Why ndaact sometimes redacts too much, and how to report a missed redaction

Why ndaact redacts more than it needs to

Now and then you will see an ordinary word replaced by a placeholder: a street name read as a surname, a town read as a company name. That is how the models are tuned. We fine-tune our detection models from a deliberately strict base, so when they are unsure, they redact.

The reason is that the two possible mistakes do not cost the same. A word redacted for nothing shows up at once in the text you send, and no data leaves with it; at worst the AI’s answer is slightly less precise. A client’s name that the model misses goes to the AI provider in clear text, and nobody notices. Given the choice, we prefer the mistake you can see and that does no harm.

This tuning lowers the risk but cannot remove it. ndaact is an aid. Read the text before you send it: what leaves your machine is your decision.

Models retrained on a regular cycle

We retrain the detection models regularly, one language at a time. Each cycle produces a new version of the French, English, German, Spanish or Italian model, and the app downloads the one for your language. Every cycle aims at two things: fewer words redacted for nothing, and no extra name let through.

The cases you report feed the next cycles. A precise report is worth more than a general impression, because it tells us which word, in which sentence, with which model version.

How to report a missed redaction

Desktop app: open the menu of the tray or menu bar icon and choose “Report a missed redaction…”. The report covers the last clipboard redaction.

Chrome extension: in the message preview, click the report link.

Either way, a window opens with the text in question. You say whether ndaact redacted too much, not enough, or something else, then you review what is about to be sent.

What is sent

• the kind of report: too much, not enough, or other;
• the original text, as you edited it in the window;
• the redacted text;
• a comment, if you write one;
• the app version and the identifier of the active detection model.

Before sending, you can swap real data for made-up data: a fake name instead of your client’s, a fake address. The case stays useful as long as the sentence keeps its structure.

Nothing leaves without you. Sending takes your click, a full preview of the content, a consent box that you tick yourself, and then the “Send” button. There is no automatic sending and no pre-ticked box.

What we do with it

• A report is kept for 90 days at most, then deleted automatically.
• The text and the comment never appear in any technical log.
• The real text is never used to train a model. We rewrite each case into an equivalent fictitious one (same sentence structure, same kinds of data, invented values). Only that fictitious case is used for training, and the original report is deleted once it has been rewritten.

The details of this processing are in the privacy policy.