Keep your data private with AI

What do you need before you start?

Gather the items below before you open the chat. You already have most of them.

• The chat you already use, signed in
• Ten minutes in that product's settings
• Your employer's or school's AI rule, if one exists

You do not need a paid plan to leave a secret out of the box. A paid or work plan changes how the company says it uses the chat. It does not make a password safe to paste. If you are new to the tools, do the 7-day start first, then come back for the settings.

How do you write down what you will not paste?

This list is the brake. It is not a legal policy. If a task needs something on the list, do that step yourself, or use the tool your employer or school already approved. The longer guide for ordinary work tasks is Use AI at work. The school version is Use AI for school the right way.

Leave these out of a personal chat, a custom instruction, a project brief, a memory, and an agent task:

• Passwords, one-time codes, and recovery keys
• A full Social Security number, passport number, driver's license number, or other government ID
• Bank account numbers, card numbers, and tax IDs
• Health details, diagnoses, insurance IDs, and medical records
• A client's, customer's, patient's, or student's confidential file
• Unreleased financials, prices, or deals
• Anything marked confidential, internal only, or covered by a contract you signed
• A coworker's review, a classmate's essay, or a private message that is not yours to share

OpenAI's scheduled tasks help says not to put health information, financial details, passwords, or account numbers in a task title or in task instructions. The same idea belongs in a normal chat. Google's Gemini Apps Privacy Hub says not to enter confidential information you would not want a reviewer to see.

A standing note is reused. Do not store this list's items in custom instructions either.

I need help with this task. Do not ask me for a password, an ID number, a full account number, or a medical or client file.

Task: [what done looks like]
Facts I can share: [the non-secret facts]
I will do this part myself: [the step that needs a secret]

If the task cannot be finished without that secret, say so and stop. Do not invent the missing fact.

How do you turn off model training in the tool you use?

Menus move. If a control named here is missing, it is not on this account. Use the switch you can see, and open the help page. Turning the switch off does not delete old chats. Delete those separately if you want them gone.

ChatGPT. OpenAI's data controls page says:

1. On the web, signed in, open your account menu, then Settings, then Data controls.
2. Turn off Improve the model for everyone, then select Done.
3. On iOS or Android, open the sidebar, your profile, Settings, then Data controls, and turn off the same switch.

When that switch is off, OpenAI says new conversations are not used to train OpenAI models. They can still appear in your chat history. A temporary chat is a different control: while it stays temporary, OpenAI says it is not used to improve the models, and it may be kept for up to 30 days for safety. If you later send feedback, such as thumbs up or thumbs down, OpenAI says the conversation tied to that feedback may still be used for training.

Claude, on a consumer plan. Anthropic's privacy settings article says this is for Claude Free, Pro, and Max, including Claude Code on those plans:

1. Select your name, then Settings, then Privacy.
2. Under Help improve our AI models, turn the switch off.

Anthropic says new chats and coding sessions are then not used for future model training. Data already in a training run that started, and in models already trained, stays in. Chats flagged by safety classifiers may still be used for trust and safety. An incognito chat is not used to improve Claude even if the switch is on. A thumbs-up or thumbs-down stores that conversation, and Anthropic says feedback may be used to train models.

Gemini, on a personal Google account. Google's activity help says:

1. On a computer, go to gemini.google.com.
2. Click Settings & help, then Activity. You can also open myactivity.google.com/product/gemini.
3. Turn Keep Activity off. You can choose Turn off, or Turn off and delete activity.

Google's Privacy Hub says that with Keep Activity off, future chats are not used to train its AI models unless you send feedback. The chats can still be saved with your account for 72 hours so Gemini can respond and for safety, including review by people for abuse. If you are 18 or over, Google says Keep Activity is on by default. Audio and Gemini Live recordings have a separate "improve Google services" box that is off by default. Leave that box off unless you mean to turn it on.

