The PDF rules. Contracts, reports, statements: it is the format people hand to an AI more than any other. But here is the catch. A PDF is never a simple block of text, and pseudonymizing one properly means first answering a single decisive question that shapes everything else about your method: does your file hold real text, or is it just a scanned image? It all flows from there.
PDF with text
The text is selectable. Detection and masking happen directly, the layout is preserved.
Direct maskingScanned PDF (image)
The text is inside the image. It must be OCR'd first, then masked, otherwise the information stays readable in the pixels.
OCR then maskingThe redaction trap
A black box over a name? Not enough. The text sits right there beneath the bar, perfectly copyable, and that intact version is exactly what slips into the AI while nobody notices a thing. Pseudonymizing means replacing the content, not hiding it. There is more in our dedicated article on why redaction fails.
The method, step by step
- Check whether the PDF is text (selectable) or scanned (image).
- For a scan, run OCR to recover the text.
- Detect and replace identifying data with tokens.
- Review the masked PDF, then use it with the AI.
- Re-identify the result locally if needed.
Safe-Doc does both. Text PDFs, scanned PDFs through OCR: it detects, genuinely replaces the identifying data and hands back a document whose layout stays faithful to the original. The full picture is in the ChatGPT and GDPR at work guide.
Test on your PDF. Text or scanned, it makes no difference: see for yourself before you send a single file to the AI.