AI checks in axesPDF

axesPDF goes beyond traditional testing mechanisms. In addition to automated testing of technical criteria according to PDF/UA and WCAG, the tool now also performs AI-supported tests that identify areas that previously required human judgment.

For example, it is not technically possible to clearly identify whether a section in a document is a heading or plain text. Conventional software cannot reliably assess this. With its new AI functions, axesPDF analyzes the document content and automatically recognizes structural elements such as headings, paragraphs, lists, and tables.

The AI thus provides valuable information about possible errors or inconsistencies in the semantic structure. For example, it indicates when a visually highlighted heading is not correctly marked up or when an image is included in the document without a tag.

AI-supported checks do not replace human judgment. They are intended to facilitate it. They help to identify potential problem areas more quickly and evaluate the quality of documents more efficiently. The AI results should therefore be understood as supporting information, not as official technical criteria.

Background

When is a PDF document accessible? This question is not so easy to answer. The document must pass no fewer than 31 tests with a total of 137 conditions. These tests are listed in detail in the Matterhorn Protocol. It is the checklist for PDF accessibility.

Of the 136 conditions in the protocol, 41 require human judgment. For example, “classic” software cannot recognize whether a paragraph belongs to a title or to the body text (i.e., whether an H or P tag would be correct). All PDF testing tools have failed to solve this problem so far.

axesPDF 2.5 introduces artificial intelligence (AI) features. With these new features, it can cover some of the Matterhorn Protocol's checkpoints that previously required human judgment. axesPDF 2.5 thus significantly reduces the manual testing effort.

Functionality

The AI built into axesPDF was trained using many accessible PDFs. It learned how to recognize semantically correct tagged areas based on these documents.

The AI runs exclusively locally, on the computer on which axesPDF is running. It does not access cloud services or the internet in any way. Accordingly, your PDF documents always remain on your computer and are not used to train the AI.

Semantic areas (structural elements)

The AI-supported checks recognize the following semantic areas of 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)

Frequently asked questions (FAQ)

Will axesPDF access the internet for AI testing?

No. The AI model is delivered with axesPDF and runs entirely locally.

Does axesPDF check the correctness of alternative texts?

No. To formulate or check alternative texts, a so-called large language model (LLM) is required, as used by ChatGPT, Gemini, and other well-known AI applications. Such a model would be far too large to deliver with axesPDF. It is not an option for axesPDF to send data to AI servers on the internet. Therefore, alternative texts are not checked.

Does axesPDF check the correctness of the reading order?

No. This should still be checked by a human. axesPDF has a screen reader preview function for this purpose.

How does AI recognize a paragraph as a heading? Is bold font decisive for this, for example? 

The AI does not automatically classify every bold font as a heading. The decisive factor is the combination of formatting, position, structure, and context in the document. Only when several characteristics indicate a heading does the AI issue a notification.

Does AI recognize the details of tables or lists?

No. AI recognizes tables or lists as a whole, but not individual elements such as rows, columns, or list items.

Does the AI in axesPDF continue to learn locally on its own?

No. The AI does not continue to learn locally. It is a static, non-adaptive model and does not change on the device.

Can an AI error be ignored so that it is no longer displayed?

No. Individual AI results cannot be manually ignored or hidden.