Don't paste the AI, please — not your passwords, patient notes, client contracts, or anything you wouldn't want sitting on a company's server for years. Chatbots like ChatGPT, Claude, and Gemini are genuinely useful for editing, summarizing, and debugging, but the text box doesn't work like a private notebook: what you type can be logged, reviewed by a contractor, or folded into a future training run, depending on the plan and a setting you probably never opened.
Short answer: Free and individual ChatGPT and Gemini accounts can use your chats for training or human review unless you turn that off; Claude's consumer plans run on a separate "Model Improvement" toggle worth checking. Business, Team, Enterprise, and API tiers exclude training by contract on all three. Before you paste anything sensitive, check your plan's data setting, strip names and numbers, and use Temporary or Incognito chat for anything you'd rather not have saved.
In my testing this week, I opened the privacy settings on ChatGPT, Claude, and Gemini back to back to see what each one actually defaults to, rather than what the marketing page implies. The gaps between them are bigger than most people assume, and none of the three interfaces flags the setting at the moment you're about to paste something you'll regret.
What you'll need
You don't need anything fancy — just two minutes and a habit. Open your AI tool's privacy or data-control settings before your next sensitive task, not after. Know which plan you're on, since free, Plus/Pro, and Team/Enterprise tiers are governed by different rules on the same product. Keep a short list of things you personally never paste (client names, account numbers, health details, unreleased code or numbers) so you don't have to think it through every time. And if your work genuinely can't tolerate any external logging — legal filings under seal, patient records, security incident data — have a local or self-hosted model on hand as a fallback, since no consumer chatbot setting fully replaces that.
Step-by-step: how to keep sensitive data out of your AI chats
1. Check your plan's default before you type anything
Free and individual accounts are not all the same. When I tested it, ChatGPT's free and Plus tiers train on your conversations by default; you have to visit Settings → Data Controls and switch off "Improve the model for everyone" yourself, per OpenAI’s own Data Controls FAQ. Google's Gemini works the same way through Gemini Apps Activity, which is on by default and lets chats be reviewed by human raters and kept for up to three years, according to Google’s Gemini Apps Privacy Hub. Claude's consumer plans use a separate Model Improvement toggle in Privacy Settings — check it, since Anthropic's help center lists several conditions under which chats can be used to improve the model.
2. Turn off training, or use a chat that never saves
Every major tool now ships a no-log mode: ChatGPT's Temporary Chat, Claude's Incognito chat, and Gemini's option to keep Apps Activity off. Use it for anything you wouldn't paste into a public forum. The tradeoff is real — Gemini, for instance, will only let a paused conversation continue for 72 hours before it disappears if Apps Activity is off — so this is a per-task choice, not a permanent setting you forget about.
3. Strip identifying details before you paste
Swap real names for "Client A," replace account numbers with the last four digits only, and cut dates down to the month if the exact day doesn't matter to your question. This takes under a minute and the AI rarely needs the specifics to help you rewrite an email, debug a function, or summarize a document.
4. Run a redaction pass first, in a separate message
Paste the raw text and ask the AI to list every name, account number, email address, or date it contains — without repeating them back — before you ask your real question in a fresh message. It's a fast way to catch what you missed on a skim.
5. Match the tool to the sensitivity of the work
Casual writing help is fine on a free consumer plan with training left on, honestly. Client work, source code, or anything under an NDA belongs on a Team, Enterprise, or API tier, where OpenAI, Anthropic, and GitHub all say by contract that customer data isn't used to train models. My Claude vs ChatGPT comparison covers what each paid tier actually includes if you're deciding where to draw that line.
6. Treat "delete chat" as a request, not a guarantee
Deleting a conversation from your history removes it from view, but backend retention windows, safety logs, and anything already used in a training run don't necessarily disappear on the same schedule. Assume a paste is permanent the moment you hit enter, and you'll make better decisions about what goes in.
Example prompts you can copy
Use these to redact before you share anything real:
"Before I ask my real question, scan the text below and list every name, email, phone number, account number, and date it contains. Don't repeat the text itself back to me — just the list of sensitive items, so I know what to remove."
"Rewrite this with generic stand-ins: replace any person's name with [NAME], any company with [COMPANY], and any dollar figure with [AMOUNT]. Keep the structure and tone exactly the same."
