Guides › Do AI providers train on your data
Do ChatGPT, Claude, Copilot and Gemini Train on Your Data? The Answer by Plan
On business plans and APIs, no by default. OpenAI, Anthropic, Microsoft and Google each state that content from their commercial products is not used to train models unless the customer opts in. On consumer plans the answer is a setting. ChatGPT Free, Plus and Pro and personal Microsoft Copilot use conversations for training unless the user turns it off, Claude Free, Pro and Max use them when the user allows it, and the consumer Gemini app uses activity, with human review, while Keep Activity is on. The plan a person is signed in to decides the answer more than the provider does, so the exposure for most companies is an employee using a personal account. This page gives the answer for each plan, the exceptions that apply on business plans, and what happens to knowledge base and connector content.
- Business plans and APIs. Not used for training by default on ChatGPT Business, Enterprise and Edu, Claude Team and Enterprise, Microsoft 365 Copilot, Gemini for Workspace and the providers' APIs.
- Consumer plans. Training is on unless the user opts out (ChatGPT, personal Copilot), on when the user allows it (Claude), or tied to an activity setting (Gemini app).
- Exceptions on business plans. Feedback you submit, explicit opt-ins, and third-party agents and apps that carry their own terms.
- Connected knowledge. Content reached through connectors and knowledge features follows the plan's rule. On business plans it is excluded from training.
The Answer by Plan
Each row paraphrases the provider's published position as of November 2026. The sources are linked at the end of the page.
| Product and plan | Used for training by default? | What changes it |
|---|---|---|
| ChatGPT Free, Plus, Pro | Yes | Turning off the setting that shares conversations to improve the model stops it for future conversations. |
| ChatGPT Business, Enterprise, Edu, Healthcare, Teachers | No | Excluded by default under OpenAI's business terms. |
| OpenAI API | No | Not used since March 1, 2023 unless the organization explicitly opts in. |
| Claude Free, Pro, Max | Only if the user allows it | The model improvement setting. Incognito chats are never used. Chats flagged for safety review may be used to improve safety systems. |
| Claude Team, Enterprise, API | No | Feedback sent with the thumbs up or down buttons can be used. Owners can turn feedback off for the organization. |
| Microsoft Copilot, personal account | Yes | A training setting in Copilot's privacy controls lets the user opt out. |
| Microsoft 365 Copilot and Copilot Chat, work account | No | Prompts, responses and Microsoft Graph data are not used to train foundation models. |
| Gemini app, personal Google Account | Yes, while Keep Activity is on | Turning Keep Activity off, or using temporary chats. A subset of chats is read by human reviewers. |
| Gemini for Google Workspace, and the Gemini app with a work or school account | No | Not used for training or human review outside your domain without permission. |
| Azure OpenAI, Amazon Bedrock, Google Cloud | No | Azure OpenAI and Google Cloud state that prompts are not used to train models without permission. Amazon Bedrock runs models in AWS accounts that model providers cannot access. |
Why the Plan Decides the Answer
Each provider runs two contracts. Consumer plans sit under consumer terms and a privacy policy, where using conversations to improve models is either the default or a choice presented to the user. Business plans and APIs sit under commercial terms and a data processing addendum, where the provider acts as a processor and the training exclusion is a contractual commitment. A paid consumer subscription is still a consumer plan. ChatGPT Plus and Pro and Claude Pro and Max do not carry business data terms.
That makes the account the thing to check. A person signed in to a workspace account on a managed device and the same person signed in to a personal account in another browser tab are under different rules, and no admin setting in the workspace reaches the personal account. The plan-by-plan safety page and the page on shadow AI cover that gap.
Exceptions That Apply on Business Plans
Feedback you send
Rating a response is a submission. Anthropic states that when a user gives feedback with the thumbs up or down buttons, it stores the whole related conversation for up to five years, de-links it from user and customer IDs, and may use it to train models. Owners of Team and Enterprise plans can turn the feedback buttons off for the organization. Microsoft says optional feedback on Copilot is used to improve the product and is not used to train foundation models, and Google says Workspace feedback improves its products without training the models behind Workspace generative AI.
Opt-ins
OpenAI offers an explicit opt-in to share API data through the dashboard, and Anthropic's commercial terms allow use where the customer chooses to permit it. These are choices someone in the organization can make, so they belong in the admin review.
Safety review
On its consumer plans, Anthropic states that conversations flagged for safety review may be used to improve its ability to detect and enforce its usage policy, including training models used by its safeguards team. Safety systems screen content on every plan. What differs on business plans is what the content may be used for afterward.
Third-party agents, apps and connectors
The provider's commitment covers the provider. A custom GPT action, an MCP server or an agent published by another company receives what it is sent under that company's terms. Microsoft notes that agents in Copilot carry their own terms of use and privacy statements, which admins can review before allowing them.
