Can AI Replace Human Accessibility Testing?

No. AI can help teams find possible accessibility issues, but it cannot replace a complete human-led WCAG audit. Some checks need an expert to understand the page, its purpose and how users interact with it.

For example, AI may find a button with a valid accessible name. But that button may open a dialog that is difficult to use with a keyboard.

AI may also suggest text for an image. An expert must check whether the image needs that text and whether the description explains its purpose.

The World Wide Web Consortium (W3C) explains that automated tools cannot check every part of web accessibility. Human judgement is still needed.

Source: W3C guidance on selecting accessibility evaluation tools .

What Does the October 2026 Research Show About AI Accessibility Testing?

A study published in Information and Software Technology tested how well Large Language Models, or LLMs, could evaluate web accessibility. LLMs are AI systems that work with language, code and other information.

The researchers tested seven LLMs using 384 short HTML test cases. The cases came from W3C Accessibility Conformance Testing (ACT) rules. The evaluation covered 39 WCAG Level A and AA success criteria.

The best model setup reached 71.52% accuracy on the selected test cases. Giving models examples improved average accuracy by about 8.2 percentage points.

Yet the models often struggled to decide whether a test rule applied to the content. The study reported applicability error rates of about 45% to 67% in the evaluated settings.

Why Do These Findings Matter?

A tool must first decide whether a rule applies. If it gets this step wrong, the final result may also be wrong.

This can lead to false alarms. It can also cause teams to miss real barriers.

The study also compared LLMs with standard automated tools. The systems did not always respond in the same way. Some standard tools could report that they could not evaluate a test. The LLMs returned a classification for every test case.

This difference matters when comparing the reported accuracy.

Research: An empirical investigation of Large Language Models for automated software compliance testing: Evidence from web accessibility .

What Can AI Accessibility Testing Tools Do?

AI tools can support many parts of an accessibility testing project. They are most useful when teams give them clear tasks and check their results.

These features may save time. They do not prove that a website or app meets WCAG requirements.

What Can AI Accessibility Testing Miss?

AI tools may miss barriers when they lack the full page, user task or interaction state. They may also make mistakes when applying WCAG rules.

1. The Purpose of an Image

AI may report an empty alternative text value as an issue. But an empty alt value can be correct for a decorative image.

2. Keyboard Interaction

AI may review a button’s code without testing what happens when a user activates it. The next step may contain a keyboard barrier.

3. Screen Reader Behaviour

Valid HTML does not always mean that a screen reader user can complete a task. Testing may need NVDA, JAWS, VoiceOver or TalkBack.

4. Dynamic Page Content

Some content appears only after a user clicks, scrolls, signs in or submits a form. A scan of the first page state may miss these changes.

5. Form Errors

AI may find an error message but fail to check whether the message is linked to the correct input.

6. Complete User Tasks

A website may pass many automated checks but still block a user from finishing a payment, booking or application.

7. Incorrect WCAG Mapping

AI may link an issue to the wrong success criterion. Experts must check the rule, its scope and any exceptions.

AI vs Automated vs Manual Accessibility Testing

These three methods have different strengths. The best approach uses them together.

Manual testing can also include people with disabilities. Their feedback can reveal barriers that technical checks alone may not show.

Five Real-World WCAG Testing Examples

Example 1: An Image With Empty Alt Text

AI finding: The image has an empty alt attribute. AI marks it as a failure.

Human review: The tester checks the image’s purpose. If it is decorative, an empty alt value may be correct.

Relevant WCAG criterion: 1.1.1 Non-text Content.

Example 2: A Dialog Opens

AI finding: The button has a name and works with a mouse.

Human review: The tester opens the dialog with a keyboard. They check focus movement, keyboard access and how the dialog closes.

Relevant WCAG criteria may include: 2.1.1 Keyboard, 2.4.3 Focus Order and 4.1.2 Name, Role, Value. The exact mapping depends on the verified failure.

Example 3: A Custom Dropdown

AI finding: The dropdown uses ARIA attributes and appears valid.

Human review: The tester checks the accessible name, role, current state and keyboard behaviour. They also check screen reader output.

Relevant WCAG criteria may include: 2.1.1 Keyboard and 4.1.2 Name, Role, Value.

Example 4: A Form Shows an Error

AI finding: An error message appears near a form field.

Human review: The tester submits invalid data and checks whether the error is clear, linked to the field and available to assistive technology.

Relevant WCAG criteria may include: 3.3.1 Error Identification, 3.3.2 Labels or Instructions and 4.1.3 Status Messages. Each criterion must be assessed separately.

Example 5: An Online Payment

AI finding: Each page has few automated issues.

Human review: The tester signs in, selects a service, enters payment details and checks the final result.

