What Not to Paste Into ChatGPT, and How to Clean It First

A customer writes in, furious. Their order arrived broken, they want their money back, and they have put their full name, phone number and the last four digits of their card in the email. Possibly the whole card number. You copy the email into ChatGPT and type "help me write a calm reply".

The reply you get back is good. It is also the moment that customer's contact details landed on a server you don't run, under terms you probably haven't read, attached to your account for as long as the chat exists.

Most people never think about this, because a chat box feels like a scratchpad. It isn't one. This post covers what to keep out of AI chats and a quick way to clean text before you paste it, so you can keep using the tools without the side effects.

How often people paste sensitive data into AI

More often than you'd guess. Harmonic Security analysed tens of thousands of prompts that employees sent to ChatGPT, Copilot, Gemini, Claude and Perplexity during 2024, and 8.5% of them contained sensitive information. Nearly half of those leaks were customer data such as billing and account details, and about a quarter were about employees.

None of those people were trying to leak anything. They were trying to get work done, and the fastest way to give an AI context is to paste everything.

Why a chat window is not a private notepad

• It stays in your history. Anyone who gets into your account later, through a reused password or a shared laptop, can scroll back and read it.
• It may be used for training. On personal ChatGPT plans, conversations can be used to improve future models unless you switch off "Improve the model for everyone" under Settings, Data controls.
• Links travel. A shared chat link shows the whole conversation, including whatever you pasted at the top.
• Work accounts have admins. On a company workspace, the people who manage it may be able to see or export chats.

The best-known example is still Samsung. In 2023, engineers in its chip division pasted internal source code and notes from a meeting into ChatGPT while asking for help, and within weeks the company restricted generative AI tools on its devices. Nobody was hacked. The data simply left the building in a chat message.

What not to paste into ChatGPT

The test I use is simple: would the person this is about be surprised to learn it was sent to an AI company? If yes, it shouldn't go in as it is. In practice that covers:

• Other people's contact details: customers, patients, tenants, job candidates, your kid's teacher. A name next to an email address or phone number is the classic case.
• Government ID numbers, such as a Social Insurance Number, Social Security number, passport or driver's licence number.
• Card numbers, bank account numbers and IBANs, including the ones in a screenshot of an invoice.
• Health information about someone who can be identified, even indirectly.
• Passwords, API keys and recovery codes. These need rotating if they leak, which is a separate headache covered in is it safe to paste API keys into ChatGPT.
• HR and performance notes, salaries, anything said to you in confidence.
• Unreleased numbers, contracts and source code your employer hasn't cleared for AI tools.

Your own information is a different call. If you want to ask about your own lab results, that's your decision to make. Most of the list above is about other people, and they don't get a say when you paste their details.

The AI doesn't need the names to help you

Go back to the angry customer. To write a good reply, the model needs to know there is a customer, an order that arrived broken and a request for a refund instead of a replacement. Who the customer is makes no difference to the wording. So you swap the details for placeholders before pasting:

Ask for the reply with that version and you get the same quality of answer. Placeholders like [NAME_1] are also easy to swap back afterwards, which plain deletion is not.

A quick cleaning routine before you hit enter

Doing this by hand works for one email. By the third one you'll miss a phone number in a signature. That's the reason we built the PII Redactor, which runs entirely in your browser, so the text you're cleaning never gets uploaded anywhere either.

1. Paste the text into the redactor. Emails, phone numbers, card numbers, ID numbers, addresses and IPs are matched by format; names are picked up from greetings, titles and sign-offs.
2. Skim the highlights. Untick anything that isn't personal, like a product name that happened to follow "Hi".
3. Add anything it couldn't know about, such as your company name or a client mentioned by first name only.
4. Copy the safe version and paste that into the AI.
5. When the answer comes back, paste it into step 2 of the redactor to put the real names and numbers back in.

If you already pasted something you shouldn't have

Delete the conversation, and if you ever shared a link to it, remove the link too. Deleting doesn't undo the fact that the data was sent, though, so deal with anything that can be reused. Cancel a card number, rotate a password or API key, and if it was customer or employee data at work, tell whoever handles privacy at your company. They'll know whether it needs reporting, and it's far better coming from you.

Then make the cleaning step a habit. A full walkthrough, including how to get the details back into the AI's answer, is in how to redact personal information before using AI. Keep the PII Redactor open in a tab next to your chat and it adds about ten seconds to each request.