Where do I start? · Is it safe? ·
Is ChatGPT safe for confidential information? The three-bucket rule
Short answer: some of it, in some tools, with one setting changed. The longer answer is a sorting job you do once, in about ten minutes, and never have to think hard about again.
What actually happens to what you paste
When you type into a free ChatGPT, Claude, or Gemini account, two things can happen to it. It's stored in your chat history, where anyone with your login can read it. And, unless you've turned it off, it can be used to train future versions of the model.
"Train" doesn't mean the next person who asks gets your client's phone number back. It means your text goes into a pile the company uses to improve the model, under a policy you didn't read. For your website copy, that's nothing. For a client's tax return, it's a problem — and a professional-obligation problem, not only a privacy one.
Business plans (ChatGPT Business and Enterprise, Claude for Work, Gemini inside Google Workspace, Microsoft 365 Copilot) contract not to train on your data. That is the main thing you're paying for, and often the only reason to.
The three buckets
Take a real week and sort everything you might paste into three piles.
Public
Anything already on the internet, or that you'd hand to a stranger: your website text, your Google reviews, a job photo where the customer said yes, your price list if it's on your site, the flyer. Paste freely, in any tool.
Internal
Yours, not secret, but not for strangers: your price list, how you schedule the crews, your proposal template, last month's numbers with the customer names taken out, the draft of an email. Fine in a free tool once training is turned off (below). Fine in any business plan.
Confidential
Other people's information, or anything you'd be in trouble for leaking: customer financials, health or employee records, anyone's phone number, street address, or card number, passwords, contract terms you've agreed to keep private, anything from a client's books. Stays out of free tools, full stop. It goes into a business plan only if that plan's terms say it won't be trained on and your own obligations — to a client, a licensing board, HIPAA — allow it.
The trick that makes most tasks work anyway: strip, then paste. A customer's request with the phone number and address deleted is internal. A contract with the names and account numbers replaced by "the client" and "[account]" is internal. You lose nothing the assistant needs. It's summarising the terms, not calling the client.
Turning training off
Where the setting lived when we wrote this, in September 2026. Menus move; if a path has changed by the time you read it, open the tool's settings and search for the word "improve" or "training".
- ChatGPT
- Profile → Settings → Data controls → Improve the model for everyone → off. Your history stays; new chats aren't used for training.
- Claude
- Settings → Privacy → Model improvement → off.
- Gemini
- myactivity.google.com → Gemini Apps Activity → Keep Activity → off. Google still keeps each conversation for 72 hours; it isn't used for training unless you send feedback on it.
- Copilot
- The free Copilot app: profile → Privacy → Training on conversation activity → off (and the voice one next to it). Microsoft 365 Copilot — the one inside your work Outlook — doesn't train on your data by contract, so there's no toggle to find.
Turn it off once on each account, then check it after a big update. Ten minutes, total. If you can't find it in two, email hello@aiandaguy.com and we'll tell you where it went.
What "safe" still doesn't mean
Turning training off makes it private. It doesn't make it right. The tool will still invent a date that wasn't in the letter, miss the one clause that matters, and add up a column wrong — confidently. So the confidential-safe version of any task still ends with you checking, against the original, every time.
And it's still in your chat history. If anyone else uses your computer, log out. If you'd rather a conversation never existed, most tools have a temporary or incognito chat that isn't saved at all; that's the right mode for a one-off question about something sensitive.
Two Workflows where this matters most: the plain-English summary of a contract or notice — the confidential case, done with names stripped — and meeting notes into actions, where the meeting decides the bucket.
Three questions people ask next
Is ChatGPT Plus safer than the free one?
No. Same setting, same rules. Plus adds capacity and features, not privacy. The privacy step is a business plan, and only if you have confidential work that can't be stripped first.
What about the AI built into my email or documents?
If it's part of a paid business suite — Gemini in Google Workspace, Microsoft 365 Copilot — it's covered by that suite's terms, which is usually the strongest privacy position you have. Read the one paragraph on training. If it says your data isn't used to train, that's usually the safest place to do confidential work, safer than pasting into a separate chat.
My bookkeeper says never use it. My lawyer says the same.
They're right about the confidential bucket in free tools, and you should listen. Ask them about the other two buckets. Most of the useful work — drafts, summaries, follow-ups, your own procedures — lives there, and none of it needs a client's data to be pasted anywhere.
Where this fits
Session 1 of an Engagement is this, done on your actual work: we sort what you'd paste into the three buckets together, find the settings, and set up the first Workflow with the rule built in. Twenty minutes with a Guide first, free, no pitch.
Book a free Fit Call