The short answer
Keeping your data private in AI tools takes three steps: switch off training, use private chat modes, and keep sensitive work on business plans. In ChatGPT, turn off Improve the model for everyone under Settings, Data controls. In Claude, switch off Help Improve our AI models under Settings, Privacy. In Gemini, turn off Keep Activity in Gemini Apps Activity. Opt-outs only apply from then on, and feedback ratings can still be used.
What you need before you start
- A list of the AI tools you use, including the ones built into other apps, and which account you use for each: personal or work.
- Ten minutes per tool to find one setting and read one help page.
- A clear idea of what counts as sensitive for you. For most people that is client names and contracts, health and money details, passwords, unreleased plans and anything about colleagues.
Privacy in AI tools comes down to three separate questions. Training: can the maker use your chats and files to improve its models? Retention: how long does it keep them, even if it does not train on them? Control: who else, such as human reviewers or your employer's admins, can see them? A tool can be fine on one and poor on another, so check all three.
Step 1: switch off training in each tool
These are the settings as the makers document them in November 2026. Personal plans are mostly opted in by default.
| Tool | Default on personal plans | Where to switch it off |
|---|---|---|
| ChatGPT (OpenAI) | Used for training on Free, Go, Plus and Pro | Web: account menu, Settings, Data controls, Improve the model for everyone. Mobile: sidebar, profile icon, Settings, Data controls. Voice has a separate control. |
| Claude (Anthropic) | Can be used on Free, Pro, Max and Ultra unless you opt out | Settings, Privacy, Help Improve our AI models (claude.ai/settings/data-privacy-controls); same path on mobile from your name menu |
| Gemini (Google) | With Keep Activity on, activity helps improve Google services including training, and human reviewers read some chats | Gemini Apps Activity at myactivity.google.com/product/gemini, Keep Activity |
| Perplexity | AI data retention on by default on Free, Pro and Max | Account settings, Preferences, AI data retention |
| Vibe, formerly Le Chat (Mistral AI) | Used by default for individual users | Mobile: Settings, Account, Data & Account Controls, untick Enable data sharing |
| Grammarly | On by default for Free, Premium and single-user Pro | Free and Premium: account.grammarly.com/security/privacy; single-user Pro: account.grammarly.com/admin/data_settings. Toggle Product Improvement and Training |
| DeepL | Free version texts, including your corrections, are used to train models | No opt-out described for Free; paid plans delete texts after translation and never train on them |
| Microsoft Copilot (personal) | The updated app (since 18 August 2026) documents no training switch | Settings, Personalization covers memories, web search and ads personalisation; keep work data out until Microsoft documents training |
| Meta AI | Chats are used to personalise ads and content; no opt-out from training in the US | UK and EU: objection form in Meta's Privacy Center |
Two rules apply almost everywhere. Opting out works from that point on: Anthropic says it stops future use of new chats, not training already done, and Perplexity says its opt-out only applies from then on. And feedback can still be used: ChatGPT says thumbs up or down ratings can still train its models, Mistral AI says a thumbs up or down rating in Vibe lets it use that rating with the related input and output, and Gemini chats you send feedback on can be used even with Keep Activity off.
Step 2: use the private chat mode for one-off sensitive questions
- ChatGPT temporary chat: not in your history, creates no new memories (and by default does not use existing ones), not used for training, but may be kept for up to 30 days for safety.
- Claude incognito chat: outside a project, click the ghost icon in the upper right of a new chat. It is not saved to history, not used for training and does not use or add to memory, but it is kept for 30 days by default. Anthropic notes it opens in the previous chat experience, so Claude cannot create files or run code there, and once closed it cannot be reopened.
- Gemini temporary chats: not used to train Google's models or reviewed by humans for improvement, but kept with your account for 72 hours.
- Grok Private Chat: the ghost icon in the top right corner; SpaceXSI says Private Chat content is not included in model training and is deleted from its systems within 30 days.
None of these modes means "deleted instantly". Treat them as "not used for training", not as "never stored".
