AI & PDF Accessibility: Hype, Help, or Hindrance?

Thesis: AI can quickly and cost-effectively generate alternative text for images, thereby reducing the workload for authors in their day-to-day work.

Where AI Can Be Used Effectively

AI-powered image descriptions offer clear advantages:

• Time savings & scalability: Alternative text can be generated much more quickly, especially for extensive documents with many images.
• Support for Beginners: Authors with little experience in writing accessible text receive solid wording suggestions right away.
• Consistency: The tone and structure of the descriptions remain more consistent across many images.

In many cases, AI can prevent relevant images from being left entirely without alternative text. In doing so, it establishes a basic level of accessibility, but it does not replace the decision as to whether an image is informative, functional, or purely decorative.

Where AI Reaches Its Limits

The quality of AI results depends entirely on the expertise of the users (prompt engineering). A simple “Generate an alt text” isn’t enough. The following must be defined in advance: How long should the description be? What details (numbers, objects) are relevant? Who is the target audience?

This gives rise to the following key risks:

• Lack of context & incorrect focus: AI does not automatically understand the editorial intent behind an image. Without precise instructions and the surrounding text to provide context, it often selects the wrong focus or misinterprets diagrams.
• Hallucinations & Bias: AI models are trained to always provide an answer. If they are not used precisely, they tend to simply invent information (“hallucinations”) or reproduce stereotypical biases (“bias”) when dealing with image content that is difficult to interpret.

Thesis: AI automatically handles tagging, structuring, and reading order, eliminating the need for manual follow-up work.

Where AI Can Be Used Effectively

Compared to traditional, rule-based algorithms, AI operates with significantly greater flexibility and context sensitivity:

• Semantic structuring: Headings, images, tables, and blocks of text are recognized both visually and semantically and assigned to the appropriate structural elements.
• Improved Reading Order: Even more complex layouts, such as multi-column text or infoboxes, can be interpreted more effectively.
• Optimized Text Recognition: Scanned documents are analyzed more precisely and converted more reliably into searchable, machine-readable text.

A key advantage: AI can efficiently ensure that a basic tag structure exists in the first place, an important first step and a solid foundation for further accessibility.

Where AI Reaches Its Limits

However, especially with complex documents, it quickly becomes apparent that AI automation is not a sure thing:

• Complex tables and formulas: While nested data matrices or linked table headers are recognized visually, they are often translated incorrectly from a structural standpoint because the AI lacks a deep semantic understanding of these complex relationships.
• Lack of contextual understanding: Current AI models cannot always reliably determine whether a graphic provides genuine added value or should be hidden from screen readers as a purely decorative element (“artifact”). When it comes to document structure, AI also sometimes tends to tag purely visual formatting, such as large, bold text, as hierarchical headings, rather than reflecting the document’s actual logical flow.
• Data Protection Risks and Compliance Violations: When dealing with confidential or personal data, the use of standard cloud-based AI systems is strictly prohibited. In such cases, dedicated enterprise solutions with strict agreements (opt-out for training data, retention periods, data processing agreements) are absolutely essential to avoid GDPR violations.

Thesis: AI detects errors that humans overlook and makes testing processes faster and more objective.

Where AI Can Be Used Effectively

Traditional testing tools follow rules rigidly. AI, on the other hand, helps in situations where human understanding of context needs to be simulated:

• Aligning Layout and Semantics: AI detects visual discrepancies, for example, when text appears to be a prominent heading but is incorrectly defined as a simple paragraph (<P>) in the tag tree.
• Quality and plausibility checks for alternative text: Only a few validators, such as the PAC PDF Accessibility Checker, also evaluate the content of alternative text. While most traditional tools typically only check whether alternative text is present, AI goes a step further: It analyzes the actual image content and assesses whether the alternative text is helpful.  This allows her to determine whether a description offers real value or consists merely of meaningless phrases such as “an image with text” or “diagram.”
• Language Accessibility: AI models can efficiently assess complex texts for comprehensibility, check for compliance with plain language and simple language guidelines, and provide specific suggestions for improvement.

Where AI Reaches Its Limits

AI is no substitute for conformance checks (such as PDF/UA or WCAG), it lacks the necessary precision and understanding of context:

• Lack of reliability in formal syntax: Generative AI cannot reliably verify the correct, nested tag structure of a document or the proper inheritance of table headers (TH). In this regard, traditional validation tools (such as the PDF Accessibility Checker (PAC)) are unbeatable and absolutely essential.
• Complex reading orders: In unconventional, magazine-style layouts with multi-column text, embedded quotes, or complex infoboxes, the AI’s visual logic still fails too often. The actual reading order for screen readers is then misinterpreted.
• Contrast in Busy Backgrounds: While areal contrast is easy to calculate mathematically, AI struggles to accurately evaluate text against color gradients, transparent layers, or complex image backgrounds because it “estimates” pixel ratios rather than measuring them precisely.
• Identification of functional design elements: Specific layout components such as “stopper elements,” decorative separators, or icons without text associations are often misclassified by the AI, it tends to mistakenly tag purely decorative elements as content-relevant.

Artificial intelligence is neither a fully automated, out-of-the-box solution for PDF/UA compliance nor mere marketing hype. Its strengths lie in its enormous scalability, processing speed, and resource savings for routine tasks. However, its weaknesses often become apparent when dealing with complex information architectures, unconventional layouts, and its inherent lack of true contextual understanding.

The key factor remains: Digital accessibility cannot simply be automated at the push of a button. Legal compliance, functional quality, and genuine, accessible usability for people with disabilities absolutely require sound human expertise.

The principle is called “human-in-the-loop”: The AI acts as a highly efficient assistant for pre-tagging and semantic analysis. It relieves experts of repetitive tasks, but at no point does it replace the in-depth expertise required for final, normative quality assurance. The AI provides the foundation, the final approval remains the responsibility of humans.