What Never to Paste Into an AI Tool

What Never to Paste Into an AI Tool

There is a moment in almost every AI workflow where it would be easier to just paste the whole thing in. The full contract. The customer spreadsheet. The real login. The actual medical report.

This lesson is about that moment, and where the line sits.

None of this is a reason to use AI less. It is a reason to use it without handing over things you cannot get back.

What actually happens to what you type

Three facts worth holding onto:

• Most consumer AI products may use your conversations to improve their models unless you turn that off, and the setting is usually on by default.
• Conversations can be reviewed by humans — for safety, abuse detection, and quality — so "nobody reads it" is not a safe assumption.
• Deleting a chat does not reliably delete everything. Copies can persist in backups, logs, and anything already used in training.

The practical version: treat anything you paste as though it has left your control. Not because a leak is likely, but because you cannot undo it if it happens.

In 2023 Samsung engineers pasted internal source code into ChatGPT to debug it. The code worked. It was also, at that point, outside the company. That is the shape of almost every real incident — nobody was careless, someone was just in a hurry.

These do not go into a general AI chat box, in any form, for any reason:

• Passwords, one-time codes, recovery phrases, API keys, and access tokens
• Bank details, card numbers, and anything that could move money
• Government identifiers — social security, passport, driver's licence, tax numbers
• Other people's personal data: customer lists, employee records, contact details, anything you were trusted with rather than given
• Medical records, diagnoses, and test results tied to a named person
• Anything under an NDA, anything marked confidential, and unreleased commercial information — pricing, forecasts, supplier terms, product plans
• Proprietary source code from an employer or client

The test is simple. If it would be a problem for this to appear in a screenshot, it does not go in the box.

These are usually fine, but only after you strip them down:

• A contract you want explained — remove the parties, the amounts, and the dates first
• A customer complaint you want help answering — replace the name, the account number, and anything identifying
• Your own health question — ask about the condition, not about your named record
• A spreadsheet you want analysed — delete the columns that identify people before you upload it

Redacting sounds tedious. It takes about thirty seconds, and it almost never costs you answer quality, because the model does not need the name to explain the clause.

How to get the help without handing over the data

Four moves that cover nearly every situation:

1. Replace real identifiers with placeholders — Client A, Vendor B, $X, City C. The reasoning transfers back perfectly.
2. Ask about the pattern, not the record. "How should a landlord respond to this kind of clause" gets you the same answer as pasting the lease.
3. Describe the structure of your data instead of uploading it — column names and row counts are usually enough for analysis advice.
4. Use a business or enterprise account for work. Those tiers typically do not train on your content and give the company a real data agreement. If your employer has one, use it instead of your personal login.

Settings worth changing today

• Turn off training on your conversations in every AI tool you use, in privacy or data controls
• Turn off chat history where you do not need it
• Check what each connector can actually reach before you approve it — a file connector with access to your whole drive sees your whole drive
• Use your work account for work and your personal account for personal, and do not blur them

If you already pasted something

Do not panic, and do not ignore it either.

1. Delete the conversation, then check whether the tool has a separate data-deletion request — deleting the chat and deleting the data are often two different actions.
2. Turn off training on your data now, so it does not apply going forward.
3. Rotate anything that is a credential — passwords, keys, and tokens should be treated as compromised and replaced, which takes minutes.
4. If it was company or client data, tell whoever owns it. This is uncomfortable for about ten minutes and much less uncomfortable than it being discovered later.

The one rule to remember

Everything in this lesson collapses into a single question, and it is worth asking before every paste:

Am I putting in something I would be fine seeing outside my control?

If yes, paste it and get on with the work. If no, spend the thirty seconds to redact it. That habit costs almost nothing and removes almost all of the risk.