Domain 1 of 4 · Chapter 2 of 2

Identify responsible AI and data protection practices

Risks in relying on Copilot responses

Suppose you ask Copilot to summarize last quarter's supplier policy and get three tidy paragraphs with numbered citations. Nothing in that polish tells you whether the paragraphs are true. Microsoft's application card for Microsoft Copilot[1] says plainly that AI can make mistakes: Copilot could fabricate content that might sound reasonable but is factually inaccurate, and even when it draws on high-authority web data, its response might misrepresent that content in a way that is not completely accurate, up to date, or reliable.

This topic follows the skill Identify responsible AI and data protection practices in the Microsoft AB-730 study guide for Exam AB-730 (skills measured)[2]: common risks, verification steps, risks to sensitive data, and how data protection restricts prompt results. The same study guide lists "Understand how Copilot works to keep your organization’s information private and secure" under Understand generative AI capabilities across Microsoft 365 experiences (Understand generative AI fundamentals); see Understand Copilot capabilities across Microsoft 365 experiences.

Inaccuracy

Grounding means providing the language model with input sources related to the prompt, such as data from Microsoft Graph or Bing, so that Copilot may deliver more accurate and relevant responses, as the application card's key terms[1] define it. Grounding lowers the risk without removing it: the same card explains that when summarizing content from various sources, Copilot may include information that isn't present in its input sources, which Microsoft calls ungrounded results. A confident tone is not evidence.

Repetition is not evidence either. The application card[1] describes AI services as fallible and probabilistic and lists inconsistent outputs under similar prompts among the suboptimal outcomes Microsoft evaluates for, so getting the same answer twice tells you nothing about whether the sources support it.

Bias and stereotypes

Despite training by the model providers and Microsoft's responsible AI controls on both prompts and outputs, the application card[1] warns that it is challenging to comprehensively block all inappropriate content, which can lead to biases, stereotypes, or ungroundedness in generated content. A response can quote its sources accurately and still frame people unfairly, so a fairness review asks a different question from a citation check. In Microsoft's responsible AI overview[3], fairness means AI systems should treat everyone fairly and avoid affecting similar groups differently; a fairness review of a Copilot answer checks its recommendations and wording against that standard before they influence decisions about people.

Over-reliance

Over-reliance (Microsoft's documentation spells it overreliance) happens when users accept incorrect or incomplete AI outputs, mainly because mistakes in AI outputs may be hard to detect, according to the application card's end-user guidance[1]; its consequences range from decreased productivity and loss of trust to financial, psychological, or physical harm. Microsoft mitigates it with disclaimers and AI disclosure, for example a notice in Word that AI-generated content may contain errors, but the same card says users should still review the accuracy of the answers. The rule behind this comes from Microsoft's accountability principle: AI systems should not be the final authority on decisions that affect people's lives, and humans keep meaningful control, as the responsible AI overview[3] puts it.

Inaccuracy and bias are properties of the output; over-reliance is a behavior of the reader.

Prompt injection: instructions that try to coerce Copilot

Prompt injection is an attempt to coerce Copilot into doing something malicious, and Microsoft's prompt defense in depth article[4] distinguishes two kinds:

  • Direct: the user types a prompt designed to coerce Copilot into doing something malicious.
  • Indirect: the user's prompt refers to instructions that try to coerce Copilot, and those instructions sit in text files, images, code, files, or encoded text, not in the prompt itself.

Who is involved

Three parties meet in an indirect attack: an attacker who controls some content, you as the person asking Copilot for help, and Copilot reading on your behalf. The attacker needs no access to your account, only content that Copilot will read; inbound email is the documented example, which is why Microsoft Defender for Office 365 Plan 2 detects prompt injection content in inbound email before it reaches a user or an AI assistant, per the prompt defense in depth article[4]. The risk matters whenever Copilot reads content you did not write, such as an external email, a shared document, or a web page. When it fires, the hidden text tries to steer the task away from what you asked, for example toward disclosing information or adding a recipient.

Microsoft's layered defenses

Jailbreak attacks are prompts designed to bypass Copilot's safeguards or induce non-compliant behavior, and to mitigate them Copilot uses jailbreak and cross-prompt injection attack (XPIA) classifiers[5] that analyze inputs to the service and help block high-risk prompts before the model runs, but Microsoft notes that these classifiers may not be available in all Copilot scenarios. The privacy article[5] files these classifiers under "Does Copilot block prompt injections (jailbreak attacks)?" and does not say which one handles the direct or the indirect kind, so read their names as Microsoft's labels, not as a mapping onto that split. The prompt defense in depth article[4] lists further protections, among others: spam mail and chats are excluded from grounding, and certain hidden Unicode instructions in copied and pasted text are detected and sanitized. Separately, administrators can configure a Microsoft Purview Data Loss Prevention (DLP) rule that keeps email from external senders out of Copilot grounding to reduce the risk of prompt injection (DLP for Microsoft 365 Copilot[6]); that rule is covered in DLP rules for Copilot: conditions and outcomes below.

Your part: compare actions with the task

None of these defenses replaces your review. Microsoft Copilot Cowork, the agent that carries out multi-step tasks for you across Microsoft 365, presents an approval dialog with a preview before sensitive actions such as sending, posting, or scheduling, and its application card[7] tells users to verify recipients, dates, and content before approving and to pause or cancel when a task heads in the wrong direction. Treat that preview as the moment to check content and destinations, not as confirmation that the work is done.

The working rule: instructions found inside retrieved content are data to evaluate, never authority to change your task or disclose information.

Choosing verification steps that fit the task

Microsoft's guidance is general: users should review the responses Copilot generates and verify that they match their expectations and requirements, per the application card[1]. The Microsoft AB-730 study guide for Exam AB-730 (skills measured)[2] states the skill as "Select verification steps appropriate to the task, including citation checks and human review." Beside those two named checks, the table adds two that rest on Microsoft's guidance to double-check facts before making decisions or acting on a response (application card[1]) and to review, edit, and verify anything Copilot creates in Excel (Get data insights with Copilot in Excel[8]):

Check What it catches Use it when
Citation check A claim its source does not support A response cites documents or web pages
Currency and authority check A faithful summary of the wrong source The answer will guide a current policy or business decision
Data and calculation check A result its source data does not support Copilot produced an analysis, total, or chart
Human review Unsuitable, incomplete, or unfair output The output reaches customers or affects people

Citation checks

Where possible, Copilot responses based on business documents include references to the sources so users can verify the response, and users are advised to check the source materials, according to the application card's safety mitigations[1]. A citation shows where to check a claim; it does not prove the claim. Open the cited item and confirm it supports the specific claim, with the same qualifications and scope. A link that opens proves only that the source exists and that you can reach it. Separate the statements a source supports from additions it does not, then revise or drop the additions. Two features sit beside the citation check without being one of the four checks: query citations extend it for answers that use web search by showing what Copilot searched for, and Safe Links is a separate security check on where a link leads.