The tester also checks loading states, errors, time limits and confirmation messages.

A complete task can fail even when individual pages appear to pass automated checks.

How Should Teams Validate AI Accessibility Findings?

Teams need a clear process to turn possible issues into trusted audit findings. The process should record evidence at every stage.

1. Define the testing scope. List the pages, user roles, workflows and WCAG version.
2. Run automated checks. Use tested accessibility rules for machine-checkable issues.
3. Review AI suggestions. Treat AI output as possible findings, not confirmed failures.
4. Check rule applicability. Confirm that the WCAG requirement applies to the content.
5. Test actual behaviour. Use a keyboard, screen readers and other tools as needed.
6. Capture clear evidence. Record the page, element, steps, result and test setup.
7. Confirm WCAG mapping. Match each failure to the correct success criterion.
8. Provide fix guidance. Explain what happened, what should happen and how to fix the issue.
9. Retest the fix. Check the updated product and confirm whether the barrier remains.
10. Prepare the final report. Include verified findings, limitations and unresolved issues.

How Should AI Accessibility Test Results Be Classified?

A testing system should not force every result into a simple pass or fail. Some issues need more review. Others may not apply.

This table describes a practical reporting workflow. It is not a replacement for the formal ACT outcome terms.

W3C ACT rules use five defined outcomes: passed, failed, inapplicable, cantTell and untested.

An ACT rule result is not the same as a full WCAG conformance decision. A passed rule may still need other tests before the related WCAG criterion can be judged.

Source: W3C ACT Rules Format .

Can AI Generate a Reliable VPAT/ACR?

AI can help draft parts of an Accessibility Conformance Report (ACR). It can organise findings, suggest remarks and summarise test evidence.

But AI should not make final conformance decisions without verified evidence.

What Is the Difference Between VPAT and ACR?

VPAT stands for Voluntary Product Accessibility Template. It is a template used to prepare an Accessibility Conformance Report.

The completed ACR explains how a product supports the accessibility criteria covered by the report.

What Must Be Checked Before an ACR Is Final?

• Product name and version.
• Modules and features tested.
• Standards covered by the report.
• Testing methods and environments.
• Verified evidence for each criterion.
• Known accessibility barriers.
• Features not included in testing.
• Clear remarks and limitations.

Common VPAT Conformance Terms

These terms must be used according to the applicable VPAT edition and its instructions. They should not be chosen only by counting issues.

An AI tool finding no issues does not prove that a criterion is supported. Untested features should not be presented as tested.

Reference: ITI VPAT resources .

How Can Organisations Keep AI Accessibility Testing Reliable?

Teams need clear rules for using AI in testing. These rules help protect user data and improve the quality of reports.

Should Sensitive Data Be Sent to AI Tools?

Not without proper controls. Test pages may contain names, account details, health data or private business records.

Teams should review data-sharing rules, service agreements, access controls and retention settings.

Where possible, use safe test data and remove personal details before sending content to an external AI service.

How Enabled.in Helps Validate AI Accessibility Testing Results

Enabled.in provides digital accessibility audits, manual WCAG testing, assistive-technology testing, remediation guidance and VPAT/ACR services.

We help organisations review possible issues and confirm actual accessibility barriers. Our testing focuses on how digital products work for people with disabilities.

Based on project needs, testing may also include persons with disabilities to help assess real user experiences.

Who Can Benefit From These Services?

• SaaS and enterprise software companies.
• EdTech and LMS providers.
• Banking and financial technology teams.
• ERP and HR software vendors.
• Healthcare technology providers.
• Government technology contractors.
• Product owners and QA teams.
• Procurement and compliance teams.

Learn more about Enabled.in digital accessibility auditing and testing services .

Frequently Asked Questions About AI Accessibility Testing

AI Can Support Accessibility Testing. Human Validation Builds Trust.

AI accessibility testing can help teams work faster. It can review code, suggest fixes and organise findings.

But AI tools may report false issues or miss real barriers. Research from October 2026 shows that even strong AI models can struggle with accessibility test rules.

Organisations need a balanced approach. Use automated checks and AI assistance, then confirm results through expert testing.

Human review, screen reader testing, complete user journeys and retesting help make accessibility reports more reliable.

This approach also supports clearer remediation plans and better VPAT/ACR documentation.

Research Sources and References

1. Information and Software Technology (2026): An empirical investigation of Large Language Models for automated software compliance testing: Evidence from web accessibility .
2. W3C Web Accessibility Initiative: Selecting Web Accessibility Evaluation Tools .
3. W3C Accessibility Conformance Testing Rules Format .
4. W3C Web Content Accessibility Guidelines 2.2 .
5. Information Technology Industry Council: VPAT Resources .