What About Knowledge Bases and Connected Data?
Content an assistant reaches through a connector or a knowledge feature follows the same rule as the conversation on that plan. OpenAI states that workspace content, including data accessed through apps, is excluded from training by default. Microsoft's commitment names data accessed through Microsoft Graph alongside prompts and responses. Google's covers Workspace content and webpage context. Anthropic states that connector data is not used for training, and adds a detail that matters on consumer plans: raw content from connectors is not included in training data when the user has allowed model improvement, though anything copied from a connected source into the conversation becomes part of the conversation.
Training is the narrowest of the questions a knowledge base raises. Connected content is still retrieved, sent to the model, stored with the conversation and governed by whatever permissions the source system has. The knowledge features guide and the connector security guide cover those.
Opting Out Does Not Reach Backward
Turning training off stops future use. Anthropic states that data already included in training runs in progress, or in models already trained, stays there, and that with model improvement allowed it may keep de-identified data for up to five years. Google states that chats already seen by human reviewers are kept for up to three years and are not deleted when the user deletes their activity. For a company, this means personal-account use discovered after the fact cannot be fully undone, which is the case for preventing it.
How to Verify It for Your Organization
- Inventory the accounts in use.Sanctioned workspaces, API organizations and the personal accounts your network or expense data shows.
- Confirm the plan behind each.Business, Enterprise, Edu or API under commercial terms, with a signed data processing addendum.
- Get the exclusion as a contract term.A default setting can change. A term in the agreement is what the review should cite.
- Turn off feedback submission where the platform allows it.Feedback is the common path from a business plan into training.
- Review opt-ins, agents and connected apps.Check the API dashboard sharing setting and the third-party tools enabled in each workspace.
- Write the tier into the acceptable use policy.Name the approved tools and plans, and state that personal accounts are not approved for company data.
The acceptable use policy template has wording for the last step.
Training Is One of Five Questions
A training exclusion says what the provider will not learn from. It says nothing about how long the content is kept, who at the provider can read it, where it is processed or what your own users were allowed to retrieve. Those are covered in the provider policy comparison, the retention periods page and the ZDR explainer.
Where Hardshell Sits
Hardshell addresses what is sent, which sits upstream of every provider policy. It connects enterprise data to ChatGPT, Claude, Microsoft 365 Copilot, Gemini and internal assistants without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before content reaches the provider, and records every retrieval under opaque IDs. Whatever a provider does with content afterward, the identifiers were not in it. Hardshell deploys as a cloud service, in a private cloud or as an offline bundle.
Frequently Asked Questions
Does ChatGPT train on my data?
On Free, Plus and Pro, yes by default, unless you turn off the setting that shares conversations for model improvement. On ChatGPT Business, Enterprise, Edu and Healthcare and on the API, OpenAI states that it does not train on your data by default.
Does Claude train on my conversations?
On Claude Free, Pro and Max, only if you allow it in the model improvement setting, and never for Incognito chats. On Claude Team, Enterprise and the API, Anthropic does not train on inputs or outputs by default. Feedback submitted with the rating buttons is the exception on every plan.
Does Microsoft Copilot train on company data?
Not when users are signed in with a work account. Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models. The personal version of Copilot does use conversations for training unless the user opts out.
Does Gemini train on Google Workspace data?
No. Google states that Workspace customer data, prompts and generated responses are not used to train models or reviewed by humans outside your domain without permission. The consumer Gemini app is different: activity is used to improve Google's models, with human review, while Keep Activity is on.
Do AI providers train on API data?
Not by default. OpenAI has excluded API data from training since March 2023 unless a customer opts in, Anthropic excludes commercial API inputs and outputs, and Azure OpenAI and Google Cloud state that prompts are not used to train models without permission.
If I opt out of training, is my earlier data removed?
Generally not. Opt-outs apply going forward. Anthropic states that data already in a training run or a trained model remains, and Google keeps chats that human reviewers have already seen for up to three years.
Related guides
- What AI providers do with your data · Training, retention, residency and admin controls side by side.
- Is it safe to put company data in ChatGPT, Claude or Copilot? · A plan-by-plan answer with the conditions attached.
- AI data retention periods by plan · How long each provider keeps conversations, files and knowledge.
Sources (accessed November 2026)
OpenAI, Enterprise privacy at OpenAI; Data controls in the OpenAI platform; OpenAI Help Center, How your data is used to improve model performance. Anthropic Privacy Center, Is my data used for model training? (consumer); Is my data used for model training? (commercial); How long do you store my data?. Microsoft Learn, Data, privacy, and security for Microsoft 365 Copilot; Microsoft Support, Privacy FAQ for Microsoft Copilot. Google, Generative AI in Google Workspace privacy hub; Gemini Apps Privacy Hub. Cloud platforms: Azure, Amazon Bedrock, Google Cloud.