Web search query citations

When an answer uses public web content, Copilot generates a short search query from your prompt and sends it to the Bing search service. Web search query citations[9] in the linked citation section of a response show the exact queries sent to Bing, along with the sites searched. They are available only in Microsoft Copilot Chat, not in the Copilot pane within an app such as Word or PowerPoint, and the queries are available in the chat thread for only 24 hours. Source citations show what an answer rests on; query citations show what Copilot looked for, which helps you judge whether it searched for the right thing.

Safe Links checks destinations, not claims

Administrators with Microsoft Defender for Office 365 can configure Safe Links so that links returned by Copilot are evaluated at the time you select them, helping identify malicious destinations (prompt defense in depth[4]). That verdict answers "is this destination dangerous?", not "does this page support the answer?". A safe page can still be outdated or irrelevant, so destination safety and evidentiary support stay two separate checks.

Currency and authority

Responses drawn even from high-authority web data might not be completely accurate, up to date, or reliable, and Microsoft encourages users to double-check facts before making decisions or acting on a response (application card[1]). So check the source's date, owner, and applicability as well as its wording: an accurate summary of a superseded travel policy, or of another organization's inapplicable public rule, is still the wrong evidence for today's decision. When sources conflict, settle the conflict with the authoritative source, not with whichever generated summary reads best.

Data and calculations

For Copilot in Excel, Microsoft's advice is to review, edit, and verify anything Copilot creates for you, as with any AI-generated content, and when Copilot summarizes text such as survey results into a workbook, you can hover over the superscript number in the analysis to see the source data Copilot referenced (Get data insights with Copilot in Excel[8]). For a derived result, check it against its source data, not only against the narrative or the chart title. Two generated descriptions that agree with each other have not been checked against the data.

Human review

Human oversight is an important safeguard because outputs may be inaccurate, incomplete, biased, misaligned, or irrelevant to your goals (application card[1]). The same card asks deployers and users to exercise caution and evaluate outcomes for consequential decisions, meaning decisions with a legal or significant impact on a person's access to education, employment, financial services, government benefits, healthcare, housing, insurance, or legal services, and to take particular care in sensitive domains such as financial services, healthcare, and housing. A citation check establishes evidence; an accountable reviewer decides whether the output is fit to publish or act on.

If a response in Teams is inaccurate, harmful, or inappropriate, Microsoft asks you to select thumbs down in the summary and add comments in the feedback form (Copilot in Teams FAQ[10]). That feedback helps improve the service; it does not correct the draft in front of you, which you still fix or withhold.

Scale the applicable checks with the consequences: even a quick internal note needs review appropriate to its content, while customer-facing guidance or a consequential decision about a person calls for particular care and human review.

Reducing sensitive-data risk in what you share

The Microsoft AB-730 study guide for Exam AB-730 (skills measured)[2] lists "Recognize and mitigate risks to sensitive data" under the skill Identify responsible AI and data protection practices. One fact frames both what you expose to Copilot and how widely its sources and its outputs are shared: Copilot only surfaces organizational data to which individual users have at least view permissions (Data, Privacy, and Security for Microsoft Copilot[5]), the permission boundary that Item-level controls on what Copilot can use below explains in full, so Copilot neither bypasses a permission nor repairs a bad one.

Oversharing is a source problem

"Sharing" has two senses here: sharing a file with people sets who can open it, while sharing data with Copilot means putting it into a prompt. Oversharing is the first kind. Copilot can make a file shared with far more people than intended easier to discover without any permission bypass, but access alone does not guarantee discovery through Copilot. Microsoft's data protection architecture overview[11] lists the controls that address this, including sharing and membership controls that affect how broadly content can be accessed and governance and lifecycle policies that reduce long-term oversharing risk. When you notice content available to the wrong audience, raise it with the content owner or an administrator so the sharing gets corrected at the source; a cleverer prompt cannot hide a broadly shared document. Check the intended audience before you forward a Copilot summary, too: being able to read the sources does not make every recipient an appropriate audience.

Share only what the task needs

Microsoft Purview DLP for Copilot prompts exists so organizations can keep sensitive information out of prompts (prompt defense in depth[4]), but it checks only the text you type: DLP can't scan the contents of files you upload directly into a prompt (DLP for Microsoft 365 Copilot[6]). If a task needs only regional totals, attaching a full customer record adds personal data that this prompt check never inspects (DLP for Microsoft 365 Copilot[6]). Omit unneeded sensitive fields, or use an approved redacted version, before you attach anything.

Use the organizational Copilot experience

The application card[1] states that its information applies to Copilot used in organizations where users are signed in with their work or school account, and it points individuals signed in with a personal Microsoft account to a separate transparency note. Enterprise data protection, Microsoft's name for the controls and commitments under the Data Protection Addendum and Product Terms, applies to Copilot and Copilot Chat as used by organizations (enterprise data protection[12]). A similarly named app signed in with a personal account is documented separately, so do not assume it carries the same commitments; check which account you are signed in with before pasting work material, and keep minimizing what you share even inside the protected experience.

Check an agent before you give it data

Copilot can reference third-party tools and services through agents, and Microsoft advises that when you use agents, you check the agent's privacy statement and terms of use to determine how it will handle your organization's data (Data, Privacy, and Security for Microsoft Copilot[5]). A user can access only the agents that an admin allows and that the user installed or was assigned, per the same article. That availability is an admin decision about the agent, not approval of every disclosure you might make to it.

Temporary chat is not a retention exemption

Starting a temporary chat lets you chat without using or updating memory; the conversation won't appear in your chat history and can't be referenced or accessed later by you or Copilot (personalize what Copilot remembers[13]). The same article warns that your organization's data retention policy might require retaining temporary chat data, and that your IT administrator might be able to access it during the retention period. Use temporary chat for a memory-free conversation, not as a channel for data you are not allowed to submit.

These practices reduce unnecessary sharing and support responsible handling of data; the next sections cover the controls that limit what Copilot may do with what it can reach.

Two kinds of control restrict Copilot results

Every control that changes a Copilot result acts either on an item Copilot wants to use or on the interaction itself.

Microsoft's prompt defense in depth article[4] describes a typical Copilot interaction as five stages:

  1. Prompt input: you enter a prompt or attach content.
  2. Prompt ingress: the prompt is sent to the Copilot orchestrator, the engine that coordinates Copilot's components (privacy article[5]).
  3. Grounding: content is retrieved from documents, messages, meetings, and other Microsoft 365 data.
  4. Web search: if enabled, web content is retrieved.
  5. Response egress: the response is displayed.

