The Role of AI in Scalable Accessibility Remediation

The numbers make the case plainly. 94.8% of homepages had detected WCAG 2 failures in 2025. ADA website accessibility lawsuits crossed the 2,000 mark in the first half of the same year. And publishers are staring down backlists of thousands of titles that need to meet standards that did not exist when they were produced.

Manual remediation was never built for this scale. At a certain point, adding more reviewers and more remediation hours stops improving outcomes. The inventory simply grows faster than the workforce can address it.

This is where AI enters the conversation. Not as a replacement for human expertise, but as the only realistic way to bring that expertise to bear at the volume publishers actually need.

What AI Can and Cannot Do

The distinction matters enormously in accessibility work, and the field has been getting clearer about it.

New tools are lowering the cost of auditing by reducing the amount of time needed to audit. This is not automating the manual evaluation. It is making it quicker to manually evaluate by automatically extracting code, screenshots, success criteria, and context. That framing is exactly right. AI handles the work that does not require judgment. Humans handle the work that does.

In practice, this means AI can reliably cover a significant portion of the accessibility workload: running automated scans against WCAG criteria, flagging low-contrast text, identifying missing form labels, detecting empty links and buttons, classifying images by type, and generating initial alt-text drafts based on the image and its surrounding content.

AI tools can now draft alt text or image descriptions. However, because results vary, and alt text should be contextual, it is necessary to review and edit outputs to ensure accuracy and relevance.

What AI cannot reliably do is provide judgment. It cannot assess whether the meaning of a complex chart has been accurately captured. It cannot evaluate whether a MathML rendering conveys the same understanding as the original equation. It cannot determine whether a reading order that passes automated validation actually makes sense to a screen reader user working through a dense scientific argument. These require domain knowledge, and in scholarly publishing, they require specialists who understand the content as well as the standards.

The Human-in-the-Loop Model

With human oversight, AI can be useful for learning about accessibility standards and best practices, remediating content, and developing alternative formats to expand access.

The human-in-the-loop model is not a compromise between AI efficiency and human quality. It is the architecture that makes both possible at scale. AI processes the volume. Humans verify the judgment calls. The result is a remediation pipeline that moves faster than any manual process while maintaining the accuracy that automated tools alone cannot guarantee.

This structure maps directly onto the layers of accessibility work that publishers face.

• At the detection layer, AI handles initial WCAG audits across large file sets, identifying issues by type and severity. Automated triage reduces the manual burden at precisely the point where it is highest.

• At the remediation layer, AI generates first-pass outputs: alt-text drafts, structural tags, reading order suggestions, colour contrast adjustments. Each is then reviewed and corrected by a human specialist. The AI draft is the starting point, not the output.

• At the validation layer, human specialists run final checks using real assistive technologies: NVDA, JAWS, VoiceOver, keyboard navigation, braille displays. These tests cannot be replicated by automated tools and remain essential regardless of how capable the AI pipeline becomes.

Why Scholarly Content Needs Specialist Oversight

The case for human-in-the-loop remediation is strongest in scholarly and STEM publishing, where content complexity is highest and the margin for error is lowest.

A generic AI model generating alt-text for a photomicrograph in a pathology textbook will produce a description of what it visually detects. A specialist reviewing that output will ask whether the description accurately conveys the diagnostic significance of the image in its clinical context. Only a specialist can answer the second question.

A chemistry structural formula, a statistical regression table, a historical map: each requires not just accessibility tagging but domain-appropriate interpretation. AI provides the scaffold. Human expertise provides the accuracy.

AI in accessibility remediation is not effective when used in isolation. The publishers who understand this treat AI and human expertise as complementary layers rather than competing approaches.

What a Good Programme Looks Like

A well-designed AI-assisted accessibility programme for publishers operates in three phases.

First, systematic triage: automated scanning across the full content inventory, categorising issues by type, severity, and content complexity. High-complexity content is flagged for specialist review from the outset.

Second, pipeline remediation: AI generates initial outputs at speed and human specialists review, correct, and approve. Turnaround times shrink significantly without sacrificing quality.

Third, validation and documentation: manual testing with real assistive technologies confirms real-world accessibility, and ACR/VPAT documentation is generated from the validated outputs.

Where S4Carlisle Fits

S4Carlisle's approach to accessibility remediation has been built on this model since before it became an industry conversation. Our GCA-certified accessible ePUB and PDF workflows combine AI-assisted processing for initial tagging, classification, and alt-text generation with specialist human review across every output, including complex STEM content where accuracy is non-negotiable.

The NINJA platform brings this capability to web platforms and Learning Management Systems, using a hybrid methodology that pairs automated testing against WCAG 2.2 Level A and AA criteria with manual review by experienced specialists and full interaction testing using real assistive technologies.

Scale is achievable. Accuracy is non-negotiable. The human-in-the-loop model is how you get both.