How to Protect Your Privacy When Using AI (Step by Step)
A guide to Phase 1 — Audit What You Actually Share, Phase 2 — Lock Down Accounts and Settings, Phase 3 — Handle Sensitive Material Safely, Phase 4 — Review, Delete and Build Lasting Habits
Every prompt you send is a small act of disclosure, and most people never check where it goes or how long it stays. You do not need to stop using AI to stay private; you need a short set of habits around what you paste, which settings are on and what you delete. This guide covers 28 concrete steps, from listing every tool that touches your data to writing a one-page personal rule you will actually follow. The guiding idea is minimisation: send the least sensitive, least identifying material that can still get the job done. Browse all illustrated guides or the AI Safety & Privacy category for more illustrated walkthroughs on Howcompass.
How do you protect Your Privacy When Using AI (Step by Step)?
Howcompass | How to Protect Your Privacy When Using AI (Step by Step)Phase 1 — Audit What You Actually Share
HowcompassList every AI tool you use
List every AI tool you use. Write down chat assistants, smart email features, editors, transcription services and browser extensions that send text to a model. You cannot protect an entry point you have not counted, and the overlooked plug-ins are usually the leaky ones. For the primary source behind this phase, see Mozilla — web and browser resources.
HowcompassClassify the data each one sees
Classify the data each one sees. Grade the material you feed each tool as public, internal, personal or confidential. Matching a tool to a sensitivity level makes it obvious when a confidential document is heading for a casual public chat.
HowcompassCheck what is stored and for how long
Check what is stored and for how long. Read the privacy terms for the retention period and whether conversations train the model, and note both in your list. 'We may review chats to improve quality' means a human could see what you wrote.
HowcompassLocate the delete and export controls
Locate the delete and export controls. Find the settings for deleting conversations and exporting your data, and bookmark their exact location. When you need to remove something urgently is the worst time to go hunting through menus.
HowcompassTrim each prompt to the minimum needed
Trim each prompt to the minimum needed. Include only the text required to complete the task and cut names, addresses and unrelated background. Shorter prompts are cheaper, often get better answers, and leave fewer fragments of your life on someone else's server.
HowcompassRedact identifiers before pasting
Redact identifiers before pasting. Replace real names, emails, phone numbers, account IDs and record numbers with clear placeholders like [CLIENT]. Keep the structure so the model still understands the task, but strip the details that identify a person.
HowcompassPrefer local or no-retention modes
Prefer local or no-retention modes. For anything sensitive, choose an on-device model, an enterprise plan with no training, or a session that promises zero retention. A public free chat box should be your default only for material you would happily publish. Related on Howcompass: start a blog from zero.
Phase 2 — Lock Down Accounts and Settings
HowcompassUse a distinct, strong password
Use a distinct, strong password. Give each AI account its own long password stored in a password manager instead of reusing your email password. Credential-stuffing attacks start with a breach somewhere else, and a unique password stops that breach reaching your chats. For the primary source behind this phase, see NIST — AI Risk Management Framework.
HowcompassTurn on two-factor authentication
Turn on two-factor authentication. Choose an authenticator app or a security key in preference to SMS, which can be intercepted through a SIM-swap attack. This is the single biggest improvement to account security for the fewest minutes of effort.
HowcompassReview connected apps and permissions
Review connected apps and permissions. Open the list of third-party apps linked to the account and revoke anything unused or over-permissioned, such as access to an entire cloud drive. Old integrations keep reading data long after you forget you approved them.
HowcompassSeparate work from personal use
Separate work from personal use. Keep distinct accounts for employment and personal tasks so a mix-up cannot send company material into a personal tool. Separation also lets you delete or audit each history on its own schedule. Browse all illustrated guides or the AI Safety & Privacy category for more.
HowcompassDisable training on your content
Disable training on your content. Switch off any option that lets your conversations improve the model, and check whether the choice applies to past data as well as future chats. If the provider gives you no such control, treat everything you type as potentially retained.
HowcompassChoose the shortest retention period
Choose the shortest retention period. Select the shortest available conversation-retention setting and decline optional human review of your content. A short window limits how much of your history exists to be exposed in a future breach.
HowcompassCheck where your data is processed
Check where your data is processed. Find out which country or region handles your prompts and attachments, since that affects which privacy laws and company rules apply. Some regulated material simply cannot be processed in certain locations. Related on Howcompass: make a monthly budget.
Phase 3 — Handle Sensitive Material Safely
HowcompassKeep regulated and secret data out entirely
Keep regulated and secret data out entirely. Never paste medical records, financial account numbers, source-code secrets or customer lists into a public AI tool. The time saved is trivial beside the cost of a breach or a compliance violation. For the primary source behind this phase, see Anthropic — AI research and documentation.