Controls attach to specific stages, and they fall into two kinds:

  • Item-level controls act at the grounding stage and decide whether Copilot may use a particular file, email, or site. The main ones are Microsoft 365 permissions; encryption, through your usage rights (what the item's encryption lets you do with its content once you can open it, such as copying text) and its protection method; DLP label exclusions (DLP rules that keep items with chosen sensitivity labels out of processing); and Restricted Content Discovery (a SharePoint setting that hides a site from organization-wide discovery).
  • Interaction-level controls act on the prompt and the response, among them an Endpoint DLP paste block, which stops sensitive content from being pasted into browser-based apps at prompt input; a DLP prompt rule, which blocks a prompt containing chosen sensitive information types (patterns such as credit card numbers) at ingress; a DLP web search rule, which keeps such a prompt away from external web search only; and content safety filters (content harm filters[5] in Microsoft's privacy article), which check prompts at ingress and generated content before it is displayed (prompt defense in depth[4]).

DLP appears in both kinds. The Copilot DLP policy location supports a condition on sensitivity labels, which is a property of an item, and a condition on sensitive information types in the typed prompt, which is a property of the interaction, and one rule cannot use both conditions (DLP for Microsoft 365 Copilot[6]). A third condition, Email is received from > External users (preview), is item-level: it excludes an external email from grounding, summarization, and citation (DLP for Microsoft 365 Copilot[6]). The figure below places each control at the stage where it acts.

Whichever kind of control applies, every Copilot prompt runs in the security context of the user who initiates it (DLP for Microsoft 365 Copilot[6]); Item-level controls on what Copilot can use explains that permission boundary.

When a result is missing or partial, ask first whether something blocked an item or the interaction, then name the specific control.

Stage 1Prompt inputInteraction-level:Endpoint DLPpaste blockStage 2Prompt ingressInteraction-level:DLP prompt rule,content safetyfiltersStage 3GroundingItem-level:permissions,encryption,DLP label exclusion,Restricted ContentDiscovery, othersStage 4Web searchInteraction-level:DLP web search ruleStage 5Response egressInteraction-level:content safetyfilters
Microsoft's five Copilot prompt stages, with the item-level and interaction-level controls that act at each stage

Item-level controls on what Copilot can use

This section covers the item-level controls that decide whether Copilot may use a particular file, email, or site, starting with the one every other control builds on.

Permissions come first

Copilot presents only data that each individual can access, using the same underlying access controls as other Microsoft 365 services, and its grounding process only accesses content that the current user is authorized to access (Data, Privacy, and Security for Microsoft Copilot[5]). Microsoft Graph, which the application card[1] describes as the gateway to data and intelligence in Microsoft 365, is how Copilot reaches that content, not a way around the boundary. Naming or linking a SharePoint document you cannot open therefore does not let Copilot summarize it, and rewording the prompt changes nothing; if you need the content, request access through its owner.

Encryption usage rights: VIEW is not enough

When data is encrypted by Microsoft Purview Information Protection, Copilot honors the usage rights granted to the user (Data, Privacy, and Security for Microsoft Copilot[5]). Usage rights are the individual permissions encryption grants on an item, such as viewing, forwarding, printing, or copying text. Unlike the Microsoft 365 permissions above, which decide whether you can reach the item at all, usage rights come with the item's encryption and decide what you may do with its content, such as copy from it. Copilot needs two of them: VIEW and EXTRACT, the usage right whose common name is Copy and which determines whether Copilot can display text from encrypted content (Purview considerations for Copilot[14]).

If content grants you VIEW but not EXTRACT, you can read it yourself, but you can't use Copilot while it is open in an app, and Copilot won't summarize it; Copilot can still reference it with a link so you can open it outside Copilot (Purview considerations for Copilot[14]). A "Do Not Forward" email is the familiar example: per the same article, it omits the FORWARD right and does not let you copy text from it. Whoever applied the encryption always holds EXTRACT as the Rights Management owner, so content you encrypted yourself stays eligible to be returned to you.

Encryption can also be applied without a sensitivity label, for example by Microsoft Purview Message Encryption or Information Rights Management (IRM). That encryption still comes from the Azure Rights Management service, so Copilot still checks VIEW and EXTRACT before returning data and links; the difference is that new items don't automatically inherit the protection (Purview for Microsoft 365 Copilot[15]).

Protection methods differ

Encryption asks a second question besides usage rights: which protection method was applied, because the methods are not interchangeable. S/MIME and password protection, for example, are encryption methods that don't use the Azure Rights Management service (Purview for Microsoft 365 Copilot[15]). Files that a user has open in an app session are what Microsoft calls data in use (Purview considerations for Copilot[14]).

Protection method What Copilot does
Double Key Encryption (DKE) Can't access it: DKE items aren't returned, and you can't use Copilot in the app while a DKE item is open (Purview considerations for Copilot[14])
Microsoft Purview Customer Key, or your own root key (BYOK) Supported: as with eDiscovery and search, these items are eligible to be returned (Purview for Microsoft 365 Copilot[15])
Password-protected document Can't be accessed unless you already opened it in the same app (data in use); the password isn't inherited by a destination item (Purview for Microsoft 365 Copilot[15])
S/MIME-protected email Not returned, and Copilot isn't available in Outlook while one is open (Purview for Microsoft 365 Copilot[15])

DKE is intended for the most sensitive data with the strictest protection requirements, which is the documented reason Copilot and agents can't access it (Purview considerations for Copilot[14]). "Encrypted" alone therefore predicts nothing: a Customer Key item stays usable while a DKE item is excluded.

Restricted Content Discovery hides, it does not revoke

Restricted Content Discovery is a SharePoint site-level setting that limits discovery of a site's content, including recently interacted files, in organization-wide search results and Copilot responses while permissions and governance are reviewed (Restrict discovery of SharePoint sites and content[16]). It doesn't change existing permissions or remove content from the Microsoft 365 search index, users who already have access can continue to open the content directly, and it doesn't affect experiences that work on content already in use, such as summarizing an open document. It also removes AI entry points such as the Copilot button from those sites, and it applies only to SharePoint sites, not OneDrive. A file missing from a broad Copilot answer may simply sit on a restricted site; that absence does not mean your access was revoked.

Each item-level control answers a different question (permissions: can you open it; encryption: do your usage rights and its protection method let Copilot use it; DLP label exclusion: is its label excluded; Restricted Content Discovery: is the site discoverable), so name the control before you predict the result; the DLP label exclusion and the external-email exclusion are covered in DLP rules for Copilot: conditions and outcomes below.

Sensitivity labels on sources and new content

This section covers what Microsoft Purview sensitivity labels do when Copilot reads labeled content and when it writes new content.

What a label is

A sensitivity label classifies a file or email and can apply protection settings such as encryption and content markings, which include headers, footers, and watermarks (Learn about sensitivity labels[17]). A label can also classify with no protection at all, giving users a visual mapping of the data's sensitivity. Because the label is stored in the item's metadata, it stays with the content wherever it is saved. Typing "Confidential" into a document body is not a label, and Microsoft warns not to confuse sensitivity labels with Outlook's built-in sensitivity levels, which indicate the sender's intention but can't provide data security (Learn about sensitivity labels[17]). Administrators set a priority order for labels, and a higher priority usually denotes more sensitive content with more restrictive permissions (Purview for Microsoft 365 Copilot[15]).

