How to protect your personal data from AI tools

Short answer

To protect your personal data from AI tools, reduce what you put into the prompt before you look for a privacy setting. Don’t paste passwords, account numbers, full medical records, government IDs, confidential work documents, or another person’s private messages; then turn off model-improvement settings where available, review chat history and memory, disconnect unneeded apps, and limit device permissions. Those controls matter, but they do not make a consumer AI chat a vault: the service still has to process what you send, and retention rules differ.

A useful default is simple: remove identifying detail first, then ask for help. In many everyday tasks, the model needs the structure of the problem, not the names, numbers, and history attached to it.

What personal data should you never put into an AI tool?

There is no single list that fits every product. Privacy terms, account types, workplace agreements, and settings all vary. But for a regular personal account, treat anything that could unlock money, an account, health information, a legal matter, or another person’s safety as out of bounds.

Avoid putting these into an ordinary AI chat or upload:

• passwords, MFA backup codes, one-time codes, API keys, and crypto-wallet recovery phrases;
• card and bank-account numbers, tax identifiers, passport details, and driver’s-license images;
• full medical files, test results tied to your identity, insurance identifiers, and detailed therapy notes;
• customer lists, unreleased financials, source code, contracts, or internal plans unless your employer has explicitly approved that tool and use case;
• a child’s identifying details, home address, real-time location, routine, or school information;
• private messages, photos, voice recordings, or documents belonging to someone else without their permission.

CISA advises people not to share personal details or confidential company information with generative AI. Anthropic’s consumer guidance similarly highlights financial information, health records, credentials, and confidential personal or business documents as information to handle with particular care.[1][5]

This does not mean AI is useless for paperwork or difficult conversations. Change the input. Instead of uploading a lease and asking, “What can go wrong here?”, ask for a checklist of terms to review: deposit, repairs, early termination, rent increases, subletting, and notice periods. Instead of pasting a medical report with your identity attached, describe the situation at a high level and ask which questions would be useful to take to a clinician. The tool can help you prepare; it should not become the place where sensitive records live.

How do you know whether a prompt is safe to send?

Pause for ten seconds before you hit Enter and ask three questions:

1. Does the model need this detail to do the job? If not, remove it.
2. Can I replace it with a role, range, or placeholder? “Client A,” “a mid-sized city,” “around $1,000,” and “a school-age child” are often enough.
3. Would I still be comfortable if this text were seen beyond my own screen? That does not mean it will be public. It means your protection should not depend on the most optimistic assumption about a service.

This quick triage keeps the decision practical:

Redaction is not the same as anonymity. A rare job title, small town, exact date, unusual health event, and dollar amount can identify a person when combined. For sensitive material, remove more than the name: delete IDs, contact details, addresses, exact dates, cloud links, location information, and distinctive background facts.

Which privacy settings should you change in ChatGPT, Gemini, Claude, and Copilot?

Menu labels change, so follow the purpose of a setting rather than a years-old screenshot. The steps below were checked on October 30, 2026 and apply to personal accounts. Work and school accounts may be governed by an administrator.

ChatGPT: turn off model improvement and use Temporary Chat for one-off tasks

1. Go to Settings → Data controls.
2. Turn off Improve the model for everyone. For new conversations in personal workspaces, this stops chats from being used to improve OpenAI models; it does not erase chats already in history.[2]
3. For a one-off question that has already been redacted, start a Temporary chat.
4. Before sending, check whether that temporary chat is personalized if you do not want it to draw on existing memories or custom instructions.
5. Do not use thumbs-up or thumbs-down feedback on a chat containing private context. OpenAI says the conversation tied to feedback may be used for training even when the general improvement setting is off.[2]

Temporary chats do not appear in normal history, do not create or update memory, and are not used to improve models. They may still be retained for up to 30 days for safety purposes. That makes the feature useful for a short, lower-risk task—not a promise of zero retention.[2]

Gemini: turn off Keep Activity and review connected apps

1. In Gemini on a computer, open Settings & help → Activity.
2. At Keep Activity, choose Turn off or Turn off and delete activity.
3. In the same area, delete chats you no longer need and shorten auto-delete if that fits your use.
4. Open Connected Apps and remove access to services Gemini does not need for the task.
5. If Gemini is your Android assistant, separately review Google app permissions for camera, microphone, contacts, notifications, and location.

Google says that when Keep Activity is off, conversations may still be kept with your account for up to 72 hours to provide the service and process feedback, but they will not appear in Gemini Apps Activity. Turning it off also limits some features. Gemini Live audio, video, and screen sharing deserve a stricter standard than a typed question because they can carry much more incidental context.[3][4]

Claude: review model-improvement permission, chat reference, and project knowledge

1. Open Settings → Privacy and check whether you have allowed chats or coding sessions to help improve Claude.
2. Turn that permission off if you do not want future conversations used for that purpose.
3. In Settings → Profile, turn off Search and reference chats if you do not want Claude to find context in your earlier conversations.
4. Delete old chats that should not be searchable. Anthropic says deletion is the way to ensure a specific prior chat is excluded from chat search.[6]
5. Do not upload sensitive material to Project knowledge “just in case.” Content stored there is used as context across chats in that project.[7]

A model-improvement choice is worth making, but it does not remove the need to minimize inputs. A provider cannot process a detail you never sent.