HowcompassAnonymise case studies and examples
Anonymise case studies and examples. Replace real clients with synthetic or heavily generalised examples, removing the combination of details that would let someone identify them. A job title, a city and an unusual revenue figure together can be enough to name a person.
HowcompassGet consent before uploading others' data
Get consent before uploading others' data. If material contains information about colleagues, clients or third parties, ask permission and state exactly what you will do with it. Do not make a privacy decision on someone else's behalf, even with good intentions.
HowcompassWrite a no-go list for your team
Write a no-go list for your team. Agree in writing which categories of data must never enter an AI tool and who may approve an exception. A shared list removes the guesswork that leads an eager colleague to paste a spreadsheet 'just to test it'.
HowcompassScan answers for leaked fragments
Scan answers for leaked fragments. Before sharing a generated answer, read it for real names, old clients or details that should not travel. Models can echo identifying material from a prompt or from context you did not realise was included.
HowcompassVerify facts before you rely on them
Verify facts before you rely on them. A privacy-conscious tool is not necessarily an accurate one, and a wrong conclusion about a person carries real consequences. Confirm any claim that affects someone against an authoritative source first.
HowcompassStore AI outputs at the right sensitivity
Store AI outputs at the right sensitivity. Save generated files with the same encryption, access controls and retention rules as the source material. A confidential summary left on a public share link undoes every precaution taken before it was written. Related on Howcompass: start a blog from zero.
Phase 4 — Review, Delete and Build Lasting Habits
HowcompassDelete old conversations monthly
Delete old conversations monthly. Set a recurring day to remove chat histories you no longer need and empty the trash if the tool keeps a separate one. A shrinking stored history is a smaller target if an account is ever compromised. For the primary source behind this phase, see USA.gov — official U.S. government services.
HowcompassRevoke access when a project ends
Revoke access when a project ends. As soon as work wraps up, remove the AI integrations, shared links and third-party permissions created for it. Unused access tends to persist indefinitely and is rarely reviewed by anyone.
HowcompassAudit sign-ins and activity logs
Audit sign-ins and activity logs. Periodically read the list of devices and times that accessed the account, and investigate anything unfamiliar immediately. Early detection turns a stolen session into a quick password reset rather than a long exposure.
HowcompassClose accounts you no longer use
Close accounts you no longer use. Before abandoning an old AI service, export anything you need and then delete the stored data and the account itself. Each retired account is one less place holding a copy of your conversations.
HowcompassWrite a one-page personal rule
Write a one-page personal rule. Condense everything into one page: what you may input, which mode to use and when you clean up. A short written rule survives busy weeks far better than a set of good intentions you cannot remember.
HowcompassShare results without permanent links
Share results without permanent links. When passing an AI result to someone, send a file or a link with an expiry date and a viewer allow-list. A permanent public link is how a useful summary quietly becomes a public document.
HowcompassRecheck settings after tool updates
Recheck settings after tool updates. Providers change defaults and terms, so spend a few minutes each quarter confirming that training, retention and sharing options have not reverted. A toggle you switched off last year may be on again after a redesign. Related on Howcompass: make a monthly budget.
Frequently Asked Questions
Question: Does deleting a conversation remove it everywhere?
Not necessarily. It usually disappears from your history, but provider backups or retained copies may persist for the stated retention window; check the policy and, where offered, request deletion of your stored data.
Question: Can I use AI safely at work?
Yes, if you follow your organisation's approved tools and rules, keep regulated or confidential data out of public services, and use an account with no training on your content. When in doubt, ask before pasting rather than after.
Question: What should I never put into an AI tool?
Passwords and API keys, full financial or medical records, government ID numbers, source-code secrets and identifiable customer lists. If a detail would embarrass you or breach a duty if it leaked, redact it or leave it out.
Warnings
- Pasting secrets into a public chat: A public chat box is not a private notebook; anything sensitive pasted there may be stored, reviewed or retained, so redact identifiers or use a local, no-retention mode instead.
- Assuming deleting a chat erases the data: Removing a conversation from your view does not always delete backups or provider copies; check the retention policy and the export-and-delete controls rather than trusting the trash icon.
- Reusing one account for everything: Mixing work, personal and experimental use in a single account makes over-sharing likely and cleanup impossible; separate accounts and revoke permissions when a project ends.
Related Howcompass Guides
References (4)
- Mozilla — web and browser resourcesmozilla.org
- NIST — AI Risk Management Frameworknist.gov
- Anthropic — AI research and documentationanthropic.com
- USA.gov — official U.S. government servicesusa.gov
Sources are listed so you can read the underlying guidance yourself. Our editorial policy explains how guides are reviewed.
Did this article help you?
0 readers found this helpful — add your vote.
No comments yet — be the first to share your results.