Labels Copilot reads

When a labeled file is open in Word, Excel, or PowerPoint, or a labeled email or event is open in Outlook, the app shows the label name and its configured markings. When the label applies encryption, you need the EXTRACT usage right as well as VIEW for Copilot to return the data, and this protection extends to data stored outside your Microsoft 365 tenant, such as local storage, network shares, and cloud storage, while it is open in an Office app (Purview for Microsoft 365 Copilot[15]). EXTRACT is the usage right that grants permission to copy text, and it determines whether Copilot can display text to you from encrypted content (Purview considerations for Copilot[14]).

Copilot Chat displays the sensitivity label for items listed in a response and its citations, and the latest response shows the highest-priority label from the data used (Purview for Microsoft 365 Copilot[15]). That badge informs you about sensitivity. Displaying a label in chat isn't the same as applying that label and its protection settings to every generated response or copied output (Microsoft Learn: data security in Copilot[18]), so check the label and protection of the item where you paste or share the text.

Labels Copilot writes: inheritance

When you use Copilot in Word, PowerPoint, or Outlook to create new content based on a labeled item, the new content automatically inherits the source label with its protection settings; when several files are used, the label with the highest priority is inherited (Purview for Microsoft 365 Copilot[15]). Examples of these creation flows include Draft with Copilot with Reference a file in Word and Create presentation from file in PowerPoint.

Two rules decide what happens when the destination already has a label or cannot take one. Unlike other automatic labeling, an inherited label replaces a lower-priority label even when someone applied it manually, but it never replaces a higher-priority label (Purview considerations for Copilot[14]). And when the inherited label can't be applied, for example because the destination is read-only, is already encrypted without the rights to change its label, or the label isn't published to the user, the text isn't added to the destination at all (Purview considerations for Copilot[14]), as the figure below shows. As with any automatic label, the user can replace an inherited label afterwards, and remove it unless mandatory labeling is in use (Purview for Microsoft 365 Copilot[15]).

Container labels stay on the container

Labels applied to containers, such as Teams, SharePoint sites, Microsoft 365 Groups, Loop workspaces, and Viva Engage, aren't inherited by the items inside them, so those items don't display the container label in Copilot and can't support label inheritance (Purview considerations for Copilot[14]). For example, channel messages summarized from a team labeled Confidential won't display that label in Copilot Chat. When a particular file needs persistent classification and protection, label the file itself.

A label applies any configured protection to what it is attached to: the source item, and the new item a supported creation flow produces, but a badge in chat, a pasted copy, or a container label does not automatically protect another item.

Copilot drafts in Word,PowerPoint, or Outlookfrom labeled sourcesTake the highest-prioritysource labelDestination already hasa higher-priority label?YesExisting labelstaysNoInherited labelcan be applied?YesNew item getsthe label andits protectionNoText is not added
Sensitivity label inheritance when Copilot in Word, PowerPoint, or Outlook drafts new content from labeled sources

DLP rules for Copilot: conditions and outcomes

This section covers Microsoft Purview Data Loss Prevention (DLP) as it applies to Copilot. DLP identifies sensitive items, monitors them, and helps protect against their leakage (Purview for Microsoft 365 Copilot[15]). For Copilot, administrators create a policy on the Microsoft 365 Copilot and Copilot Chat policy location, and each rule pairs a condition with an action (DLP for Microsoft 365 Copilot[6]). A sensitive information type (SIT) is a pattern DLP detects, such as credit card numbers, passport numbers, or Social Security numbers, including custom SITs that your organization defines. The location supports four condition-and-action pairs, each with a different outcome (DLP for Microsoft 365 Copilot[6]):

Condition Action What the user experiences
Content contains sensitivity labels Prevent Copilot from processing content Copilot doesn't use the content of the labeled file or email; the item can still appear in citations
Content contains SITs in the typed prompt Processing prompts No response; the prompt isn't used for internal or web searches
Content contains SITs in the typed prompt Performing web searches No external web search for that prompt; the response still uses allowed internal Microsoft 365 sources
Email is received from external users Prevent Copilot from processing content External email is excluded from grounding, summarization, and citation; the email itself stays accessible

DLP label exclusion

With the sensitivity-label condition, identified items still appear in the citations of a response, but Copilot doesn't access or use the content of the item in the response (DLP for Microsoft 365 Copilot[6]). A visible citation therefore does not prove the item's text was summarized. When such a file is open in Word, Excel, or PowerPoint, the Copilot skills in the app are disabled except for certain experiences that don't reference file content or don't use large language models, and because the policy is evaluated when the file opens, a label applied mid-session takes effect the next time the file is opened. Your access does not override the exclusion: prompts run as you, and the DLP rule then excludes the item from processing even though you can open it.

Sensitive information types in prompts: two different actions

The two SIT actions, which Two kinds of control restrict Copilot results calls the DLP web search rule and the DLP prompt rule, are easy to confuse, so keep their outcomes apart. The web search action blocks only external web search as a grounding source for that prompt, and Copilot continues to generate the response using allowed internal Microsoft 365 data sources (DLP for Microsoft 365 Copilot[6]). The prompt-processing action stops the response entirely and keeps the prompt out of both internal and web searches; the block message can vary, and during preview Word, Excel, and PowerPoint might not clearly identify the organizational policy. Microsoft documents the prompt-processing action as a preview feature that is rolling out to tenants, available in Microsoft 365 Copilot, Copilot Chat, Cowork, and Copilot in Word, Excel, and PowerPoint. A web-search restriction alone never means the whole prompt fails.

External email exclusion

A rule with the condition Email is received from > External users, documented as a preview, excludes email from senders outside the organization's accepted domains from grounding, summarization, and citation, while the user's access to the email itself isn't affected (DLP for Microsoft 365 Copilot[6]). The policy evaluates email metadata only, comparing the sender domain with the tenant's accepted domains, and doesn't inspect the body. An external email missing from a Copilot inbox summary was excluded because of where it came from, not because its content was found to be malicious. Microsoft positions this rule as a way to reduce the risk of prompt injection from untrusted content, which ties it back to Prompt injection: instructions that try to coerce Copilot above.

What prompt DLP does not inspect

DLP can't scan the contents of files that you upload directly into a prompt; it only checks the text you type into the prompt itself (DLP for Microsoft 365 Copilot[6]). A successful upload is therefore not a sensitive-information scan. Separately, Microsoft 365 Copilot Chat on the web supports Endpoint DLP capabilities that block paste of sensitive content and block files that carry a specified sensitivity label (Purview for Microsoft 365 Copilot[15]); those are different controls from the prompt-text check.

Name the DLP condition before you predict the outcome: labels exclude items, SIT rules act on the prompt (either all of it or only its web search), and the external-email rule looks only at the sender's domain.

