The AI training opt out that matters most is the plan your staff use, not the brand of assistant. On the four big assistants, business plans do not train on your chats by default. Personal plans can, unless someone changes a setting.
That gap is where small organizations get caught. A staff member opens a personal ChatGPT, Claude, Gemini or Copilot account. Then they paste in a donor list or a member email. Whether that text can train the vendor’s models depends on that one account’s settings.
The short version: move work to business plans where you can. On any personal account used for work, turn the training setting off. Tell staff that rating a reply can send the whole conversation to the vendor, even after an opt out. Use temporary or incognito chats for anything sensitive.
This post covers what each vendor says, plan by plan, and where each setting lives. It ends with a short check for your team. It is about what you type into an assistant. Keeping AI crawlers off your website is a separate control, covered in our post on Google’s shared AI training control.
What "training on your data" means here
When a vendor says it may use your content to improve its models, your prompts, files and replies can become training data. Each vendor explains the details on its own help pages, and the wording differs.
Three paths come up again and again.
- The main setting. A toggle that decides whether new conversations can be used.
- Feedback. Rating a reply can send the full conversation, on some plans even after you opt out.
- Safety review. Some vendors can use flagged chats for safety work.
An AI training opt out usually covers future chats only. Anthropic says data already in a training run, or in a trained model, stays there. Assume the same for others unless they say otherwise.
Personal plans vs business plans
Every vendor in this post draws the same line. Personal plans are where training can happen. Business, enterprise and education plans, and the developer APIs, do not train on your content by default.
That makes plan choice the first decision for a nonprofit or agency. A business plan takes the question off each staff member’s plate. A personal plan leaves the AI training opt out to whoever set up the account. Our post on checking an AI vendor’s usage policy during procurement covers the contract side.
AI training opt out settings by tool, as of November 2026
We checked each vendor’s own help pages on November 2, 2026. These settings change often, so treat the table as dated.
| Assistant | Personal plans | Business plans and API | Where to turn training off |
|---|---|---|---|
| ChatGPT (OpenAI) | Free, Plus and Pro may be used for training | Business, Enterprise, Edu and API not used by default | Settings, Data controls, “Improve the model for everyone,” or the Privacy Portal |
| Claude (Anthropic) | Free, Pro, Max and Ultra used only if you allow it, plus flagged chats | Claude for Work, API and Claude Gov not used by default | Settings, Privacy, “Help Improve our AI models” |
| Gemini app (Google) | Used to improve services, including training, when Keep Activity is on | Qualifying Workspace editions not used outside your domain without permission | Settings and help, Activity, turn off Keep Activity |
| Gemini API (Google) | Unpaid tier used to improve products | Paid tier not used to improve products | Use a paid-tier project |
| Microsoft Copilot | Personal accounts used unless excluded or opted out | Work and school accounts not used to train foundation models | Privacy settings in the Copilot app (see below) |
ChatGPT: one setting covers chats and Codex tasks
On Free, Plus and Pro, OpenAI says it may use your content to train its models. That is from its help article, "How your data is used to improve model performance".
To opt out, go to Settings, then Data controls, and turn off "Improve the model for everyone." Or select "Do not train on my content" in OpenAI’s Privacy Portal. OpenAI says either one covers ChatGPT conversations and Codex tasks. After you opt out, new conversations are not used.
Codex has one more switch. OpenAI says a separate "Include environments" setting controls training on full environments. The ChatGPT setting does not change it.
On business plans, OpenAI says it does not use ChatGPT Business, Enterprise, Edu or API content by default. API organization owners can choose to turn on data sharing. Our post on ChatGPT Enterprise and Edu admin changes covers more workspace settings.
Claude: a privacy toggle on personal plans
Anthropic’s Privacy Center covers Claude Free, Pro and Max in "Is my data used for model training?" That includes Claude Code used from those accounts. The page predates Anthropic’s October 30 plan rename. Ultra is the former Max 20x plan, which the page covers as Max, as we explained in the Ultra rename. Anthropic says it uses chats if you choose to allow it. It may also use chats flagged for safety review, or chats from programs you join.
On personal plans, Claude’s AI training opt out is under Settings, then Privacy. Anthropic’s settings article names the toggle "Help Improve our AI models" in the app. Turn it off, and Anthropic says new chats will not be used for future training. Stored chats also drop out of future training runs.
Anthropic says raw connector content, such as Google Drive files, is not included unless you copy it into the chat. With the setting on, de-identified chats may be kept for up to 5 years.
Anthropic’s commercial training article covers Claude for Work, the API and Claude Gov. By default, Anthropic says it will not train on their inputs or outputs. Nonprofits on Team and Enterprise fall under this, as our Claude for Nonprofits plugin post explains.
Gemini: Keep Activity decides it on personal accounts
Google’s Gemini Apps Privacy Hub says Google uses your activity to improve its services, including training generative AI models. Some chats are read by human reviewers. Google keeps reviewed chats for up to three years.
On personal accounts, Gemini’s AI training opt out is the Keep Activity setting. Google’s activity help page gives the steps. Go to gemini.google.com, open Settings and help, choose Activity, and turn it off. Google still keeps chats with your account for up to 72 hours to run the service.