Microsoft Copilot: clear history and keep personal and work use separate

In the current personal Copilot experience, you can delete an individual chat from the menu beside it and clear all history in the Microsoft Privacy Dashboard: Privacy → Empower your productivity → Copilot → Your Copilot app activity history.[8] Avoid sharing a conversation link that includes personal material: anyone with that link can see the snapshot of the conversation created at the time you shared it.[9]

Microsoft 365 Copilot in a work or school organization is a different context. Microsoft says prompts and responses are not used to train foundation models, but they can be logged and an IT administrator may have audit tools. That is not a blanket invitation to upload anything; it is a reason to follow your organization’s data classification and AI policy.[10]

Why is “don’t train on my chats” different from “don’t keep my data”?

AI privacy has several separate layers:

• model improvement or training: whether eligible future content can help improve a model;
• chat history: whether you can reopen a conversation later and continue its context;
• memory and personalization: whether the product saves useful facts for future answers;
• retention and safety review: how long data can be held to operate the service, prevent abuse, or meet legal obligations;
• permissions and integrations: whether the tool can reach email, files, calendars, your screen, camera, microphone, or contacts.

Turning off one layer does not automatically switch off the others. In ChatGPT, for example, opting out of model improvement does not delete history. In Gemini, turning off Keep Activity does not mean the service stops processing the conversation immediately.[2][3] The sensible order is: share less, choose the right mode and settings, then delete what no longer needs to remain.

How can you handle files, voice, and AI browsers more safely?

A file can reveal much more than the text you see. It may contain author names, comments, revision history, sheet tabs, hidden columns, folder names, or photo location metadata. Before uploading, make a copy. Remove identifying columns, replace real names, clear comments, and check that your screenshot does not include browser tabs, notification banners, an address bar, or someone else’s message.

Be stricter with voice and screen sharing. A spoken request can capture background conversations, names, television audio, or notifications. Stop microphone access when you are done, and do not share a screen while email, chat, password management, banking, or calendars are open. The FTC notes that voice assistants can activate by mistake on sounds resembling a wake word; if a device offers a physical microphone control, use it for sensitive conversations.[11]

For an AI browser or agent, do not hand it your main browser profile—the one already signed in to banking, email, cloud storage, and work systems. Make a separate profile with no saved passwords, and enable extensions or connected tools only for the specific task. That is not an extreme measure; it is basic context separation. See “Don’t give an AI browser the keys to your whole life”. Keep credentials in a dedicated tool rather than a note or prompt, too: why a password manager helps—and how to switch.

A 15-minute, no-purchase experiment: make one clean prompt

Choose a real task you would normally hand to an AI tool with the original source material: a client email, meeting notes, or an explanation of a bill. You do not need new software or a subscription.

1. Make two copies of the draft. Keep the original private and label the other one “AI copy.”
2. In the AI copy, replace names with roles, exact amounts with ranges, dates with relative time (“last month”), and company names with “Company A” and “Vendor B.”
3. Remove signatures, contract numbers, contact details, cloud links, and anything that does not change the task.
4. Turn off model improvement for new chats in the service you use and choose a temporary mode if it offers one.
5. Ask the tool to complete the task using the cleaned version. Compare the result with what you expected from the original.

For many writing, planning, and analysis tasks, structure matters more than biography. Save the version that worked as a reusable prompt template. Privacy becomes less about a dramatic quarterly settings audit and more about a normal way of framing work.

When this doesn’t apply

Do not try to self-redact material where a mistake could seriously harm someone or breach a legal, professional, or contractual obligation: medical records, children’s information, legal case files, regulated financial data, or customer data are common examples. In those cases, either do not use a public AI service or use only an organization-approved system with clear contractual terms, access controls, retention practices, and internal authorization.

This approach also has limits when the original data is essential for a decision. AI can help you make a review checklist or explain a process, but the primary record should stay in the system designed for that data. For workplace use, decide in advance what the tool may handle and how much checking time the result deserves; our guide “Before You Use AI, Set a Checking Budget” can help establish that boundary.

FAQ

Is using an AI tool without an account private?

It may reduce the connection between a prompt and your profile or saved history, but it does not make the input anonymous or confidential. Keep sensitive information out either way.

Does deleting a chat erase it forever?

Deletion removes the conversation from your visible history and starts the provider’s deletion process, but it does not necessarily mean immediate removal from every backup, safety, or legally required system. Prevention is more reliable than cleanup.

Can I upload a contract if I turn off AI training?

Not automatically. A training opt-out addresses only one use of future content. You still need permission from the relevant party, a suitable account, appropriate retention and sharing controls, and a reason for every part of the document you upload.

Does a paid AI plan automatically protect my data better?

No. Some plans have different terms and controls, but price alone does not determine privacy. Check the specific product, account type, settings, contract, and your employer’s policy.

Should I delete old chats or change my habits first?

Do both, but start with the habit. Deleting old chats reduces what has accumulated; data minimization stops unnecessary sensitive material from accumulating next week.