Automatic filters on prompts and responses

This section covers controls that act on the content of prompts and responses rather than on which data Copilot may use. Copilot's content safety filters identify harmful content in prompts or generated responses, in categories such as Hate and Fairness, Sexual, Violence, and Self-harm (Data, Privacy, and Security for Microsoft Copilot[5]). Classifiers flag potentially harmful content, and flags lead to mitigations such as not returning generated content to the user or diverting the user to a different topic; when crisis content is detected, the system disengages and provides crisis support resources instead (application card[1]).

A safety block is not an access problem

You can hold full permission to a source file and still receive no answer, because the filter evaluates the content of the prompt or response, not your rights to the source. Requesting file access, changing a label, or asking for a DLP exception addresses a different layer and does not change a content safety outcome. The two kinds of review also differ: a content safety filter is automatic and acts before delivery, while the fairness review in Risks in relying on Copilot responses is a human judgment of an answer that was delivered.

Workplace-harm filters restrict certain uses

Some Copilot scenarios add filters against workplace harms, which Microsoft defines as inferences, judgments, or evaluations about an employee's performance, attitude, internal or emotional state, or personal characteristics based on their workplace communication, and Microsoft restricts generative AI from being used for these purposes (Data, Privacy, and Security for Microsoft Copilot[5]). Summarizing what a meeting discussed is a normal task; scoring a colleague's motivation from their messages is the restricted use, even though Copilot can read those messages.

Content safety filters judge content, and workplace-harm filters judge purpose; neither is a permission, a label, or a DLP rule.

Copyright and ownership of Copilot output

This section covers intellectual property in what Copilot generates. Microsoft doesn't claim ownership of the output of the service, but it also doesn't determine whether a customer's output is copyright protected or enforceable against other users, because generative AI may produce similar responses to similar prompts from multiple customers, who may then claim rights in the same or substantially similar content (Data, Privacy, and Security for Microsoft Copilot[5]).

Two protections sit around that position, and both have limits. Copilot provides detection for protected material, which includes text subject to copyright and code subject to licensing restrictions, but this detection may not be available in all Copilot scenarios (Data, Privacy, and Security for Microsoft Copilot[5]). And under Microsoft's commitment for commercial customers, if a third party sues a commercial customer for copyright infringement for using Copilot or its output, Microsoft will defend the customer and pay any adverse judgments or settlements, as long as the customer used the guardrails and content filters built into the products, per the same article; enterprise data protection[12] names this the Customer Copyright Commitment.

Neither the absence of a Microsoft ownership claim nor the commitment is a blanket rights clearance: review how you reuse generated output, especially when it closely resembles existing work.

Microsoft's six responsible AI principles

This section maps a business concern to the responsible AI principle that names it. As part of its commitment to responsible AI, Microsoft adheres to six core principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability (application card[1]). Microsoft's responsible AI overview[3] describes them as follows, with the inclusiveness entry from Microsoft's responsible AI guidance for agent design[19]:

Principle The concern it names
Fairness AI systems should treat everyone fairly and avoid affecting similar groups differently
Inclusiveness AI systems should empower all users and accommodate diverse needs
Reliability and safety Systems operate reliably, safely, and consistently, respond safely to unexpected conditions, and resist harmful manipulation
Privacy and security Protecting privacy and securing personal and business information
Transparency Helping stakeholders understand how and why AI systems work
Accountability AI systems are not the final authority on decisions that affect people's lives, and humans keep meaningful control

Applied to Copilot, the mapping looks like this. A concern about keeping customer records out of a prompt is privacy and security, not transparency. Copilot's AI disclosure and limitation notices, which the application card[1] describes as helping users understand the system's capabilities and limitations, are transparency in practice. A review for unequal treatment of people in a recommendation is fairness, and keeping a person responsible for a decision that Copilot only informed is accountability.

When a scenario describes a concern, match it to the principle by what is at risk (people's treatment, whether all users are served, dependable behavior, information, understanding, or responsibility), not by which word sounds closest.

How these practices show up in questions

Suppose a prompt that contains a credit card number still returns an answer from internal files: a DLP rule that blocks only web searches lets Copilot keep generating the response from allowed internal Microsoft 365 data sources (DLP for Microsoft 365 Copilot[6]), one of the look-alike outcomes the pairs below separate by mechanism.

  • Citation shown vs content used. Under a DLP rule that excludes a sensitivity label, an item can be cited while its content was never used (DLP rules for Copilot: conditions and outcomes). A tempting wrong answer treats the citation as proof of summarization.
  • Readable vs summarizable. VIEW lets you read an encrypted item; Copilot also needs EXTRACT (Item-level controls on what Copilot can use).
  • Hidden vs revoked. Restricted Content Discovery hides a site from organization-wide discovery and leaves permissions unchanged (Item-level controls on what Copilot can use).
  • Web search blocked vs prompt blocked. The web search action still returns an internal answer; the prompt-processing action returns nothing (DLP rules for Copilot: conditions and outcomes).
  • Label displayed vs label applied. Copilot Chat displays the highest-priority label; supported creation flows in Word, PowerPoint, and Outlook apply it to a new item (Sensitivity labels on sources and new content).
  • Container label vs item label. A team labeled Confidential does not label the files and messages inside it (Sensitivity labels on sources and new content).
  • Outlook sensitivity level vs sensitivity label. Outlook's built-in Confidential level signals intent and provides no data security (Sensitivity labels on sources and new content).
  • DKE vs Customer Key. Double Key Encryption excludes Copilot; Customer Key and BYOK items stay eligible (Item-level controls on what Copilot can use).
  • Temporary chat vs no retention. A temporary chat doesn't use or update memory and doesn't appear in your chat history, but your organization's retention policy might still require retaining its data (personalize what Copilot remembers[13]; Reducing sensitive-data risk in what you share).
  • Safety refusal vs access denial. A content safety outcome is not changed by requesting file access (Automatic filters on prompts and responses).
  • Safe destination vs supported claim. Safe Links checks where a link leads, not whether the page supports the answer (Choosing verification steps that fit the task).
  • Unexpected action vs finished task. An approval preview is the moment to check recipients and content against what you asked for (Prompt injection: instructions that try to coerce Copilot).

When a question describes a missing or partial result, identify the kind of control first (item-level or interaction-level, as set out in Two kinds of control restrict Copilot results), then the specific control.