Step 3: strip out what the AI does not need
Most tasks work just as well with placeholders. Replace names with roles ("Client A", "the supplier"), remove account numbers and addresses, round exact figures if the point survives, and never paste passwords, API keys or ID numbers into any chat. If the task genuinely needs the real data, it belongs in a work plan your organisation controls, not a personal account.
Step 4: move work data to a work plan
Business plans are where the strong defaults are. ChatGPT Business, Enterprise and Edu content is not used for training by default. Claude Team and Enterprise content is not used for training by default, and Enterprise adds custom data retention. Google Workspace accounts get no human review and no training outside your domain without permission. Microsoft 365 Copilot with a work account runs under enterprise data protection, and prompts are not used to train foundation models. Perplexity Enterprise data is never used for training. For a ranked view, see the best private AI for sensitive data.
Worked example: getting help with a client contract safely
You want a plain-English summary of a 12-page supplier contract and a list of risky clauses, and you only have a personal Claude Pro account.
- Confirm Help Improve our AI models is off in Settings, Privacy.
- Copy the contract into a text file and replace the supplier's name with "Supplier", your company with "Customer", and remove signatures, addresses and bank details. Keep the clause numbers.
- Start an incognito chat with the ghost icon and paste the cleaned text with the review prompt below.
- Read the flagged clauses yourself in the original, and send anything that matters to whoever signs contracts. The AI's job here is to point, not to decide.
If your company has Claude Team, ChatGPT Business or a Workspace account with Gemini, use that instead and skip the redaction step only if your policy allows it.
Common mistakes and how to fix them
- Assuming paid means private. Plus, Pro, Max and Ultra are personal plans with training on by default. Fix: switch it off, or use a business plan.
- Thinking the opt-out cleans up the past. Fix: switch it off before you share anything sensitive, and delete old chats you regret.
- Rating answers in sensitive chats. Fix: skip the feedback buttons. In Gemini Notebook on a personal account, Google says content is not used to directly train its foundation models unless you give feedback.
- Connecting every app at once. Each connector gives the AI more to read. Fix: connect one account, check what it can see, then add the next.
- Forgetting the free translator. DeepL Free trains on what you paste, with no opt-out. Fix: use a paid plan for confidential documents.
Which tool to use when privacy matters most
For personal use, pick a tool with a clear switch and turn it off: ChatGPT, Claude and Perplexity all document one. For work, the answer is whichever business plan your company already pays for, because that is where contracts, admin controls and retention settings live. If you are choosing one, Claude Team or Enterprise is our pick for sensitive documents, and Vibe by Mistral AI suits teams that want EU hosting. Then read how to use AI at work safely, compare options in ChatGPT vs Claude, and find team-ready prompt sets in Yuzoki kits.
Prompts to copy
Redact before you paste
I will paste a document. Do not summarise it yet. List every piece of personal or confidential information in it (names, companies, addresses, account numbers, amounts, dates of birth) and suggest a neutral placeholder for each, such as Client A or Supplier. [paste text]
You get: A redaction checklist; works best in a private chat on a tool with training switched off
Review a contract without the names
This is a supplier contract with names removed. Summarise it in plain English in ten lines, then list any clauses on liability, termination, automatic renewal, payment terms or data handling that a small business should question, with the clause number and one sentence on why. [paste cleaned contract]
You get: A plain summary and the clauses to read yourself
Check a tool before you trust it
Search the web for the official privacy and data-use pages of [tool name]. Tell me, with links to the maker's own pages and their dates: whether my chats are used for training on [plan], where the setting is, how long chats are kept, and whether humans can review them. Say plainly if you could not find an official answer.
You get: A sourced privacy check of any new tool
Write a safe version of a sensitive email
Rewrite this email so it keeps its meaning but contains no names, figures or details that identify a person or company. Mark each replaced item in square brackets so I can put the real details back after. [paste email]
You get: A shareable draft you can complete outside the AI tool