AI-assisted checks

Background

When is a PDF document accessible? This question is not easy to answer. To be considered accessible, a document must pass no fewer than 31 checks containing a total of 136 conditions. These checks are listed in detail in the so-called Matterhorn Protocol, the official standard for PDF accessibility.

Of the 136 conditions in the protocol, 41 require human judgment. A “classic” software tool, for example, cannot determine whether a paragraph belongs to a heading or to the body text (that is, whether an H or a P tag would be correct). Until now, all PDF accessibility testing tools have failed to address this problem.

PAC 2026 includes artificial intelligence (AI) functions. With these new features, PAC can now cover part of those Matterhorn Protocol checkpoints that previously required human judgment. This significantly reduces the amount of manual checking and makes professional PDF/UA testing much more accessible—especially for beginners.

How It Works

The AI built into PAC 2026 has been trained on a large number of accessible PDF documents. By learning from these examples, it has developed an understanding of how different content areas are semantically tagged.

Semantic areas (structural elements)

The AI-assisted checks recognize the following semantic areas within a document:

  • Body text (P)
  • Headings (H)
  • Lists (L)
  • Tables (Table)
  • Images and graphics (Figure)
  • Tables of contents (TOC)
  • Captions (Caption)
  • Footnotes and notes (Note)

Results

When the AI has analyzed a document, PAC 2026 displays three result categories:

Match of structural elements

The AI-assisted check has semantically identified an area in the document, and the PDF contains a tag for the same area.

If the AI tag and the document tag do not match, PAC 2026 reports an error or a warning, depending on the AI’s level of confidence.

Example: The AI-assisted check detects a heading (H), but the document contains a paragraph (P) tag at the same location.

Unrecognized structural elements

The document contains a structural element at a given position, but the AI-assisted check does not detect any corresponding content there. PAC 2026 displays a warning about this inconsistency.

Example: An empty image has been tagged as a “Figure.”

Missing structural elements

The AI-assisted check has semantically identified a region in the document, but this region is not tagged.

Example: A paragraph or image is not tagged.

Frequently asked questions (FAQ)

Does PAC 2026 access the Internet for AI checks?

No.
The AI model is delivered with PAC 2026 and runs entirely locally. PAC does not access the Internet in any way.

Does PAC 2026 check the correctness of alternative texts?

No.
Formulating or evaluating alternative text requires a so-called Large Language Model (LLM), such as those used by ChatGPT, Gemini, and other well-known AI applications. Such a model would be far too large to include with PAC, and sending PAC data to AI servers on the Internet is not an option. Therefore, PAC does not check alternative texts.

Does PAC 2026 check the correctness of the reading order?

No.
This should still be verified by a human. For this purpose, PAC provides the screen reader preview function.