# Anonymize Text > Replace names, phone numbers, emails and IDs in text with numbered tokens before you paste it into an AI, then restore them in the reply. Runs in your browser. - URL: https://bytescope.dev/en/anonymize-text/ - Locale: en · Also: [ja](https://bytescope.dev/ja/anonymize-text.md) · [zh-Hant](https://bytescope.dev/zh-Hant/anonymize-text.md) - Runs: Runs in your browser. Your text is not uploaded. The optional name finder downloads its model file once from huggingface.co, and the ONNX Runtime script loads from cdn.jsdelivr.net. - Requires: Any modern browser. The rule-based finder works everywhere and instantly. The optional AI name finder downloads a model of about 189 MB once from huggingface.co and runs on WebGPU where the browser has it, otherwise on the CPU, which is slower. The ONNX Runtime script loads from cdn.jsdelivr.net. Your text stays in the tab. - Updated: 2026-09-08 ## What it does ### What it finds Two kinds of detection run on the text. Rules catch the patterns with a fixed shape and a check digit: email addresses, phone numbers, Taiwan national ID numbers, Japanese My Number, card numbers (Luhn), IBAN, IP and MAC addresses, and secret-looking values in links. They run as you type. The optional AI name finder reads the whole text for people, addresses, companies and ID-like numbers that have no fixed shape. Names are usually found; addresses and company names are often missed, so check them yourself. ### Replace, copy, restore Each value gets a numbered token, such as [Person1] or [Email2]. The same value always gets the same token, so the AI can still tell two people apart. Copy the result into the AI. When its reply comes back with the tokens in it, paste the reply into the restore box and the tokens change back. The original values live only in this tab's memory; you can download them as a JSON table, and they are gone once the tab closes. Mask mode writes a solid block instead, and that cannot be restored. ### What it cannot do Automatic detection always misses something, so this page saves work but does not make a text safe. Even with every name removed, a job title, a company and an event together can still point to one person. Your company's data rules decide what you may share with an AI, and they come first. ### Where the model comes from The name finder is GLiNER-X small by Knowledgator, licensed Apache-2.0: a quantized model and its tokenizer files, about 189 MB in total. It downloads once from huggingface.co and stays in this browser's storage. It works best on English and Chinese text in our tests; Japanese addresses and company names are missed more often. ## Limits Does not guarantee that the text is clean. Automatic detection always misses some names, addresses and company names, and the remaining content (job title, company, event) can still point to a person. It does not work on images, scanned PDFs or screenshots. ## FAQ ### Is my text uploaded anywhere? No. The rules and the AI name finder both run in your browser, and the text is never sent to a server. The only network traffic is the one-time model download from huggingface.co and the ONNX Runtime script from cdn.jsdelivr.net. Closing the tab removes the replacement table. ### Why did it miss a name or an address? Automatic detection is a best guess. Names are usually found, but addresses and company names are often missed, and unusual names can slip through. Read the text yourself before you paste it, and mark anything it missed by selecting it in the preview. ### Can I get the original text back? Yes, in token mode. Paste the AI's reply into the restore box and each token is changed back to its original value. The AI sometimes adds a space or changes a bracket, and the restore step accepts the common variants. Tokens it cannot match are listed so you can check them. Mask mode cannot be restored. ### Is the output safe to paste into any AI? It identifies people less, but it is not safe by default. Check the output for anything detection missed, ask whether the rest of the text still points to a person, and follow your company's data policy. ### Does it work on images or PDFs? No. It works on pasted text and on .txt, .md and .eml files. For screenshots, use Redact Screenshot instead. ## Related tools - [Remove EXIF](https://bytescope.dev/en/remove-exif.md) — Strip location and camera details from a photo. - [Redact Screenshot](https://bytescope.dev/en/redact-screenshot.md) — Black out parts of a screenshot before you share it.