Item-level vs interaction-level controls on a Copilot result

ControlControl typeWhat Copilot cannot doWhat still happens
Microsoft 365 permissionsItem-levelUse or show an item you lack at least view permission toItems you can view stay available; a prompt cannot widen access
Encryption: no EXTRACT usage rightItem-levelReturn the item's textCopilot can reference the item with a link you open outside Copilot
Encryption: Double Key EncryptionItem-levelReturn the item, or run in an app while a DKE item is openItems encrypted with Customer Key or your own key (BYOK) stay eligible
DLP label exclusionItem-levelUse the content of items carrying that labelThe item can still appear as a citation
Restricted Content DiscoveryItem-levelSurface the site's content in organization-wide search and Copilot discoveryPermissions are unchanged; summarizing an open document still works
DLP prompt rule (sensitive information types)Interaction-levelRespond to the prompt or use it for internal or web searchesThe block message can vary; during preview, Word, Excel, and PowerPoint might not clearly identify the organizational policy
DLP web search rule (sensitive information types)Interaction-levelSend the prompt to external web searchThe response still uses allowed internal Microsoft 365 sources
Content safety filtersInteraction-levelReturn content flagged as harmful, such as violence or self-harm contentCopilot can redirect the conversation to a different topic

Decision tree

Copilot skipped an item you expected?Can you openthe item yourself?NoMicrosoft 365permissionsYesNo EXTRACT right, or aprotection methodCopilot can't use?YesEncryptionNoIts label excludedby a DLP rule?YesDLP labelexclusionNoSite has RestrictedContent Discovery?YesRestrictedContentDiscoveryNoCheck other item-level controls,then interaction-level controls

Sharp facts the exam loves — give these one last read before exam day.

Cheat sheet

Sharp facts the exam loves — scan these before test day.

A fluent Copilot response can contain fabricated or unsupported information

Copilot can produce content that sounds reasonable but is factually incorrect, including claims not present in the sources used for grounding. It can also misrepresent a reliable source, so grounding reduces risk without guaranteeing accuracy. Treat a factual claim as something to verify against evidence, not something established by confident wording or polished presentation. Repeating an identical prompt need not produce identical wording or selected points: generation is probabilistic, and in a live meeting or chat the available conversation context also changes. Agreement with an earlier generated answer is therefore not a substitute for checking the underlying evidence.

Trap Using confident wording to rank an answer above contradictory source evidence, which measures fluency rather than factual support.

7 questions test this
Responsible AI safeguards do not eliminate bias from generated content

A response that Copilot has returned can still contain bias or stereotypes despite training and responsible AI controls. Human fairness review examines recommendations and language for unjustified differences in how people or groups are treated, not just grammatical quality or factual errors. Accurately quoted source material can still need this review before influencing people-related decisions. This assessment of returned content differs from an automatic safeguard withholding harmful output before delivery.

Trap Using a successful factual citation check as the entire fairness review, which tests source support without assessing unequal treatment or stereotyping.

7 questions test this
Over-reliance occurs when users accept AI output instead of exercising judgment

Over-reliance is accepting incorrect or incomplete AI output because it appears credible or its mistakes are hard to detect. Keep human responsibility for deciding whether to use the answer and whether its reasoning is adequate for the task. An AI assistant can support a decision, but it should not become the final authority for consequential decisions merely because it produced a recommendation.

Trap Using Copilot's recommendation as the approval decision itself, which substitutes generated advice for accountable human judgment.

7 questions test this
Copilot restricts certain employee judgments inferred from workplace communication

Microsoft restricts certain Copilot scenarios that infer or evaluate an employee's performance, attitude, emotional state, or personal characteristics from workplace communication. Summarizing what was discussed is different from judging a person's motivation or emotional state from their messages. Access to the communications does not make those unsupported people assessments an appropriate use of the system.

Trap Using a meeting summary to justify an inferred employee-attitude score, which converts discussion content into an unsupported personal assessment.

7 questions test this
Prompt injection tries to make untrusted instructions control the AI's behavior

Prompt injection attempts to redirect Copilot away from its intended instructions or safeguards. A direct attack appears in the user's prompt; an indirect attack is embedded in referenced material such as a document, email, image, or web content. Recognize instructions inside retrieved content as potentially untrusted data rather than as authority to change the task or disclose information.

Trap Treating instructions embedded in a retrieved policy attachment as authorized user directions, which elevates source content into control over the assistant.

7 questions test this
Unexpected AI actions require task-boundary review before approval

When Copilot or Cowork proposes an unexpected disclosure, recipient, or action after reading external material, compare it with the work you actually authorized before proceeding. Pause or cancel work that has gone in the wrong direction and review the source and proposed action; built-in injection protections reduce risk but do not replace that review. An action preview is the opportunity to check content and destinations, not merely a confirmation that the task has finished.

Trap Using factual correctness of a draft as the only approval check, which leaves an unexpected recipient or disclosure outside the review.

7 questions test this
Citation verification checks whether the source actually supports the claim

When verifying a sourced Copilot response, open the cited material and check that it supports the particular claim, including its qualifications and scope. A working link establishes access to a source, not that the generated interpretation is correct or that every claim has evidence. Separate supported statements from unsupported additions and revise or withhold additions that cannot be substantiated.

Trap Using the presence of a clickable citation as proof of the entire answer, which checks link existence rather than claim-level support.

7 questions test this
A source must be current and authoritative for the decision being made

When an answer will guide a current policy or business decision, check the source's authority, date, and applicability as well as the quoted content. An old policy or a public rule for another organization can be accurately summarized yet be the wrong evidence for your decision. If sources conflict, resolve the conflict using the applicable authoritative information rather than relying on the most fluent generated summary.

Trap Using a correctly quoted superseded policy for a current decision, which verifies faithful quotation but not whether the rule still applies.

7 questions test this
Human review should match the consequences and audience of the output

Before publishing customer-facing guidance or acting on consequential AI advice, have an appropriate human review factual accuracy, completeness, suitability, and applicable requirements. Sensitive domains such as healthcare, finance, legal matters, and employment require particular care and meaningful human oversight. Citation checks help establish evidence, while a responsible reviewer decides whether the output is appropriate to release or act on. In Teams, also report inaccurate, harmful, or inappropriate generated content with the response's thumbs-down button and relevant comments in the feedback form; reporting helps improve the service but does not replace correcting or withholding unsuitable output before use.

Trap Using a citation-only check as publication approval for customer guidance, which leaves suitability, completeness, and consequences unassessed.

7 questions test this
Derived AI conclusions need checks of both source data and reasoning

When Copilot produces an analysis, verify the underlying data and the steps that connect it to the conclusion. For a spreadsheet result, that includes checking the relevant figures, formula logic, units, and scope rather than checking only the final prose or chart. Accurate source values can still lead to a wrong conclusion if the calculation or interpretation is wrong; missing inputs can also make a plausible result incomplete.

Trap Using agreement between the chart title and narrative as validation of an analysis, which compares two generated representations without checking the data or calculation.

7 questions test this
Web-search query citations reveal what Copilot searched rather than just what it cited

When checking how public information entered an answer, inspect web-search query citations in Copilot Chat's linked citation section to see the exact generated Bing queries and sites searched. The generated queries are available in the chat thread for 24 hours and are not available in the Copilot pane in apps such as Word or PowerPoint. Source citations help verify claims, while query citations help explain retrieval choices and refine the search context.

