AI Data Redactor

What this tool does

CM8-331 takes text you want help with and gives you back a version that is safe to show a chat assistant — every name, email address, telephone number, postal address, identifier, credential and term of your own replaced by a consistent stand-in. It keeps a key on your machine, so when the assistant answers you can paste the reply back and get your real names returned in one step.

The whole exchange happens inside this one file. There is no account, no upload and no network request of any kind. That is not a marketing position: a tool that anonymises data by sending it somewhere would be self-defeating.

The four-step loop

1. Paste your text. Replacement happens as you type. Nothing is sent anywhere and the box is never saved.
2. Copy the anonymised version and read it. This is the moment to notice anything the patterns could not catch.
3. Use it with the assistant exactly as you would have used the original.
4. Paste the reply back into the restore panel. Every stand-in becomes the real value again, and you have a usable answer with none of the exposure.

Why consistency matters more than concealment

The obvious way to redact is to black things out. It is also the way that produces useless answers: an assistant shown ████ emailed ████ about ████ has been handed a puzzle rather than a problem, and the reply reflects it.

Consistent substitution keeps the structure of the text intact. The same person is the same stand-in every time they appear, so the assistant can follow who did what to whom, which paragraph refers back to which, and what the argument actually is. The answer you get is close to the answer you would have had from the original — and nothing identifying ever left the building.

Stand-ins are assigned once and kept. Editing your text does not renumber people who were already mapped, because a renumbered key would silently corrupt any reply you restored afterwards.

Choosing the stand-ins

Four styles are available in Settings.

• Realistic stand-ins (the default) put ordinary fictional names, addresses and numbers in place of yours. Assistants handle these far better than placeholders, because they are just language. Email stand-ins use the example.invalid domain, which by design can never resolve, and telephone stand-ins use a range reserved for fiction — so a stand-in accidentally left in a real document cannot reach a real person.
• Bracketed tags such as [PERSON_1] are unmistakable, which makes them the right choice when the anonymised text will be reviewed or filed, or when you want to be certain at a glance that a substitution happened.
• Plain words and short codes sit between the two.

One rule overrides the setting: credentials, card numbers, bank accounts, national identifiers and high-entropy strings always become tags, never realistic values. Generating a plausible-looking secret would create a new hazard — a string that reads like a live credential, sitting in a document, with nothing to say it is fictional.

Your own sensitive terms

No pattern can know that Halyard is a product launch, that Meridian Foods is your largest customer, or that a particular site name is commercially sensitive. Only you know that, so the tool asks.

Put one term per line. Matching ignores case and only matches whole words, so adding a short term will not corrupt longer words that contain it. Add => replacement after a term to choose the stand-in yourself, which is useful when you want the assistant to see a meaningful category — Meridian Foods => Customer A reads better in an answer than a fictional company name.

Terms are matched before every built-in detector and outrank them, because your assertion that something is sensitive beats any inference this tool makes. They are saved with your settings, so keep the list to terms you are content to have stored.

The key, and how to treat it

The mapping key contains your real data. That is not a flaw — it is what makes the process reversible — but it means the key deserves exactly the care the original text does. Anyone who has it can undo every replacement.

By default the key is held in this tab only and disappears when you close it. That is the safe default and the reason it is the default. Switching on Keep the mapping key in this browser writes it to local storage in plain text so you can restore replies days later; it is a real convenience and a real exposure, and the tool says so on the screen where the key is shown rather than burying it here.

Two controls sit above the key. Forget unused entries drops the ones that no longer appear in your text — tidy, but it also breaks restoring older replies. Forget the whole key discards everything immediately.

The printed report includes the key in full, because a report that omitted it could not be used to audit what was done. Print it only when you need that, and handle the paper accordingly.

Restoring the reply

Paste the assistant's answer into the restore panel. Stand-ins are replaced longest-first, so [PERSON_10] is never mangled by [PERSON_1], and matching is literal — no pattern is ever built out of your data.

Two things are worth knowing. If the reply paraphrased rather than quoted your stand-ins, nothing will be restored, and that is normal — read the answer on its own terms. And if the reply contains placeholders the key does not know, the tool says so and leaves them alone: those were invented by the assistant, not by you.

What anonymising cannot do

Replacing identifiers is not the same as making text anonymous, and the gap between the two has caused real incidents.

• Context identifies people. "The only forklift driver on nights who speaks Portuguese" survives every substitution and names somebody to everyone at that site.
• Small populations are transparent. In a team of six, "one of the two team leaders" is a coin toss, not anonymity.
• Combinations re-identify. A date, an order size and a region can be unique even when each part is not.
• Commercially sensitive facts have no shape. An unannounced closure or a lost tender is ordinary English and matches nothing.
• The tool over-reaches too. The unlabelled-name detector will replace product names and places. Untick a type or switch the detector off when that happens.

Read the anonymised version before sending it. The right question is not "have the names gone?" but "could somebody who knows this organisation still tell who this is about?"

This tool and the checker

Its sibling, Can I Paste This Into AI?, answers a different question. That tool assesses: it scores the risk, explains what it found and keeps a log that checking happened. This one transforms: it produces usable anonymised text and puts the answer back together afterwards. They share the same detection engine deliberately, so the two can never disagree about what counts as a phone number — a redactor that found less than the checker which sent you to it would be a trap.

The rules it applies

FAQ

Does anything leave my computer? No. The file contains no network code of any kind, and this is verified mechanically before release. Disconnect from the network and it behaves identically.

Why is the key not saved by default? Because it is a complete record of the sensitive data you were trying to protect. Making that opt-in is the only defensible default.

I restored a reply and got the wrong names. The key changed between anonymising and restoring — usually because Forget unused entries was pressed, or because the text was edited in a different session with the key not kept. Restore against the same key you anonymised with.

Can I use the same stand-ins across two documents? Yes, as long as the key survives between them: keep the key switched on, or do both in one sitting.

Why are currency amounts left alone by default? Because replacing them usually destroys the thing you wanted help with, and a changed figure can make an assistant's arithmetic confidently wrong. Switch it on when the figures themselves are the secret.

Is the anonymised text safe to publish? No. It is safer to ask a question with. Publication is a much higher bar and needs a judgement about re-identification that no tool can make for you.

Saving your work

Your settings, your own terms and your choice of which types to replace are written to this browser's local storage. The text boxes are not, by design — neither your original nor the reply. The mapping key is written only if you switch that on.

Treat Export .json as the real save, and note that it follows the same rule: the key is included only when you have asked for it to be kept. Export CSV gives you the key as a spreadsheet, which is convenient and sensitive in equal measure. Reset asks twice, then erases everything this tool has stored, including the key. There is no undo.

Accuracy & disclaimer

Anonymisation by pattern is not anonymity. This tool replaces what it can recognise; it cannot recognise a person described rather than named, a site identifiable from its own figures, or a codename that means nothing outside your building. Re-identification from apparently anonymous text is a well-documented problem, and small populations make it easy.

The mapping key is exactly as sensitive as the original text, because it contains it. Nothing here is legal advice, and producing anonymised text is not in itself compliance with any particular obligation. Where a disclosure would breach a contract, a data protection law or a duty of confidence, that decision belongs to the people accountable for it.