issue truly affects accessibility.

If a development team were to blindly trust the automated report, they could spend significant time fixing problems that do not exist while potentially missing barriers that genuinely affect users.

Accessibility Requires Human Judgement

Accessibility isn't just a code-compliance check; it’s an empathy-driven design practice. Automated tools can flag missing alt text, but they cannot assess whether that text provides a meaningful experience. True accessibility requires understanding user intent, cognitive load, and the nuances of assistive technology—elements that defy simple pattern matching.

Similarly, AI can generate code that appears accessible because it follows certain patterns. Yet true accessibility often requires understanding:

• User intent
• Cognitive load
• Keyboard interaction expectations
• Screen reader behaviour
• Real-world user workflows
• Differences between assistive technologies

An accessibility expert may discover issues that no automated tool identifies, such as confusing navigation, unclear instructions, poor reading order, or interactions that become difficult for users with cognitive disabilities.

AI as a Partner, Not a Replacement

Rather than replacing experts, AI should be viewed as a high-speed partner. It can accelerate work by drafting code, explaining standards, and identifying potential problem areas for human review.

• Generating initial code drafts
• Suggesting accessibility improvements
• Explaining standards and guidelines
• Identifying potential problem areas
• Creating testing checklists
• Helping developers learn accessibility concepts

When knowledgeable accessibility professionals are involved, they can validate AI recommendations, identify mistakes, and provide better prompts and instructions. This creates a feedback loop where the AI becomes more useful because the human expert directs it toward better outcomes.

In practice, the most successful approach is often a combination of:

1. AI-generated assistance
2. Automated accessibility testing
3. Manual accessibility review
4. Testing with assistive technologies
5. Input from people with disabilities

Each layer compensates for the weaknesses of the others.

Looking Ahead

AI will undoubtedly improve over time. Accessibility awareness is increasing, training data is becoming more sophisticated, and newer tools are beginning to incorporate accessibility-specific knowledge. However, we are not yet at the point where accessibility is simply "built in" to AI-generated solutions.

As AI evolves, speed must not be confused with accuracy. Accessible-looking code is not always truly accessible, and human oversight remains the cornerstone of inclusive design.

The future of accessible technology will likely involve AI playing an important supporting role. Yet for the foreseeable future, accessibility expertise remains essential.

Conclusion

AI is a powerful starting point, not an authority. As long as its training data reflects an imperfect, often inaccessible web, AI-generated solutions will inherit many of those same flaws. Until accessibility is fundamentally baked into the systems that train and govern AI, success will depend on combining automation with human expertise. AI can accelerate accessibility efforts, but it cannot replace the knowledge, judgement, and lived experience required to create truly inclusive experiences.