Trap Using a cited article alone to establish the exact Bing query Copilot sent, which provides source content rather than the generated search terms.

7 questions test this

Microsoft Defender for Office 365 Safe Links can provide additional protection when users open links returned by Copilot, evaluating URLs at click time for malicious destinations. This is distinct from opening and reading a legitimate source to confirm that it supports an answer. A safe destination can contain irrelevant or outdated information, while a convincing answer does not itself establish that a linked destination is safe.

Trap Using a successful Safe Links check as factual validation of an answer, which evaluates destination safety rather than evidentiary support.

7 questions test this
Provide only the sensitive information needed for an authorized AI task

When sensitive information is unnecessary to the task, omit it or use an approved redacted or de-identified version before adding content to Copilot. Include only the relevant material instead of a full record containing unrelated personal or confidential details. Enterprise protections support authorized processing, but they do not remove the need to follow organizational rules for what may be submitted.

Trap Using an entire customer record for a request that needs only anonymized totals, which supplies unrelated personal information without improving the task.

6 questions test this
Use the organizational Copilot experience for work-data protection commitments

When working with organizational content, choose the approved Copilot experience signed in with the appropriate work or school account to obtain its organizational protection context. Enterprise data protection commitments do not establish that a personal-account session or another public AI service has the same controls. Verify the account and permitted use before supplying work material, and still minimize sensitive inputs. This account-and-policy check is separate from whether a given Chat license automatically retrieves work content or requires an attachment.

Trap Using a personal AI account because the product has a similar Copilot name, which substitutes branding for the organizational account and protection context.

5 questions test this
Overbroad source permissions remain an exposure risk even when Copilot enforces them correctly

Copilot can make information easier to discover when it is already broadly accessible, so overly permissive sharing can expose sensitive material without a permission bypass. When content is available to an unintended audience, involve its owner or the appropriate administrator to correct source permissions and sharing. Review the intended recipients before distributing a generated summary, because access available to its creator does not establish an appropriate audience for the output.

Trap Using a stronger prompt to hide a broadly shared source document, which leaves the underlying permissions and discovery risk unchanged.

6 questions test this
An agent's data handling must be checked before it receives organizational information

When using an agent, review its privacy statement, terms of use, and relevant data access before involving sensitive organizational information. Agents can connect to external tools and services, so their particular handling must be evaluated rather than inferred solely from their presence inside Copilot. Use an allowed agent only for data and actions permitted by organizational policy; availability is not a substitute for assessing the proposed disclosure.

Trap Using an agent's installation status as proof that every sensitive-data use is permitted, which bypasses the agent-specific data handling and task assessment.

5 questions test this
Temporary chat avoids personalization and user chat history without overriding organizational retention

When a conversation should neither use nor update personalized memory, start a temporary chat. Its content does not appear in your chat history and cannot be referenced or accessed later by you or Copilot. Your organization's retention policy can nevertheless require that temporary chat data be retained, and an IT administrator might access it during that period. Choose temporary chat for a memory-free conversation, not as permission to submit otherwise prohibited sensitive data or as a guarantee of immediate deletion.

Trap Using temporary chat as a guaranteed zero-retention channel for restricted records, which confuses exclusion from personalization with organizational data retention.

6 questions test this
Copilot can surface organizational data only within the current user's access boundary

Microsoft 365 Copilot uses existing Microsoft 365 permissions and surfaces organizational data only when the individual user has at least view access. Microsoft Graph and semantic indexing do not confer additional permissions, so naming or linking an inaccessible SharePoint document does not enable its summary. If access is legitimately needed, use the normal owner or access-management process; a different prompt cannot grant it.

Trap Using Microsoft Graph grounding to obtain an inaccessible SharePoint summary, which treats the retrieval gateway as an elevation of the user's permissions.

6 questions test this
Opening encrypted content is not sufficient if Copilot lacks the user's EXTRACT right

For content encrypted through Microsoft Purview Information Protection, Copilot generally needs the user's VIEW and EXTRACT usage rights to return its text. EXTRACT corresponds to copying or extracting content, so a user may be able to read an item while Copilot cannot summarize it. Copilot can still reference an eligible item with a link for the user to open outside Copilot; a link alone does not show that its protected text was processed.

Trap Using successful manual opening of an encrypted file as proof of Copilot summarization rights, which checks VIEW without checking EXTRACT.

6 questions test this
Restricted Content Discovery limits organizational discovery rather than changing permissions

SharePoint Restricted Content Discovery can keep a site's content out of organization-wide search and Copilot discovery while its governance is reviewed. Existing permissions remain unchanged, so authorized users can still access content directly, and the restriction does not block supported experiences that summarize an already open document. Missing results in broad discovery therefore do not necessarily mean the user has lost file access.

Trap Using absence from organization-wide Copilot results as proof that a file's permissions were revoked, which confuses discoverability with direct access.

6 questions test this
Encryption methods differ in whether Copilot can process the protected content

Double Key Encryption is intended for especially sensitive content, and Copilot cannot access DKE-protected items, including when one is open in an app. This differs from supported service encryption such as Microsoft Purview Customer Key or Azure Rights Management bring-your-own-key protection, which does not categorically exclude content from Copilot. Identify the protection method and applicable rights before concluding that all encrypted data is either available or unavailable.

Trap Using Customer Key protection as if it imposed the same AI exclusion as Double Key Encryption, which treats different encryption models as interchangeable.

7 questions test this
Password-protected files and S/MIME email have distinct Copilot support boundaries

Copilot cannot access a password-protected document unless the user has already opened it in the same app as data in use, and the password is not inherited by a generated destination item. S/MIME-protected emails are not returned by Copilot, and Copilot is unavailable in Outlook when an S/MIME-protected email is open. Opening an item therefore has different consequences depending on its protection method.

Trap Using the opened-password-document exception for S/MIME email, which applies one content-protection method's support boundary to another.

7 questions test this
A sensitivity label classifies content and enforces only its configured protections

Microsoft Purview sensitivity labels can classify a file or email, apply markings, and apply encryption and usage restrictions when configured to do so. Apply the organization's appropriate label through the supported app rather than merely typing a confidentiality notice into the content. A label can classify without encryption, and Outlook's built-in sensitivity levels are not equivalent to Purview sensitivity-label security. For an eligible protected file opened in a supported Office app, configured label encryption and usage restrictions still constrain Copilot even when the file is stored outside the Microsoft 365 tenant, such as on local storage, a network share, or other cloud storage.

Trap Using Outlook's built-in Confidential sensitivity level as encryption, which marks the sender's intent without providing Purview information protection.

