What should I share, and what can it do?

Writing an invitation needs the event details. It doesn't need your address book. Start by separating the information a task needs from the access an assistant happens to offer. A public notice and a private conversation deserve different decisions.

Ask 2 questions: what information can the product use, and what actions can it take? They need different checks.

The words "the AI remembers" can hide several different processes:

• Conversation context: material available while producing the current answer, such as your request and earlier messages.
• History: saved conversations you may be able to revisit.
• Memory: a product feature that saves selected information for use in later conversations.
• Training: a process used to develop or improve a model. Depending on the service and your settings, conversations may later contribute to that process.
• Connected tools: services the assistant can use, which may receive information needed for their part of a task.

These are separate. Keeping a chat in history does not tell you whether it is used for training. Turning off training does not necessarily delete history or prevent information from reaching a connected service. A correction in a conversation can guide the next answer without immediately retraining the underlying model.

A temporary or incognito chat may still be kept for a period. Deleting a conversation may leave a separate saved memory or information in a connected service. Check the product's current controls for each. OpenAI's memory guide explains how remembered information and its sources interact in ChatGPT.

Use the provider's current explanation for your account. Personal services and employer-managed accounts can have different settings and terms.

OpenAI's data controls guide describes controls for model improvement and saved chats. Its data-use explanation also discusses exceptions, including submitted feedback. Anthropic's personal-account training explanation describes permission, feedback, and safety-review cases for Claude. These are vendor help pages; check them when settings or policies change.

If you use another product, start with its own guidance: Google's Gemini Privacy Hub or Microsoft's privacy guide for personal Copilot accounts. Choose guidance for your actual account; work and school rules may differ.

For a practice exercise, use fictional names and details. If a real task involves work documents or another person's information, first check the applicable rules and whether you have permission. Removing a name may still leave recognizable details.

An assistant may simply suggest words. A product connected to other software may also send a message, change a file, or complete another step. An agent is commonly used to mean an AI system that can choose steps and use tools toward a goal. The term does not guarantee that it acts reliably or has permission for every action.

Anthropic's explanation of agents describes how model-directed tool use differs from a fixed workflow. It is optional technical background.

Before allowing an action, check the exact result, destination, access, and consequences. "Write an invitation" should produce a draft. Sending it requires authorization for the recipients and final message. If the task only needs a draft, leaving sending access disconnected reduces what can go wrong.

Use this fictional task: "Draft an invitation for a book swap on Saturday, November 14, from 10 a.m. to noon at the community hall. Entry is free. People can bring up to 5 books. Children must come with an adult."

Does this task require your address book, a private message history, or permission to send messages?

Answer: none of those is needed to draft the invitation. The supplied event details are enough. Before sending a real invitation, review its wording and choose its recipients.

For a closer look at limiting what software can decide, explore bounded decisions.