ערוך

שתף באמצעות

Privacy, security, and responsible use of Copilot in Fabric

  • מאמר

Before your business starts using Copilot in Fabric, you may have questions about how it works, how it keeps your business data secure and adheres to privacy requirements, and how to use generative AI responsibly.

This article provides answers to common questions related to business data security and privacy to help your organization get started with Copilot in Fabric. The article Privacy, security, and responsible use for Copilot in Power BI (preview) provides an overview of Copilot in Power BI. Read on for details about Copilot for Fabric.

Note

Your business data is secure

  • Copilot features use Azure OpenAI Service, which is fully controlled by Microsoft. Your data isn't used to train models and isn't available to other customers.
  • You retain control over where your data is processed. Data processed by Copilot in Fabric stays within your tenant's geographic region, unless you explicitly allow data to be processed outside your region—for example, to let your users use Copilot when Azure OpenAI isn't available in your region or availability is limited due to high demand. Learn more about admin settings for Copilot.
  • Copilot does not store your data for abuse monitoring. To enhance privacy and trust, we’ve updated our approach to abuse monitoring: previously, we retained data from Copilot in Fabric, containing prompt inputs and outputs, for up to 30 days to check for abuse or misuse. Following customer feedback, we’ve eliminated this 30-day retention. Now, we no longer store prompt related data, demonstrating our unwavering commitment to your privacy and security.

Check Copilot outputs before you use them

  • Copilot responses can include inaccurate or low-quality content, so make sure to review outputs before you use them in your work.
  • People who can meaningfully evaluate the content's accuracy and appropriateness should review the outputs.
  • Today, Copilot features work best in the English language. Other languages may not perform as well.

Important

Review the supplemental preview terms for Fabric, which includes terms of use for Microsoft Generative AI Service Previews.

How Copilot works

In this article, Copilot refers to a range of generative AI features and capabilities in Fabric that are powered by Azure OpenAI Service.

In general, these features are designed to generate natural language, code, or other content based on:

(a) inputs you provide, and,

(b) grounding data that the feature has access to.

For example, Power BI, Data Factory, and data science offer Copilot chats where you can ask questions and get responses that are contextualized on your data. Copilot for Power BI can also create reports and other visualizations. Copilot for Data Factory can transform your data and explain what steps it has applied. Data science offers Copilot features outside of the chat pane, such as custom IPython magic commands in notebooks. Copilot chats may be added to other experiences in Fabric, along with other features that are powered by Azure OpenAI under the hood.

This information is sent to Azure OpenAI Service, where it's processed and an output is generated. Therefore, data processed by Azure OpenAI can include:

Grounding data may include a combination of dataset schema, specific data points, and other information relevant to the user's current task. Review each experience section for details on what data is accessible to Copilot features in that scenario.

Interactions with Copilot are specific to each user. This means that Copilot can only access data that the current user has permission to access, and its outputs are only visible to that user unless that user shares the output with others, such as sharing a generated Power BI report or generated code. Copilot doesn't use data from other users in the same tenant or other tenants.

Copilot uses Azure OpenAI—not the publicly available OpenAI services—to process all data, including user inputs, grounding data, and Copilot outputs. Copilot currently uses a combination of GPT models, including GPT 3.5. Microsoft hosts the OpenAI models in the Microsoft Azure environment, and the Service doesn't interact with any services by OpenAI, such as ChatGPT or the OpenAI API. Your data isn't used to train models and isn't available to other customers. Learn more about Azure OpenAI.