Work and school accounts follow different rules. Google’s Workspace generative AI privacy hub covers business, education and public sector editions. It says their content is not reviewed by people or used for training outside your domain without permission. That hub excludes Gemini for Education, so schools should check those terms. See our post on Gemini Connected Apps on a work account for related settings.
Developers have a separate line. Google’s Gemini API terms say unpaid services use prompts and responses to improve Google products. Human reviewers may read them. On paid services, Google says it does not use them for that. Logging is a separate issue, covered in our post on Gemini API logging defaults.
Microsoft Copilot: personal accounts vs work accounts
Microsoft’s Privacy FAQ for Microsoft Copilot says it uses Copilot data for AI training. It lists who is excluded:
- People signed in with an organizational Entra ID account
- People not signed in, and signed-in users under 18
- Copilot users in Microsoft 365 Personal or Family apps
- Users in some countries, and anyone who opted out
There is a catch. That FAQ says it applies only to the older Copilot app. Its companion page names two toggles in Privacy settings: "Training on conversation activity" and "Training on voice conversations." Microsoft released an updated Copilot app on August 18, 2026. As of November 2, its privacy controls page for that app lists no training toggle. On the new app, check its settings yourself.
For work accounts, Microsoft’s Microsoft 365 Copilot privacy page is clear. It says prompts, responses and Microsoft Graph data "aren’t used to train foundation LLMs, including those used by Microsoft Copilot." That is a strong reason to have staff sign in with work accounts.
Feedback buttons can undo an AI training opt out
Staff miss this one. On personal ChatGPT plans, OpenAI says rating a reply means the whole conversation may be used for training. That holds even after you opt out. So a personal AI training opt out does not cover rated chats.
Anthropic says a thumbs rating stores the whole conversation for up to 5 years and may be used for training. That applies to personal and commercial plans. It removes your user ID first.
Business admins have a switch. On ChatGPT Business, Enterprise and Edu, OpenAI says feedback does not permit training. Admins can also turn feedback off in workspace settings. On Claude Team and Enterprise, owners can turn off the "Rate chats" setting. It sits under Organization settings, then Data and Privacy.
Temporary and incognito chats
Each assistant has a mode for chats that are not saved to history. Use it for one-off sensitive questions.
- ChatGPT temporary chats are not used to improve OpenAI models while they stay temporary. Saving one makes it a regular chat.
- Claude incognito chats are not used to improve Claude, Anthropic says, even if the training setting is on.
- Gemini temporary chats are not used to train Google’s models, Google says. It keeps them for 72 hours.
Tell staff not to rate sensitive chats in these modes either.
A 30-minute AI training opt out check for your team
You can run this in one staff meeting.
- List every AI assistant staff use for work, and whether each account is personal or business.
- Move regular users to a business plan where your budget allows.
- On each personal account still in use, apply the AI training opt out from the table above.
- On personal ChatGPT accounts that use Codex, check "Include environments" too.
- Ask staff not to rate chats that hold donor, member, student or client data.
- Show staff how to open a temporary or incognito chat.
- On business plans, decide whether to turn off feedback for the whole workspace.
- Recheck in six months, since vendor pages change.
For the records side of Gemini in Workspace, see our post on Google Vault and Gemini exports.
Frequently Asked Questions
Does ChatGPT train on my conversations?
On Free, Plus and Pro, OpenAI says it may, unless you turn off “Improve the model for everyone” or use the Privacy Portal. OpenAI says it does not use ChatGPT Business, Enterprise, Edu or API content by default.
Does Claude train on my chats?
On Free, Pro, Max and Ultra, Anthropic says it uses chats only if you allow it, plus chats flagged for safety review. On Claude for Work and the API, it does not train on inputs or outputs by default.
Does Gemini train on my chats?
On personal accounts, Google says it uses Gemini activity to improve its services, including training, when Keep Activity is on. For qualifying Workspace editions, it says content is not used for training outside your domain without permission.
Does Microsoft Copilot train on my chats?
Microsoft says it uses Copilot data from personal accounts for training, unless you are excluded or opt out. It says work and school account prompts are not used to train foundation models.
Does an AI training opt out delete what I already shared?
No. Vendors describe it as covering new chats. Anthropic says data already in training, or in trained models, stays there.
Does the thumbs up or down button matter?
Yes. OpenAI says rating a reply on a personal plan can send the whole conversation to training, even after you opt out. Anthropic stores rated conversations for up to 5 years.
Are temporary or incognito chats used for training?
OpenAI, Anthropic and Google each say their temporary or incognito chats are not used to train their models. Google keeps Gemini temporary chats for 72 hours.
Is the free Gemini API safe for client data?
Google’s terms say unpaid Gemini API services use prompts and responses to improve products, and human reviewers may read them. The paid tier does not use them for product improvement.
Is this the same as blocking AI crawlers on our website?
No. These settings cover what you type into an assistant. Website crawling for AI training is controlled through robots.txt, which we cover in our post on Google’s shared AI training control.
What is the simplest fix for a small nonprofit?
Put regular AI users on a business plan, turn off training on any personal account used for work, and tell staff not to rate chats that hold sensitive data.