Is It Safe to Paste Work Documents Into ChatGPT?
You've got a client contract open in one window and a chatbot in the other. You know it could summarize the thing in fifteen seconds. Your cursor hovers over Ctrl+V, and then you stop and think: wait, should I actually be doing this?
Short answer: it's usually safe to paste your own work into ChatGPT, Claude or Gemini if you've turned off model training and the text doesn't belong to someone else. Anything confidential, regulated or under an NDA — client records, signed contracts, anything with credentials — should be redacted first or kept out entirely. Here's the three-bucket test I use.
What actually happens to what you type
Start with the baseline fact that everything else follows from: when you paste text into a web-based AI tool, that text leaves your computer and goes to somebody else's server. It gets processed there. It is stored there, at least for a while.
Beyond that, two things vary and both are worth knowing:
Whether your conversations are used to improve the model. Consumer tiers have historically defaulted to yes on this, with a setting you can turn off. Business, team, and enterprise tiers generally default to no. API access is usually treated differently again.
How long it's retained. Even when your chats aren't used for training, they're typically kept for some period for abuse monitoring and support.
Here's the part I want to be straight with you about: I'm not going to quote you specific retention windows or policy clauses, because these change often enough that anything I write today could be wrong by the time you read it. What I can tell you is exactly where to look, which is the next-to-last section of this article. Check your own tool's settings page. It takes two minutes and it's the only version of this that's actually current.
The three-bucket test
Rather than agonizing case by case, sort what you're about to paste into one of three buckets. This is the mental model I actually use, and it resolves most decisions in about three seconds.
The three-bucket test at a glance. The table below has the detail.
| Paste freely | Redact first | Don't paste at all | Why |
|---|---|---|---|
| Public information; your own rough notes; generic questions about how to do something | Your own documents that happen to contain names, addresses, or figures | Anything under a confidentiality agreement you signed | The risk scales with how identifiable the information is and who it belongs to |
| A draft you wrote, with no third-party details in it | A spreadsheet where the structure matters but the specific values don't | Customer, patient, or client records; anything covered by privacy regulation | Other people's data isn't yours to take a risk with |
| Marketing copy, job postings, public policies | An email thread where you need the tone, not the identities | Passwords, API keys, account numbers, anything that grants access | Credentials are a different category entirely — treat a paste as a disclosure |
Notice that the middle column is the biggest one in real life, and it's the one most people skip straight past in either direction. They either paste the whole unredacted thing or they give up on the tool entirely. The middle path is usually available and usually easy.
How to redact without ruining the answer
The thing that makes redaction practical is this: for most tasks, the AI does not need the real names to do the reasoning. It needs the shape of the problem.
If you want a contract clause explained, "Party A" works exactly as well as the company's actual name. If you want help with a difficult email, "my manager" is as good as your manager's name. If you want a spreadsheet formula, the column headers matter and the customer names don't.
So swap in placeholders before you paste. A prompt like this does the job:
Below is a section of an agreement. I've replaced the company names with [PARTY A] and [PARTY B] and the dollar amounts with [AMOUNT]. Explain in plain English what obligations [PARTY A] is taking on, and flag anything that looks unusual or one-sided.
You get the same quality of answer, and nothing identifying left your machine. Then you map the placeholders back in your own head.
A second technique worth knowing: describe instead of pasting. Often you don't need the document in the chat at all. "I have a lease with a clause requiring 90 days' notice and an automatic annual renewal — what should I be watching for?" gets you most of the value with none of the exposure. The same logic applies when you're cleaning up a messy spreadsheet: you can often describe the structure and get a formula back without ever uploading the file.
When you do need the whole document in there — and sometimes you do, which is the case for summarizing a long PDF — that's the moment to make sure the settings in the next section are set the way you want them.
The two ChatGPT settings to change today (and where to find them)
Open whichever tool you use most and find its settings. You are looking for two things, which are usually within a screen or two of each other:
A data-controls or training toggle. Wording varies — "improve the model for everyone," "use my data for training," "data controls." Turning it off means your conversations shouldn't be used to train future models. If you can't find it, search the tool's help pages for "training data."
