Remove PII before you paste into ChatGPT

How it works

1. Open this page — the detectors are already set for “Remove PII before ChatGPT”.
2. Paste your text, or drop a .txt, .md, .csv, .json or .log file onto the panel.
3. Check the highlights: every match is coloured by type, and one click keeps a type as-is.
4. Copy the safe version and paste it into the AI tool of your choice.
5. Paste the model’s reply back into step 3 to restore the real values.

Why the step matters more than it used to

Consumer chatbots are trained on what you type unless you change a setting. OpenAI’s policy is explicit that ChatGPT improves by training on conversations unless you opt out, while ChatGPT Business, Enterprise, Edu and the API do not train on your content by default. Anthropic changed its consumer terms in August 2025 so that Free, Pro and Max chats are used for training unless you opt out, with retention of up to five years; Team and Enterprise plans sit under different, commercial terms.

Then there is litigation. In the New York Times copyright case, OpenAI was ordered to preserve consumer output logs — including chats users had deleted — and only came out from under that obligation on 26 September 2025. Logs already captured under the order still exist, and in July 2026 the publishers moved for sanctions over the way deletion continued. Nothing about that story is unusual for a fast-growing platform; it just means the honest planning assumption is that a pasted paragraph may outlive the conversation.

What a good scrub actually looks like

The temptation is to delete everything sensitive and paste the wreckage. Models write worse answers from mangled input, so people quietly stop scrubbing. Placeholders solve that: [NAME_1] asked [COMPANY_1] for a refund on card [CARD_1] keeps every relationship in the sentence — who did what to whom, how many distinct people are involved, which value repeats — while carrying none of the identifying content.

The second half of the trick is the round trip. Most redaction tools hand you a cleaned string and walk away, which means you then retype the real names into the model’s answer by hand. Here the mapping stays in the page, so pasting the reply into step 3 restores every value at once. That is only possible in a local tool: a hosted service would have to keep your mapping on its servers, which is precisely the thing you were trying to avoid.

Tips

• Turn on All dates if you are working under HIPAA Safe Harbor — dates are identifiers there, and they are off by default because ordinary text is full of harmless ones.
• Put internal codenames, unusual surnames and product names into custom terms: no word list can know them, and they are often the most identifying thing in the text.
• Use Fake values instead of tokens when you want the model to write natural prose back — a letter addressed to “Jordan Blake” reads better than one addressed to “[NAME_1]”.