What to Keep Private When You Talk to AI

Share enough to get help while keeping unnecessary personal information out of the task.

A volunteer coordinator wants help writing a thank-you note. She copies a donor's entire record into an assistant: name, address, phone number, giving history, payment details, and a private note about a family illness.

The writing task needed very little of that information.

It could have begun with: “Draft a warm thank-you for a donation of winter coats. Leave the recipient's name blank.”

That simple choice is the starting point for sensible privacy. Decide what the task needs before you upload what you happen to have.

By the end of this lesson, you will know how to reduce an input, distinguish important settings, and check the information a connected tool can reach.

First, ask what information is necessary

For a thank-you draft, the assistant may need the donation type, the message's purpose, and the desired tone. It may not need an address, account number, or private family circumstances.

For a student practice exercise, it may need a sample paragraph and learning goal. It does not need classmates' grades or a photograph of a school identification card.

For a fictional budget exercise, it needs amounts and categories. Real bank credentials add no educational value.

Try this question before sharing:

“Could I get a useful result with less information or a made-up example?”

Often, the answer is yes.

Recognize information that deserves special care

Some information is obviously sensitive: passwords, verification codes, payment details, medical records, and identification numbers.

Other information becomes sensitive through context. A staff schedule can reveal who will be alone in a building. A short account of a family crisis may identify someone in a small community. A customer note might contain details that person never expected to leave the business.

Treat information about other people with particular care. Access to a record does not automatically give you permission to upload it to any service you choose.

Workplace, school, clinical, legal, and other professional tasks may have specific approved systems and handling rules. Follow the rules that apply to the actual task rather than assuming a personal account is suitable.

Consider this fictional source note:

Jordan Ellis, account 847291, donated six winter coats. Home address and phone number are in the record. Jordan mentioned a private family difficulty during the visit. Please write a short thank-you.

For the writing task, a reduced input could be:

Draft a short, warm thank-you for a donation of six winter coats. Leave the donor's name as [Name]. Say that the coats will support our winter clothing program. Do not include any personal story or additional claims.

Even that program statement should be used only if it is true in a real case. Here it is part of the fictional exercise.

You can add the person's name in your own document afterward. The assistant does not need the entire record to help with the wording.

Replacing a name does not always make a document anonymous. Unusual events, dates, locations, and combinations of details can still identify someone. Remove irrelevant context as well as obvious identifiers.

Privacy settings answer different questions

It is easy to think of privacy as one switch. Instead, separate these questions:

Question	What you are checking
Will this appear in conversation history?	Whether the service keeps a conversation available in your account.
Can it personalize future responses?	Whether memory or similar features may use information later.
Can it help train or improve models?	The provider's model-improvement controls for your account and use.
How long may information remain?	Retention and deletion rules, including exceptions.
Can other people or services receive it?	Sharing, workspace access, and connected applications.

Changing one setting does not necessarily change all the others.

For a concrete example, OpenAI's current Data Controls instructions describe turning off “Improve the model for everyone” under account settings. The guide explains that conversations can remain in history while being excluded from model training under that setting. Check the instructions for your actual account. OpenAI Data Controls FAQ.

This is a specific product example, not a universal menu path for every assistant.

Understand what “temporary” means in the service you use

The word temporary can sound like a guarantee that no information is ever retained or used in any other way. Read the actual feature description.

OpenAI's detailed Temporary Chat guidance, checked for this article, describes a non-personalized default and an optional personalized mode. It says an unsaved temporary chat does not create new memories, may be retained for safety purposes for up to 30 days, and follows different conditions if you save it as a regular chat. Connected services can have their own policies. Temporary Chat FAQ.

The useful habit is to check the selected mode and its rules. Do not rely on an old screenshot or assume that another provider uses the same definition.

For sensitive work, a temporary mode is not a substitute for permission to use the service in the first place.

Check connected tools before granting access

An assistant may offer to connect to documents, email, calendars, or other services. Before doing so, ask what it will be able to read or change and whether that access is necessary.

For a packing checklist, you probably do not need to connect your inbox. For a document task, access to one selected file may be sufficient where the product supports that scope.

Information passed to another service may be handled under that service's policies. Google's Gemini privacy documentation, for example, explains that interactions with connected third-party services involve those services' own privacy practices. Gemini Apps Privacy Hub.

Review sharing controls too. A link to a conversation or shared project may expose more material than the final paragraph you wanted someone to read.

Stop adding more information to the conversation. Review the service's deletion and access controls. If a workplace or school record is involved, follow the relevant reporting process.

If you exposed a credential, use the issuing service's process to revoke or replace it. Deleting the conversation alone is not a substitute for securing the account.

Avoid treating deletion as a promise that every copy in every connected service disappears at once. Check the relevant documentation and the places where the information was actually sent.

Try it: make the input smaller

Write a fictional record containing a person's name, contact details, a purchase or donation, and an unrelated private note. Choose one task, such as drafting an acknowledgment.

Now create a reduced version containing only what that task needs. Replace identifying details with placeholders where useful.

You pass the exercise when you can explain why each remaining piece of information is necessary and identify any detail that could still reveal the person's identity.

For an extra challenge, compare two approaches: uploading a whole document and providing a short, checked excerpt. Which gives the assistant enough context with less exposure?

Privacy becomes easier to manage when you begin with the job and deliberately choose the information that belongs in it.