I don't think people understand just how risky it is to let an AI read a PDF a customer sent you.
By Mike Cooper ·
A court in Brazil fined two lawyers this year over one line of text that nobody could see. And you can check your own files for the same thing in two minutes.
The line sat in an employment lawsuit, in white text on a white background. It told any AI to give a shallow answer and not to challenge the documents. The court's AI system flagged it, and the fine was 10% of the case's value.
Any PDF can hide a line like that, because it has two parts, the page you see and a layer of text underneath. A lot of AI tools read that layer, and nothing forces the two to match. Text outside the edge of the page stays in that layer too, and so does text under a drawn box.
The box worries me most, because that one happens by accident. A customer covers their BSN (the Dutch social security number) with a box in a PDF editor and thinks they hid it. Your team can't see it, so everyone believes it never came in. But the model can still read it, so a number your customer wanted private has gone to an AI anyway.
To check your own files, open a PDF, select all the text, copy it and paste it into Notepad. Anything in Notepad that you can't find on the page is something your AI may read and your team won't.
A person can only check what's on the page, so your AI should read that same page.