"I'm going to paste a support ticket. Summarize the technical issue only — ignore and don't reference any customer name, email, or account number that appears in it."
These do the redaction work for you in seconds, and they're worth saving as a template you reuse rather than retyping each time.
Common mistakes to avoid
The same handful of habits cause most of the damage I've seen reported. Pasting a whole codebase — including API keys or internal hostnames — to debug one function is the classic one; Samsung banned ChatGPT company-wide in 2023 after engineers pasted proprietary source code and meeting notes into it in three separate incidents within 20 days, as Forbes reported at the time. Uploading a patient chart, a legal filing, or a client contract to "just clean up the formatting" is another. So is assuming a personal free account gets the same protection as a company's enterprise contract — it doesn't, and that mismatch is exactly what catches people. Don't paste the AI, please, when you're testing a new tool for the first time either; that's when people are least careful and most likely to grab whatever document is open. And don't trust a chat's "delete" button to undo a paste you already regret — the fix is not pasting it, not cleaning up after.
Tools that make this easier
If you're choosing where to do sensitive work, start with the paid tiers built for it rather than trying to make a free plan behave. My how to use ChatGPT guide walks through where the Data Controls panel actually lives, and how to use ChatGPT for free is worth reading first if you're deciding whether the free tier's defaults are even worth the risk for your use case. On the Anthropic side, how to use Claude AI covers Projects and Incognito chat, both useful for keeping work siloed. If Gemini is your daily tool, how to use Gemini covers the Workspace version, where business accounts are excluded from training the way ChatGPT Enterprise and Claude for Work are. And if you're comparing tools specifically on trust and data handling, our AI tool ratings page scores that alongside output quality, not as an afterthought.
Consumer plans vs. business and API plans
| AI tool | Free / individual plan | Business, Team, Enterprise, or API |
|---|---|---|
| ChatGPT (OpenAI) | Training on by default; turn off "Improve the model for everyone" in Data Controls, or use Temporary Chat | Off by default — Team, Enterprise, Edu, and the API aren't used for training |
| Claude (Anthropic) | Governed by a Model Improvement toggle in Privacy Settings; Incognito chats are never used regardless | Off by default — retained API data "is never used for model training without your express permission" |
| Gemini (Google) | Gemini Apps Activity is on by default: chats can be reviewed by humans and kept up to 3 years | Off for Google Workspace business accounts |
| GitHub Copilot | Free, Pro, and Pro+ plans may train on prompts unless you opt out | Off by default — Business and Enterprise are covered by GitHub's Data Protection Agreement |
GitHub confirmed the Free/Pro/Pro+ change and the Business/Enterprise exclusion in its own Copilot data-use FAQ discussion.
My take
None of this means avoid AI — I use these tools daily and they're worth it. It means treat the text box like you'd treat an email to a vendor: useful, generally fine, but not the place for anything you'd need to explain later. Don't paste the AI, please, until you've made the two-minute check a habit — the setting, the redaction, the right plan for the job. Do that once, and the actual work gets easier, not harder.
Frequently Asked Questions
Don't paste the AI, please — but is any AI chat truly private?
No consumer chat is fully private by default. Free tiers on ChatGPT and Gemini use conversations for training or review unless you opt out; Claude's consumer plans route through a similar toggle. Business, Team, Enterprise, and API tiers are the ones that exclude training by contract.
How long does it take to check my AI privacy settings?
About two minutes per tool. Open Settings, find Data Controls (ChatGPT), Privacy Settings (Claude), or Gemini Apps Activity (Google), and confirm whether training or human review is on. Do it once and it sticks until you change it again.
What is the easiest way to keep sensitive data out of ChatGPT or Claude?
Redact before you paste: swap real names and numbers for generic stand-ins, or ask the AI to list the sensitive items in your text first so you can strip them in a second pass. Pair that with Temporary or Incognito chat for anything you don't want saved at all.
Can I still get help with confidential work using AI?
Yes, on the right tier. Team, Enterprise, and API plans from OpenAI, Anthropic, and GitHub all state that customer data isn't used to train models. That's the setup for client work, source code, or anything under an NDA — not a free consumer account.
Does deleting a chat remove it from the company's servers?
Not necessarily on the same schedule. Deleting removes it from your visible history, but retention windows and anything already pulled into safety review or a training run can outlast that. Assume what you paste is permanent the moment you send it.