Privacy settings pass: 15 minutes, once a quarter

Use it when it is a new quarter, you have started using another assistant, or a vendor tells you its terms have changed.

Not when you want to know what is safe to paste. That is What not to paste.

You need the apps you use, and a password manager if you have one. Fifteen minutes should be enough for three assistants (our estimate). Do the pass below for each one, using its section further down for where to click.

The pass

#	Step	Time	Check
1	List the assistants you use and how you sign in to each	2 min	Every assistant is on the list, including ones built into other apps.
2	Training. Find the setting that lets your chats improve the model (the AI system behind the app), and choose on purpose	3 min	You can say for each one whether it is on or off. If it fails: note which one you could not find, and use its section below.
3	History. Set the shortest retention (how long chats are kept) you can live with, and find the temporary or incognito mode	2 min	You know how long a normal chat and a temporary chat are kept.
4	Memory. Open what the assistant remembers about you, and delete anything you would not want quoted back. Include anything you imported from another assistant: Gemini can import memories and whole chats, and Claude can import memories	3 min	You have read every saved item, imported ones too.
5	Connectors and apps (links that let the assistant read or act in other services, such as your email). List what is connected, remove what you do not use, and choose the most cautious permission the tool offers	3 min	Each remaining connector has a reason you can state.
6	Feedback buttons, keys and limits. Learn what rating a reply does, keep API keys (secret codes that let a program use your account) in a password manager, and set a spend limit where one is offered	1 min	Your keys are not in a notes file.
7	Put the next pass in your calendar	1 min	It has a date.

Where to click, by assistant (read 2026-10-04 to 2026-10-10)

Find the one or two you use and skip the rest. Menus move, and when yours looks different the vendor's page wins. We state a default only where the vendor does.

Claude (Free, Pro, Max)

Training and retention: Where to look: Settings > Privacy, then the "Help Improve our AI models" toggle. Your chats are used for training only if you allow it. Existing users were asked in a pop-up in 2025 (deadline 8 Oct 2025), and new users choose when they sign up. Allowing it extends retention to five years, otherwise 30 days. Incognito chats are never used for training. Claude for Work, Government, Education and API use are excluded.
Memory: Where to look: Settings > Memory (or Settings > Capabilities). Memory from chat history is available on all plans, including free, and is on by default for Free, Pro and Max. Each project has its own memory. When a chat expires or is deleted, memories generated from it are not removed, but you can delete individual memories at any time. Incognito chats are not saved to history or memory, and are kept for 30 days by default.
Connectors: Not read-only. Claude's Gmail connector can search, read, draft, send, reply to and forward emails, and by default it asks for approval before each action. On Team and Enterprise plans an owner can set its write tools (the ones that send or change things) to Blocked for everyone under Customize > Connectors > Tool permissions, and individual users can turn off specific tools from the chat. It sees attachment metadata and not attachment contents (October 2026).

ChatGPT

Training and retention: Where to look: Settings > Data controls > Improve the model for everyone (phone: sidebar > profile icon > Settings > Data controls). Turn off "Improve the model for everyone" if you do not want chats used for training. Temporary chats may be kept up to 30 days, unless you save them, and are not used for training; a saved temporary chat follows your model-improvement setting. Rating a reply can send the whole chat to training. Business, Enterprise and Edu workspaces are excluded by default, and API data is kept up to 30 days.
Memory: Where to look: Settings > Personalization > Memory. Logs of deleted saved memories may be kept for up to 30 days for safety and debugging. In OpenAI's words, "Deleting the original chat does not automatically delete a separate saved memory".
Apps: ChatGPT apps have no read-only mode. In Settings, under Apps or Plugins as available, the permission levels are "Always ask", "Allow read actions", "Allow low-risk actions" and "Allow all actions", and OpenAI says the last carries elevated risk. Apps can take actions such as creating or updating information (October 2026).

Gemini

Training and retention: Where to look: Settings & help > Activity, then turn Keep Activity off. Gemini Apps train on chats by default if you are 18 or over (the Keep Activity setting), and auto-delete them after 18 months (you can choose 3 or 36 months, or never). Human-reviewed chats can be kept up to 3 years. With Keep Activity off, chats are kept up to 72 hours, and temporary chats are not used for training.
Memory: Where to look: Settings & help > Personal Intelligence, then the Memory toggle. Google said in August 2025 that memory of past chats would be on by default. The feature is now called Memory, and Google's help page lists age 18 or over, a personal Google Account and Keep Activity on as requirements. Temporary Chats are kept up to 72 hours.

