Microsoft 365 AI Search: Are OneDrive Files Used to Train?

Updated 5 November 2026 · Standard formula, shown below

No: Microsoft says data from Microsoft 365 Copilot—including prompts, responses, and information it accesses from Microsoft Graph, such as files in OneDrive—is not used to train the foundation models that power its generative AI services. That does not mean the files are invisible to AI search: when an authorized user asks a question, Microsoft 365 Copilot may retrieve relevant content to answer it, subject to the user’s existing permissions.

What “AI search” means in Microsoft 365

People use “Microsoft 365 AI search” to refer to several related experiences, including Microsoft 365 Copilot, Copilot Chat, and AI-assisted search or answers within Microsoft 365. The exact features available depend on the product, license, configuration, and how an organization has deployed them. The important distinction is between using a file to answer a request and using it as training material for a general-purpose model.

For example, an employee might ask Copilot to summarize a project document stored in OneDrive. The service can search for and use that document as context for that employee’s answer. This is retrieval at the time of the request; it is not the same as adding the document to a model’s training set so the model learns from it for future users.

How Copilot can use OneDrive content

Microsoft 365 Copilot can ground responses in work information available to the signed-in user through Microsoft Graph. Depending on the request and enabled features, that information may include files in OneDrive and SharePoint, emails, calendar details, chats, and other organizational content. A user’s prompt determines what the assistant is being asked to find or do.

In practical terms, a question such as “What were the main decisions in the budget review document?” may lead Copilot to locate a relevant document the user can access, extract useful passages, and generate a summary. The document informs that response, but Microsoft states that Microsoft 365 customer data is not used to train the foundation models.

Retrieval: Finding relevant information available to the user for a particular request.
Grounding: Using retrieved information as context for a response, sometimes with references or citations.
Training: Using examples to change or develop a model’s learned parameters. Microsoft says Microsoft 365 Copilot prompts, responses, and Graph data are not used to train its foundation models.

What the no-training statement does—and does not—cover

Microsoft’s statement addresses training of the foundation models that power Microsoft 365 Copilot. It should not be stretched into a claim that no data is ever processed, stored, or managed. To provide the service, Copilot processes prompts and relevant context; Microsoft 365 also has service records and compliance, security, and administrative controls whose exact behavior depends on the service and settings.

Nor does the statement mean that a private file can never appear in an AI-generated answer. If an account can access a file and asks a question that calls for it, Copilot may use the file as context. “Not used to train” is about model training, not a promise that the content will never be retrieved, summarized, or displayed to the person who is allowed to access it.

Microsoft’s commercial data protection terms and product documentation can differ from consumer Copilot terms. Check the terms for the specific account and Copilot experience in use, especially if switching between a work account and a personal account. Product features and policies can change, so current Microsoft documentation and an organization’s administrator are the best sources for account-specific details.

Permissions are still the key privacy control

Copilot is designed to respect the permissions of the signed-in user: it should not make a file available to someone who lacks access merely because that person asks an AI question. But this does not fix overly broad or mistaken permissions. If a document has been shared with a large group, or a folder has inherited access unexpectedly, Copilot may be able to surface that document to users who already have those rights.

This is why organizations should treat AI readiness as an information-governance task, not just a model-training question. Useful checks include:

Review OneDrive and SharePoint sharing links, group membership, and inherited permissions.
Remove access that is no longer needed and confirm sensitive files are shared only with intended people.
Use available sensitivity labels, retention policies, and data-loss-prevention controls as appropriate to the organization.
Ask an administrator how Copilot access, audit, and compliance settings are configured for the specific tenant.

Personal OneDrive versus work OneDrive

“Private OneDrive file” can mean a file in a personal Microsoft account or a work or school OneDrive account. These are not interchangeable contexts. Microsoft 365 Copilot for an organization is associated with the work or school identity and the organization’s data boundaries and policies. A personal Copilot experience may have different terms and data handling rules.

Before entering a prompt or opening a Copilot feature, check which account is signed in and which product is being used. If a work document is in a personal account, or a personal document is in a work account, the relevant access and privacy expectations may differ. Do not assume that the phrase “Microsoft Copilot” alone identifies the applicable protections.

What users and administrators can do

Individual users generally do not need to move files to a special location to prevent foundation-model training under Microsoft’s stated Microsoft 365 Copilot policy. The more useful steps are to keep permissions accurate and avoid sharing sensitive content with people who should not see it. Users can also avoid asking Copilot to retrieve information they do not need for a task.

Administrators should review Microsoft’s current documentation for the precise Copilot product, licensing, and tenant settings in use. They can assess oversharing, apply existing information-protection policies, communicate acceptable-use expectations, and verify how retention and audit requirements apply. If a feature or integration connects an additional service, review that service’s own data terms too.

Frequently Asked Questions

Can Microsoft 365 Copilot read my OneDrive files?
It may retrieve and use files that the signed-in user is authorized to access when they are relevant to a request and the feature supports that access. It does not mean Copilot indiscriminately reads every file.

Does Copilot use my OneDrive files to train ChatGPT or other models?
Microsoft says Microsoft 365 Copilot prompts, responses, and Microsoft Graph data are not used to train the foundation models that power its generative AI services. Check the terms for any separate consumer product or third-party service.

Can a coworker get my private file by asking Copilot?
Copilot is intended to respect the coworker’s existing permissions. If that person already has access because of a shared link, group, or inherited permission, the file may be available to them; review sharing settings if that is not intended.

Are Copilot prompts and responses stored?
Processing and record handling depend on the specific Copilot experience and Microsoft 365 settings. The no-training statement does not itself mean that prompts and responses are never retained. Consult current product documentation and your administrator’s policies.

Does this apply to personal OneDrive?
Do not assume that work-account protections apply identically to a personal Microsoft account. Check the privacy terms for the specific Copilot product and account you are using.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.