A detailed guide clarifying how AI systems should process PDF documents has been published by The PDF Association. The FAQ warns that common ingestion practices risk information loss and hallucinations.

It offers practical guidance for information managers preparing document collections for AI ingestion.

The resource argues PDF content is highly valued by AI systems because PDFs function as the established “document of record” in human communication. This contrasts with HTML web pages, which the FAQ describes as transactional, short, and subject to change.

Citing a recent HuggingFace blog post, the Association claims PDF content offers higher information density and is inherently long-context.

A central warning concerns converting PDFs to plain text or Markdown before ingestion. The FAQ calls this approach “inevitably lossy” and an unnecessary “dumbing down” process that risks increasing AI hallucinations.

Stripping semantic information such as strikethrough, superscript, or table structure removes intended meaning. The FAQ uses the example of plain “22” versus a superscripted “2²” to illustrate how conversion loses mathematical, chemical, or footnote context.

The FAQ argues AI systems should also avoid processing PDFs page by page, because content routinely breaks across page boundaries. Isolating individual pages reduces understanding and raises hallucination risk.

Tagged PDF as key enabler

For records managers and digital transformation teams, the FAQ places particular emphasis on Tagged PDF. Tagged documents provide logical reading order, natural language indicators, table structure, and alt-text for images.

These tags give AI systems an unpaginated logical structure, avoiding the need to interpret pagination artefacts. The Association recommends born-digital documents conforming to WTPDF, PDF/UA-1, or PDF/UA-2.

Tables receive specific attention. Because PDF does not natively define tables in its graphical model, AI systems processing untagged PDFs must rely on error-prone visual analysis. Tagged PDF tables, by contrast, express semantics that align closely with HTML.

AI ingestion systems should capture all PDF components, including annotations, embedded files, bookmarks, and XMP metadata. Ignoring this information, it argues, reduces understanding and increases hallucination risk.

The FAQ cautions that PDFs containing unresolved redaction annotations may expose personally identifiable information during AI ingestion. Such annotations flag content for removal but do not purge it.

The guide also addresses security permissions. Some AI systems ignore author-set restrictions indicating that text and graphics should not be reused.

The PDF Association is the industry body representing PDF technology vendors and promotes ISO-standardised PDF features.

Read the full FAQ at https://pdfa.org/faq-ai-and-pdf/