Imagine you run the front desk at a small clinic. A patient’s insurer wants a screenshot of a message thread to process a claim. The thread has the patient’s name, phone number, home address and policy number in it, and you need to send everything except those. It’s late, there are twelve of these to do, and there’s an AI chat window open in the next tab that would find every name in a second. You’d just have to paste the screenshot in.

​
And it’s not just clinics

The same moment happens wherever people share documents that contain other people’s details:
  • a paralegal sending a contract excerpt to opposing counsel,
  • a support agent attaching a chat log to a bug report,
  • an HR officer forwarding a payslip,
  • an engineer pasting a log file into a ticket, or
  • a journalist publishing a leaked memo without exposing its sources.
The document has to go out; parts of it must not.

​
People are already pasting it in

They reach for cloud AI in exactly this moment. In a January 2025 survey of 1,000 US employees by TELUS Digital, 57% of those using public AI assistants at work admitted entering sensitive information into them. The most common kind was personal data such as names, addresses, emails and phone numbers (31%), followed by customer information (21%). The cost of getting it wrong keeps rising. IBM’s Cost of a Data Breach Report 2026 puts the global average cost of a breach at $4.99 million, a record high and 12% more than the year before. Professional rules haven’t relaxed for AI either. When the American Bar Association issued its first guidance on lawyers’ use of AI, Formal Opinion 512 (July 2024), it reminded lawyers that under Model Rule 1.6 they must keep all information relating to a client’s representation confidential unless the client gives informed consent, and that this applies when they use generative AI. And covering text by hand has its own famous failure. In January 2019, lawyers for Paul Manafort filed a court document whose black bars sat on top of text that was still there: a Guardian reporter copied and pasted the “redacted” passages into a new document and read them.

​
So what can you actually do?

  • Don’t use AI. Read every line and cover it by hand. It works, but it’s slow, and attention fades by the twelfth document. Misses are silent.
  • Use a cloud AI tool. It’s fast and good at spotting names. But the document, with everything you meant to hide, is sent to someone else’s servers before anything is hidden. What happens to it next depends on that provider’s retention and training terms, your plan, and your settings.
  • Use a PDF editor’s box tool. Unless it truly removes the content, the text under the box is still in the file, as the Manafort filing showed.
  • Use AI on your own computer. The document never leaves the machine. Until recently, models small enough to run there weren’t good enough to trust.

​
Which leaves a few questions

What if the AI that reads your documents ran on your own machine, and the data never left it? Is local AI now good enough to find the names, addresses and account numbers for you? Could you keep the final say on every single one, so a wrong guess costs a click instead of a leak? And could you build it yourself, with tools you already know?

Read on

Let’s find out by building it.