Other assistants

Assistant	Training and retention	Where to look
Microsoft Copilot	You can switch off training on text and voice chats; opting out covers your future chats. Even then, Microsoft still uses chats for ads, safety and compliance. History is kept 18 months. Microsoft's page does not state the default, so check your own setting. It also says these controls apply to the older Copilot app, and a newer app has been out since 18 Aug 2026.	Microsoft's Copilot privacy controls page
Meta AI	From 16 Dec 2025, Meta uses AI chats to personalise ads in most regions, with no specific opt-out. Sensitive topics such as health and religion are excluded.	Meta's announcement
Perplexity	AI data retention is on by default, and opting out applies only to data collected afterwards. Enterprise data is never used for training.	Perplexity's data collection help page
Grok	xAI says "You control whether your data is used for training Grok", and turning it off keeps your new chats out of training. The default is not stated, so check your own setting. Deleted chats and Private Chats are removed within 30 days, except data kept for safety or legal reasons.	Grok app: Settings > Data Controls > "Improve the model". Grok.com: Settings > Data > "Improve the Model". On X: Settings > Privacy & Safety > Grok & Third-party Collaborators > Data Sharing
Mistral (Mistral Vibe, formerly Le Chat, renamed in August 2026)	Standard plans train on chats unless you opt out. The API has a separate toggle.	Phone app: Settings > Account > Data & Account Controls, then clear "Enable data sharing". Team admins: Manage > Vibe > Privacy
DeepSeek	DeepSeek trains on user data, stores it in China, asks users not to share sensitive personal data, and offers a training opt-out by email.	DeepSeek's privacy policy

Connectors and keys

Neither of the two connectors whose help pages we read (Claude's Gmail connector and ChatGPT apps, above) is read-only. Our suggestion is to start from the most cautious level, which for ChatGPT apps is "Always ask", and to block any write tool you do not need. For agents (AI tools that take actions for you) that read the web or your files, see Letting an agent near your email, browser and files.

If your provider lets you set a monthly spend limit on an API or agent account, set one before you let anything run unattended. We give no figure because limits and settings differ by provider and change.

How to tell it worked

For each assistant you use, you can say whether training is on or off, and how long a normal and a temporary chat are kept.
You have read every saved memory item and deleted the ones you would not want quoted back.
Every remaining connector has a reason you can state, and none has a write tool you do not need.
The next pass is in your calendar.

Sources

Anthropic. Announcement of updates to its consumer terms, Aug 2025, and Claude privacy page on model training. Vendor. Link · Link.
OpenAI. Data controls FAQ, enterprise privacy page and ChatGPT release notes. Vendor. Link · Link · Link.
Google. Gemini Apps privacy help page, Gemini activity help page, Memory help page, and announcement of temporary chats and privacy controls. Vendor. Link · Link · Link · Link.
Microsoft. Copilot privacy controls. Vendor. Link.
Meta. Announcement on improving recommendations with AI chats, Oct 2025. Vendor. Link.
Perplexity. Data collection help page. Vendor. Link.
X. Help page about Grok; xAI legal FAQ, read 2026-10-18. Vendor. Link · Link.
Mistral. Help articles on opting out of training, and on Le Chat becoming Vibe. Vendor. Link · Link.
DeepSeek. Privacy policy. Vendor. Link.
Anthropic Help. Claude memory, and incognito chats. Vendor. Link · Link.
OpenAI Help. Memory FAQ. Vendor. Link.
Anthropic Help. Two help articles on Claude's Gmail connector and its tool permissions. Vendor. Link · Link.
OpenAI Help. ChatGPT apps permissions. Vendor. Link.
Google Help. Import memory and chats to Gemini; Anthropic's memory import, as described by Tom's Guide, March 2026. Vendor and news. Link · Link

Changes

v1.010 Oct 2026first version

The pass

Where to click, by assistant (read 2026-10-04 to 2026-10-10)

Claude (Free, Pro, Max)

ChatGPT

Gemini

Other assistants

Connectors and keys

How to tell it worked

Sources

Changes

Related