Alt text, structure recognition, character mapping: where AI really helps with accessible PDFs, where it must not be the one to decide, what data protection and labelling mean – and how DokAudit uses it.
For accessible documents, AI is a good tool for suggestions: alt text, structure recognition in difficult layouts and, as a last resort, mapping characters that a PDF does not reveal. It is not an auditor: whether an alt text is factually correct or whether a document meets the standard is not for the AI to decide. For public administrations, two further questions arise – which content goes to which AI service, and whether AI-generated text has to be labelled.
Where AI helps well
- Drafts for alt text: AI usually recognises well what an image shows. For large collections of documents this is a quick start – not a finished result (Writing good alt text).
- Structure in difficult layouts: multi-column typesetting, boxes, tables without rules. Where fixed rules reach their limits, a model that sees the page as an image can suggest a plausible reading order.
- Form fields in flat forms: recognising boxes and lines as fields – as a suggestion that a person corrects.
- Unreadable characters: if an embedded font lacks its Unicode mapping, individual characters can be recognised by their shape. This should be the last step after unambiguous methods (Case study: fonts without Unicode).
- Pointers for review: AI can draw attention to places a person should look at – for example an alt text that does not match the image.
Where AI must not decide
- Factual correctness: figures from charts, names, places, the legal meaning of a map. AI misreads – and sounds convincing while doing so.
- Purpose of an image: whether a photo carries meaning or is decoration follows from the document and its intent, not from the image.
- Statements about people: assumptions about the age, origin or feelings of the people pictured do not belong in an alt text.
- Conformance verdict: whether a document meets PDF/UA and the WCAG success criteria rests on transparent checking rules and human testing. The accessibility statement is the responsibility of the public body, not of a model.
Good practice: suggestion, review, evidence
- Fixed rules first, AI only where it is needed – this is more transparent, faster and more economical with data.
- Re-check every AI change with a validator afterwards and accept it only if the result does not get worse.
- Mark AI suggestions as suggestions in the workflow and have them approved by someone who knows the content.
- Spot checks targeted where errors hurt: charts, maps, people, amounts.
- Record what the AI did – for queries and for your own quality assurance.
Data protection
If the AI runs as a cloud service, document content is transferred – image sections, page images, blocks of text. For public administrations this is a question of the General Data Protection Regulation (GDPR): legal basis, processing on behalf of the controller and, where applicable, transfer to a third country (Art. 44 et seq. GDPR). Documents containing personal data are particularly critical, for example official decisions, minutes with names or non-public meeting documents. Clarify with your data protection officer:
- What content is transferred – whole documents or only individual sections?
- Where is it processed, and who is the provider?
- Is the data used for training, and when is it deleted?
- Can the AI be switched off for the organisation or for individual documents?
- Is there an alternative with processing in the EU or without AI?
Labelling
The EU AI Act (Regulation (EU) 2024/1689) contains transparency obligations in Art. 50. Among other things, deployers of an AI system that generates text published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated. According to the wording of the regulation, this does not apply where the content has undergone human review or editorial control and a person holds editorial responsibility. Whether alt text in a PDF falls under this at all has not been conclusively settled. In practice this means: anyone who reviews and approves AI suggestions on their merits is on the safe side – and gets better quality anyway. If in doubt, your legal department should assess your own practice.
Common mistakes
- AI alt text is accepted without review.
- ‘Checked by AI’ is taken as proof of accessibility.
- Sensitive documents are sent to a cloud service without data protection having been involved.
- The AI restructures a document that was already correctly structured – and makes it worse.
- Alt text contains speculation about the people pictured.
How DokAudit uses AI
DokAudit only uses AI when it is needed. Fixed rules run first: image captions are adopted as alt text, structures are generated by rules, characters are identified by outline comparison and learned mappings. AI comes into play for multi-column layouts, unclear tables, alt text suggestions, field recognition in scanned forms and, as a last resort, character mapping. An AI-generated structure is only accepted if the standards check does not rate it worse; afterwards DokAudit checks again with veraPDF. Alt texts remain suggestions, and the test report shows where a review by someone with subject knowledge is needed.
In the Hub, ‘AI when needed’ is the default setting; administrators can switch off AI for the whole organisation. Which content goes to which AI provider is set out in the privacy policy – currently this is a provider in the USA; on request, a provider with processing in the EU can be used. The platform itself runs in data centres in Germany.
Sources:
As of: 10/2026. This article gives a general overview and is not legal advice. The legal and standards texts in force are authoritative; in individual cases, the law of the German states (Länder) may differ.