The Copilot process

These features follow the same general process:

  1. Copilot receives a prompt from a user. This prompt could be in the form of a question that a user types into a chat pane, or in the form of an action such as selecting a button that says "Create a report."
  2. Copilot preprocesses the prompt through an approach called grounding. Depending on the scenario, this might include retrieving relevant data such as dataset schema or chat history from the user's current session with Copilot. Grounding improves the specificity of the prompt, so the user gets responses that are relevant and actionable to their specific task. Data retrieval is scoped to data that is accessible to the authenticated user based on their permissions. See the section What data does Copilot use and how is it processed? in this article for more information.
  3. Copilot takes the response from Azure OpenAI and postprocesses it. Depending on the scenario, this postprocessing might include responsible AI checks, filtering with Azure content moderation, or additional business-specific constraints.
  4. Copilot returns a response to the user in the form of natural language, code, or other content. For example, a response might be in the form of a chat message or generated code, or it might be a contextually appropriate form such as a Power BI report or a Synapse notebook cell.
  5. The user reviews the response before using it. Copilot responses can include inaccurate or low-quality content, so it's important for subject matter experts to check outputs before using or sharing them.

Just as each experience in Fabric is built for certain scenarios and personas—from data engineers to data analysts—each Copilot feature in Fabric has also been built with unique scenarios and users in mind. For capabilities, intended uses, and limitations of each feature, review the section for the experience you're working in.

Definitions

Prompt or input

The text or action submitted to Copilot by a user. This could be in the form of a question that a user types into a chat pane, or in the form of an action such as selecting a button that says "Create a report."

Grounding

A preprocessing technique where Copilot retrieves additional data that's contextual to the user's prompt, and then sends that data along with the user's prompt to Azure OpenAI in order to generate a more relevant and actionable response.

Response or output

The content that Copilot returns to a user. For example, a response might be in the form of a chat message or generated code, or it might be contextually appropriate content such as a Power BI report or a Synapse notebook cell.

What data does Copilot use and how is it processed?

To generate a response, Copilot uses:

  • The user's prompt or input and, when appropriate,
  • Additional data that is retrieved through the grounding process.

This information is sent to Azure OpenAI Service, where it's processed and an output is generated. Therefore, data processed by Azure OpenAI can include:

  • The user's prompt or input.
  • Grounding data.
  • The AI response or output.

Grounding data may include a combination of dataset schema, specific data points, and other information relevant to the user's current task. Review each experience section for details on what data is accessible to Copilot features in that scenario.

Interactions with Copilot are specific to each user. This means that Copilot can only access data that the current user has permission to access, and its outputs are only visible to that user unless that user shares the output with others, such as sharing a generated Power BI report or generated code. Copilot doesn't use data from other users in the same tenant or other tenants.

Copilot uses Azure OpenAI—not OpenAI's publicly available services—to process all data, including user inputs, grounding data, and Copilot outputs. Copilot currently uses a combination of GPT models, including GPT 3.5. Microsoft hosts the OpenAI models in Microsoft's Azure environment and the Service doesn't interact with any services by OpenAI (for example, ChatGPT or the OpenAI API). Your data isn't used to train models and isn't available to other customers. Learn more about Azure OpenAI.

Data residency and compliance

You retain control over where your data is processed. Data processed by Copilot in Fabric stays within your tenant's geographic region, unless you explicitly allow data to be processed outside your region—for example, to let your users use Copilot when Azure OpenAI isn't available in your region or availability is limited due to high demand. (See where Azure OpenAI is currently available.)

To allow data to be processed elsewhere, your admin can turn on the setting Data sent to Azure OpenAI can be processed outside your tenant's geographic region, compliance boundary, or national cloud instance. Learn more about admin settings for Copilot.

What should I know to use Copilot responsibly?

Microsoft is committed to ensuring that our AI systems are guided by our AI principles and Responsible AI Standard. These principles include empowering our customers to use these systems effectively and in line with their intended uses. Our approach to responsible AI is continually evolving to proactively address emerging issues.

Copilot features in Fabric are built to meet the Responsible AI Standard, which means that they're reviewed by multidisciplinary teams for potential harms, and then refined to include mitigations for those harms.

Before you use Copilot, keep in mind the limitations of Copilot:

  • Copilot responses can include inaccurate or low-quality content, so make sure to review outputs before using them in your work.
  • People who are able to meaningfully evaluate the content's accuracy and appropriateness should review the outputs.
  • Currently, Copilot features work best in the English language. Other languages may not perform as well.

Copilot for Fabric workloads

Privacy, security, and responsible use for:

Related content