A history or retention setting. Some tools let you disable chat history, which usually shortens how long conversations are kept. Read the fine print on this one — see the caution below about temporary chats.
Do this once per tool and you've handled the biggest lever available to you. It's a two-minute job and almost nobody does it.
For the three tools most people use, the pages to look at are OpenAI's data controls FAQ for ChatGPT, Anthropic's privacy center for Claude, and Gemini Apps Activity for Gemini. The wording on each changes now and then, so read the page rather than trusting my summary of it.
I had been using this account for months and had never opened this screen; the toggle was On when I first looked. Switching it off took ten seconds. Yours may look different — the labels move around — but the toggle you're looking for is the first one. Note that turning it off only covers future chats; anything already sent isn't pulled back.
When your employer's rules are the real answer
This is the part that catches people, and it has nothing to do with technology.
Plenty of workplaces have an approved AI tool, or a policy about which data can go into one. If yours does, that policy is your answer, and it overrides every judgment call in this article. Pasting company information into an unapproved tool can be a policy violation even when it's perfectly safe in a technical sense — and "but the data wasn't sensitive" is not a defense against "you weren't supposed to use that tool."
If you don't know whether your workplace has a policy, that's worth five minutes of asking. The answer is often "yes, and we pay for a business tier you can just use," which solves the problem entirely.
The mistakes that actually cause leaks
Screenshots carry more than you meant to share. You crop to the paragraph you care about and forget the browser tabs, the client name in the window title, the notification that slid in at the corner. Look at the whole image before you upload it.
Uploaded files bring their history with them. A document can carry author names, tracked changes, earlier revisions, and comments that aren't visible in the view you're looking at. If it matters, paste the text rather than the file.
"Temporary chat" is not the same as "not stored." Temporary modes generally mean the conversation won't appear in your history or be used for training. That is not the same claim as the text never touching a server or being retained at all. Useful feature; read what it actually promises.
Free tools with no company behind them. The big names have published policies, legal exposure, and something to lose. A free summarizer you found through a search result may have none of those. Be much more conservative about what you hand to a tool you can't name the company behind.
The browser extension you forgot you installed. Extensions that add AI features to your email or documents can have broad read access to whatever page you're on. Worth an audit now and then.
Other people's information is not yours to gamble with. This is the one I'd hold hardest. If the data belongs to a client, a patient, a coworker, or a customer, the question isn't whether you're comfortable with the risk. They didn't get asked.
A rule you can follow
Most of what you'd want an AI tool for at work is fine to paste, some of it needs thirty seconds of redacting first, and a narrow slice shouldn't go in at all. The slice that shouldn't is defined less by how secret the information feels and more by whether it belongs to someone else or grants access to something.
Turn off the training toggle, learn the placeholder habit, find out what your workplace actually allows, and you can use these tools on real work without the low-grade unease. That unease is worth listening to once — and then worth resolving, so you can get on with it.
One last thing, unrelated to privacy but related to trusting these tools: getting a safe answer and getting a correct one are different problems. Checking whether the answer is actually true is its own skill.
Related reading
- How to Fact-Check an AI Answer in Under Two Minutes — which parts of an answer to distrust and four checks that work.
- How to Spot AI Scams and Deepfakes Before They Cost You — cloned voices, fake invoices and the new tells.
- How to Turn a Long PDF Into a Clear Summary With AI — three methods, including what to do when the PDF is too long.
Keep reading
Written by Mitch, a software analyst who tests software for a living. Every guide here comes from actually using the tool on a real task — including the parts where it falls over. Tested on ChatGPT Plus, Claude Max and Gemini (free). More about this site · Corrections: acheatsheet@gmail.com.
Get the next guide by email
A new tested how-to when there is one — usually a couple a month. No hype, unsubscribe any time.
Tested in real accounts. No affiliate links.
