Share via

Microsoft 365 CoPilot Privacy Policy – "Foundation LLM"

Eeeo334 0 Reputation points
2026-05-22T16:19:11.0766667+00:00

Microsoft 365 CoPilot privacy policy says "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLM." The term "foundation LLM" is then repeated throughout. Anyone know what this gobbledegook means? Is data being used to train alternative/secondary LLMs?

Microsoft Copilot | Microsoft 365 Copilot | Development
0 comments No comments

1 answer

Sort by: Most helpful
  1. AI answer

    2026-05-22T16:19:30.7133333+00:00

    In this context, “foundation LLM” means the underlying large language models that power Microsoft 365 Copilot (for example, the base models hosted via Azure OpenAI Service and other Microsoft‑hosted LLMs). These are the core, general‑purpose models that Copilot relies on before any grounding in Microsoft Graph data or user context.

    The key point of the statement

    “Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs”

    is:

    1. Prompts entered into Microsoft 365 Copilot.
    2. Copilot’s responses.
    3. Organizational data that Copilot accesses via Microsoft Graph (emails, documents, chats, etc.).

    are not used as training data to further train or improve those underlying foundation models.

    Additional clarifications from the documentation:

    • Microsoft 365 Copilot connects LLMs to organizational data via Microsoft Graph to generate contextual responses, but that data stays within the Microsoft 365 service boundary and is processed under existing privacy, security, and compliance commitments.
    • For Microsoft 365 Copilot and Microsoft 365 Copilot Chat, prompts, responses, and Graph data are explicitly stated as not used to train foundation models.
    • The same pattern applies in specialized Copilot scenarios such as Viva Glint: raw survey responses, prompts, and respondent attributes processed or stored there are also not used to train foundation LLMs.
    • When using external data via Microsoft 365 Copilot connectors, prompts, responses, and data accessed through Microsoft Graph likewise aren’t used to train foundation LLMs.

    The provided documentation does not describe any separate or “secondary” LLMs being trained on this data. It consistently states that user prompts, responses, and organizational data (including data from Viva Glint and from connectors) are not used to train the foundation LLMs that power Copilot.


    References:

    AI-generated content may be incorrect. Read our transparency notes for more information.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.