Copilot, signed in with a personal Microsoft account. Microsoft's current activity history page for the newer Copilot app says prompts, responses, and file contents in that app are not used to train foundation models. The matching privacy controls page lists memory, chat history, web search, and ads. It does not list a training toggle. An older privacy controls article does list Training on conversation activity, and Microsoft says that article applies to the older app. Use the page that matches the app you have. If you see Training on conversation activity, turn it off. If you do not, do not hunt for a switch the current page does not name.

Guides for each tool: ChatGPT, Claude, Gemini, Copilot.

How do you use a work or school account when the data is not yours?

A personal switch is the wrong place for someone else's confidential file. When a work or school plan exists, use that plan, and follow the rule your admin wrote.

• ChatGPT Business, Enterprise, Edu, and ChatGPT for Healthcare: OpenAI's data controls say OpenAI does not use that workspace content to train its models by default. Your organization's retention and access rules still apply.
• Claude for Work and the Anthropic API: Anthropic's commercial training article says inputs and outputs are not used to train models by default. Feedback you send, such as thumbs up or down, can still be used. An owner can turn off the rating control.
• Gemini with a work or school Google account: Google's Workspace privacy hub says Workspace does not use customer prompts to train generative models without the customer's permission. Google's activity help says a work or school account's Gemini activity settings are controlled by the admin. You may not be able to change them.
• Copilot signed in with a work or school account: Microsoft's data protection page says prompts and responses are not used to train foundation models, and that your IT admin can view logged prompts and responses.

None of those sentences means your manager cannot see a work chat. Ask before you paste internal material into a personal account. If the company has no written rule, ask your manager. Until you know, practice on text you could forward outside the company.

Here is a task that uses only facts I am allowed to share.
If you need a customer name, a student record, an unreleased number, or a file marked confidential, stop and tell me to use the work tool instead.

Task: [the job]
Facts I can share: [paste]
Tool I am allowed to use: [work account, or "personal, because this text is mine"]

How do you read the result before it leaves the chat?

A setting does not check the draft. You do. If the reply repeats a secret you should not have pasted, delete the chat after you have what you need, and do not paste that secret again. If you are about to hand the same material to an agent that can send mail or touch files, use Hand a task to an AI agent and keep send and pay behind your approval. Approve the plan before it continues, and check each phase before the next one starts.

Look at the draft below. List any password, ID number, account number, health detail, or confidential name that should not leave this chat.
Do not rewrite the draft yet.
Draft:
[paste your draft]

Delete every line it flags. Then decide whether the rest can be sent.

What mistakes should you avoid?

These are the misses that waste the result. Skip them before you send anything.

Common mistakes

• Pasting a customer spreadsheet, a student's file, or a medical note "just this once" into a personal account.
• Putting a password in custom instructions or in a scheduled task so you do not have to type it again.
• Trusting a training switch you have not opened. Read the setting on your account.
• Sending thumbs-up on a chat that contains something private. Several vendors say that feedback can be used to improve the model even after you opt out.
• Using a personal chat to get around a missing button in the work tool.

What should you keep private?

Work and school logs can be seen by the organization that owns the account. A personal account is the wrong workaround. Read the vendor page linked in the steps, and follow your employer's or teacher's rule when the two differ. The project memory guide is for facts you want the next chat to keep. It is not a place for secrets.

What should you check before you use it?

Answer each line below. If one is still a no, do not use the result yet.

Director check: look before you use it

• Is this fact on the do-not-paste list? If you are unsure, leave it out.
• Did you open the training setting, or are you remembering a screenshot? The switch on your account is the one that counts.
• Does this material belong to an employer, a client, or a class? Use the account they gave you.
• Could you forward this chat to your manager or your teacher? If not, something in it should not be there.

What is one extra tip?

When the task is allowed and the secrets are out, go back to a normal guide and do Step 1 only. A standing note about tone belongs in custom instructions. A secret never does.