Will They Train on Our Data?

Two more things to know up front:

• Training off doesn’t mean deleted. Every provider keeps your chats for some period, even with training off. That’s called retention, and it’s covered below.
• The thumbs up/down buttons are a loophole. On Claude, ChatGPT, and Gemini, rating a response can send that whole conversation to the vendor for improvement or training, even when your training setting is off. Don’t rate office chats.

A large language model (the “brain,” or model, behind a tool like Claude or ChatGPT) is trained ahead of time. The company feeds it an enormous amount of text over months, and the result is a fixed model. That model is then released and used by millions of people.

When you type into a chat or start a folder project, the model does not learn from you on the spot. It reads your message, and any file you attach or it opens in your project folder, only inside that conversation. Close the chat and start a new one, and the model doesn’t “remember” what you said, unless you’ve turned on a memory feature (more on that below).

So what does “your chats may be used for training” mean? It means the company may later include your conversations, usually with your name and account details removed, in the data it uses to build a future version of the model. What you shared doesn’t show up in a stranger’s chat tomorrow. But you’ve handed over control of where that information goes, for years. For an election office, that’s the problem: you can’t un-send something that turned out to be exempt.

Each table below comes from the vendor’s own pages. The date checked and the links are under each table.

Anthropic (Claude)

OpenAI (ChatGPT)

Microsoft (Copilot)

Microsoft renamed Microsoft 365 Copilot to Microsoft Copilot in 2026, and the free chat is now Microsoft Copilot Chat. Microsoft says the rename didn’t change any privacy terms. A tenant is your office’s private space in Microsoft’s cloud (your email, OneDrive, SharePoint, and Teams). When you’re signed in with your county work account, Copilot Chat runs under your organization’s Microsoft agreement. Your prompts and answers stay inside your county’s Microsoft 365 service and aren’t used to train Microsoft’s AI models. The exception is web search: short search queries can go to Microsoft’s Bing, which follows different rules (see Copilot web search below).

Google (Gemini)

These steps come from each vendor’s help pages, and you do them in your account’s Settings: Claude’s in the Claude desktop app, ChatGPT’s wherever your account menu shows Data controls. The setting follows your account, so it covers quick chats and folder projects alike. Vendors move settings around, so if a label doesn’t match exactly, look for the closest match in the same menu.

Claude (Free, Pro, Max, Ultra)

Chats can be used to train models if the setting is on. Check yours and turn it off.

1. Open the Claude desktop app and click your name to open the menu.
2. Choose Settings.
3. Choose Privacy.
4. Turn off Help Improve our AI models. (Anthropic’s help page uses this name; the switch in the app may be worded a little differently.)

Turning it off stops new chats from being used, and Anthropic says it will stop using your earlier chats in future training runs. For a one-off chat you don’t want saved, use an incognito chat (the ghost icon at the top right of a new chat). Incognito chats are never used for training. On the new Claude experience, incognito chats open in the older chat view, so Claude can’t make files or run code in them. Source: How do I change my model improvement privacy settings?

Claude Team or Enterprise: training is off by default. An owner can also turn off the rating buttons with the Rate chats setting under Organization settings → Data and Privacy.

ChatGPT (Free, Go, Plus, Pro)

1. Open chatgpt.com in your browser (or the ChatGPT app, if your account menu shows Data controls) and open your account menu.
2. Choose Settings, then Data controls.
3. Select Improve the model for everyone, turn it off, and select Done.

On a phone: open the sidebar, tap your profile icon, then Data controls, and turn off the same setting. Coding tasks in Codex follow the same switch, with a separate Include environments setting. For a one-off chat, use a Temporary Chat: it isn’t used for training and isn’t saved to your history. Source: Data controls FAQ

ChatGPT Business or Enterprise: not used for training by default.

Microsoft Copilot Chat (work account)

1. Sign in with your county work account, not a personal Microsoft account.
2. Look for the green shield next to the New Chat button. That shield means enterprise data protection is on: your prompts and answers stay inside your county’s Microsoft 365 and aren’t used to train Microsoft’s AI. The exception is web search: short search queries can go to Microsoft’s Bing, which follows different rules (see Copilot web search below).
3. No shield? Stop and ask your IT staff or provider before you use it for anything that isn’t public.

There’s no training switch to find, because work-account chats aren’t used to train the models. Source: Copilot Chat privacy and protections

Gemini

• Google Workspace (work account): your content isn’t used to train models outside your organization without permission.
• Personal Gemini: go to gemini.google.com → Settings & help → Activity, and turn off Keep Activity. Source: Gemini Apps privacy

Even with training off, every provider keeps your data for some period. Here are the basics as of our last check:

• Claude: deleted chats are removed from Anthropic’s systems within 30 days. Incognito chats are kept 30 days. Chats flagged by safety systems can be kept up to 2 years. With training on, data may be kept (de-identified) up to 5 years.
• ChatGPT: saved chats stay until you delete them; deleted chats are removed within 30 days. Temporary chats may be kept up to 30 days for safety.
• Microsoft Copilot: prompts and responses are stored inside your own Microsoft 365 tenant. Your administrators control retention with Microsoft Purview, the same way they do for email.
• Gemini in Workspace: your admins set retention (90 days to indefinite). Google Vault, the tool for retaining and searching data, comes with Business Plus and Enterprise, not Starter or Standard.

• Claude for Excel keeps its chat history in your browser, deletes inputs from Anthropic’s servers within 30 days, and doesn’t follow an Enterprise plan’s custom retention or audit logs. Enterprise sessions can appear in the Compliance API (in beta), but not in audit logs. (Claude for Excel)
• Cowork on a local computer: Anthropic says history from local Cowork sessions isn’t covered by its standard retention and can’t be centrally managed or deleted by admins. (Cowork on Team and Enterprise)
• Zero data retention (ZDR) customers: since June 9, 2026, Anthropic keeps prompts and outputs for its most capable “covered” models for 30 days for safety work, even under ZDR. (Covered model retention)
• Copilot web search: short search queries go to Bing with user and tenant IDs removed, under Microsoft’s consumer terms rather than your organization’s data agreement. In GCC and DoD tenants, web search is off until an admin turns it on. (Enterprise data protection)

• Anthropic: Consumer Terms · Commercial Terms · Privacy Policy
• OpenAI: Terms of Use · Privacy Policy · Enterprise privacy
• Microsoft: Copilot privacy · Data Protection Addendum
• Google: Workspace generative AI privacy hub · Gemini Apps privacy