Free tool
De-identify text before you paste it into AI.
Names, organisations, phone numbers, and Australian identifiers are replaced with consistent tokens, entirely in your browser. The AI gets your full problem. It never learns who it is about. When the reply comes back, restore the real names locally.
First use downloads a language model (about 110 MB) to your browser. It is cached, so this happens once. The download is what makes on-device masking possible.
Downloading model for first use…
The language model could not be loaded, so no masking was done. This tool refuses to produce output without full name detection, because a partially masked document that looks safe is worse than no output at all.
De-identified output
Review before you share. If anything was missed, edit your text above and de-identify again. The token map below stays in your browser only.
Held in your browser's memory. Download it if you want to restore names in a later session.
Restore the real names
Paste the AI's reply here and the tokens are swapped back to the originals, locally. Lost the session? Load a saved map file:
Restored text
How it works
Layer 1
Pattern matching
Australian-format phone numbers, emails, street addresses, Medicare numbers, ABNs, TFNs, BSBs, and card numbers are caught by deterministic rules.
Layer 2
Neural name detection
A named-entity model (BERT, running as WebAssembly in your browser) finds people, organisations, and places that no pattern rule could.
Layer 3
Propagation
Every detected value is swept across the whole document, including slug forms in email addresses and file names, so one missed mention doesn't undo the rest.
The same identifier always becomes the same token, so "Sarah" is PER_x everywhere she appears and the AI can reason about her consistently. This architecture runs in production in the clinical and coaching systems we build; the design is documented in our essay on on-device de-identification and grounded in a fact-checked literature review of two decades of clinical de-identification research.
Don't take our word for it
Most AI privacy is a policy. This is architecture, and architecture can be checked. Open your browser's developer tools (F12), select the Network tab, and de-identify something. After the one-time model download, you will see no request carrying your text. There is no server to send it to: the masking code, the model, and the token map all live on your device.
Honest about limits
No automated de-identification is perfect, including this one. Unusual names, nicknames, highly distinctive descriptions ("the only orthodontist in Kununurra"), and identifiers in formats we don't recognise can slip through. Always review the output before sharing it.
This tool is best-effort de-identification, not a compliance guarantee, and nothing on this page is legal advice. If you handle regulated data, your obligations under the Privacy Act and your professional standards are yours to assess. What this tool changes is your starting position: instead of deciding whether to paste raw client data into an AI, you are deciding whether a reviewed, masked version still says too much.
Questions
Is it safe to paste client information into ChatGPT or Claude?
Not raw. Text you paste into a consumer AI tool leaves your device and is processed on the provider's servers, and depending on your settings it may be retained or used for training. De-identifying first means the AI still gets the full context of your problem, but never sees who it is about.
Does this tool upload my text anywhere?
No. Detection and masking run entirely in your browser. The only network request is the one-time download of the language model itself. You can verify this: open your browser's developer tools, watch the network tab, and mask something.
What does it detect?
Names of people, organisations, and places (via a neural named-entity model), plus Australian-format phone numbers, email addresses, street addresses, Medicare numbers, ABNs, TFNs, BSBs, card numbers, and LinkedIn URLs (via pattern matching). Dates are left alone by default because they carry useful context for the AI, with an option to mask them too.
Why is there a large download the first time?
The name-detection model (about 110 MB) downloads to your browser on first use and is cached for next time. That download is the price of the privacy: because the model runs on your device, your text never has to leave it.
Can I restore the real names after the AI responds?
Yes. The tool keeps a local map of every token it created. Paste the AI's reply into the restore panel and the real names are put back, in your browser. You can also download the map as a file and reload it later.
Is this a compliance guarantee?
No. It is a best-effort tool that catches most identifiers most of the time, and you should review the output before sharing it. No automated de-identification is perfect, and this tool does not constitute legal or compliance advice.
This is what privacy-first AI looks like in practice
The same privacy architecture in this free tool goes into the systems we build for Perth professional services and trades businesses. If your team is weighing up how to use AI on client work without the client data leaving the building, that is exactly the conversation an AI Working Session is for.