5 questions test this
Supported Copilot creation flows inherit the highest-priority source sensitivity label

In supported creation flows in Word, PowerPoint, and Outlook, the destination's new content inherits the eligible source sensitivity label and its protection settings. With multiple sources, the highest-priority eligible label is used; inheritance can replace a lower-priority existing label, including one manually applied, but does not replace a higher-priority label with a lower one. If required inheritance cannot be applied, such as to a read-only destination, the protected text is not added rather than silently losing protection. This is application of protection to a destination item, not merely display of a label in a chat answer. When a source instead uses Azure Rights Management encryption without a sensitivity label, Copilot still checks the user's VIEW and EXTRACT usage rights, but new items do not automatically inherit that protection.

Trap Using a lower-priority label already on the destination to retain weaker protection, which overlooks supported inheritance from higher-priority source content.

6 questions test this
A sensitivity label displayed in chat provides context without protecting every copied output

Copilot Chat displays sensitivity context on sources and can show the highest-priority label from the content used in a response. That badge informs the reader; it is not the destination-item label inheritance performed by supported creation flows, nor does it automatically protect every response or pasted copy. Before sharing copied content, check the destination's actual label, protection, and audience rather than relying on the badge shown in chat.

Trap Using the chat response's Confidential badge as proof that a pasted email is encrypted, which confuses a sensitivity indicator with protection on the destination item.

6 questions test this
A site's or team's sensitivity label does not automatically label every item inside it

Sensitivity labels applied to containers such as Teams or SharePoint sites govern the container but are not automatically inherited by its individual files or messages. Copilot therefore cannot rely on the container label as if every retrieved item carried that label for display or content-label inheritance. Check and apply appropriate item-level labeling when a particular file needs persistent classification and protection.

Trap Using a team's Confidential container label as proof that a downloaded document carries the same item label, which mistakes container governance for persistent file labeling.

6 questions test this
DLP can prevent Copilot from processing labeled content even when the user can open it

An applicable Microsoft Purview DLP policy can exclude files and emails with specified sensitivity labels from Copilot processing. Identified items can still appear as citations or links, but their contents are not used in the generated response; content-dependent in-app capabilities can also be disabled for a blocked open file. Additional access permission or an explicit file reference does not override that processing restriction.

Trap Using a visible citation as proof that DLP-protected text was summarized, which confuses a reference to the item with processing its contents.

7 questions test this
Sensitive-data DLP can block a prompt entirely or restrict only its external web search

A configured sensitive-information-type rule can block external web search for a prompt while Copilot continues using allowed internal Microsoft 365 sources. A separate prompt-processing restriction prevents a response and use of that sensitive prompt for internal or external searches; this prompt-blocking capability is in preview and depends on rollout. These outcomes differ, so a web-search restriction does not by itself mean the entire prompt must fail.

Trap Using a web-search-only DLP restriction as proof that no internal response is possible, which assigns the broader prompt-blocking effect to a narrower control.

6 questions test this
External-email DLP exclusions can remove mail from Copilot grounding without hiding the email

The preview DLP control for external email can exclude received external messages from Copilot grounding, summarization, and citation while leaving the user's direct access to those messages unchanged. The control uses the sender domain relative to the organization's accepted domains, not inspection of the email body for malicious text. A readable email may therefore be absent from Copilot results because of its external origin, not because its content has been proven unsafe.

Trap Using an email's omission from Copilot as proof that malware was found in its body, which mistakes a sender-domain exclusion for content inspection.

7 questions test this
Prompt DLP checks typed text rather than scanning the contents of directly uploaded files

The Microsoft 365 Copilot and Copilot Chat DLP policy location checks the text entered into the prompt for configured sensitive information types; it does not scan the contents of files uploaded directly into that prompt for sensitive data. A prompt restriction therefore cannot be assumed to certify an attached document as safe to submit. Check the attachment's permitted use and sensitive content before providing it, and distinguish this prompt-text control from separate supported Endpoint DLP restrictions on pasting content or uploading labeled files.

Trap Using successful attachment submission as proof of a sensitive-information scan, which assigns uploaded-file inspection to a control that evaluates prompt text.

7 questions test this
Content-safety safeguards can restrict an answer even when its source is accessible

Copilot's automated responsible AI safeguards check prompts and generated responses for harmful content, including hate and unfairness, sexual content, violence, and self-harm content. A safeguard can withhold generated content or redirect the interaction before delivery even when the user has permission to access a source. This system-enforced safety outcome differs from human fairness review of a returned answer and from a missing file permission, an information-protection usage right, or a DLP restriction on sensitive data.

Trap Using a file-access request to remedy a harmful-content refusal, which changes source authorization rather than the response's safety boundary.

7 questions test this
Generated output still requires attention to intellectual-property conditions

Microsoft does not claim ownership of Copilot output, but that does not establish that the output is unique, copyright-protected, or free of third-party rights. Copilot provides protected-material detection in applicable scenarios, and Microsoft's Customer Copyright Commitment for commercial customers is conditional on using the built-in guardrails and content filters. Review appropriate use rather than treating either AI generation or the commitment as an unconditional rights clearance.

Trap Using Microsoft's lack of ownership as proof that no third-party rights can apply, which confuses the service provider's claim with all other intellectual-property claims.

7 questions test this
Responsible AI principles address different kinds of business risk

Microsoft's six responsible AI principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Protecting personal and business information maps to privacy and security, equitable treatment maps to fairness, and helping people understand capabilities and limitations maps to transparency. Reliability and safety concern dependable, safe behavior; inclusiveness concerns empowering people broadly, while accountability keeps responsibility and meaningful control with people and organizations.

Trap Using transparency as the primary principle for restricting disclosure of customer records, which substitutes explainability for privacy and security.

7 questions test this

References

  1. https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-application-card
  2. https://learn.microsoft.com/en-us/credentials/certifications/resources/study-guides/ab-730
  3. https://learn.microsoft.com/en-us/azure/machine-learning/concept-responsible-ai
  4. https://learn.microsoft.com/en-us/copilot/plan/copilot-prompt-defense-in-depth
  5. https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy
  6. https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about
  7. https://learn.microsoft.com/en-us/copilot/cowork/responsible-ai/copilot-cowork-application-card
  8. https://support.microsoft.com/en-us/office/identify-insights-with-copilot-in-excel-52d97339-86c0-431c-b46c-e7b07b2898dd
  9. https://learn.microsoft.com/en-us/copilot/plan/manage-public-web-access
  10. https://support.microsoft.com/en-us/office/frequently-asked-questions-about-copilot-in-microsoft-teams-e8737767-4087-4ae6-b1d8-10264152b05a
  11. https://learn.microsoft.com/en-us/copilot/plan/architecture-data-protection-auditing
  12. https://learn.microsoft.com/en-us/copilot/plan/enterprise-data-protection
  13. https://support.microsoft.com/en-us/microsoft-365-copilot/personalize-what-microsoft-365-copilot-remembers
  14. https://learn.microsoft.com/en-us/purview/ai-m365-copilot-considerations
  15. https://learn.microsoft.com/en-us/purview/ai-m365-copilot
  16. https://learn.microsoft.com/en-us/sharepoint/restricted-content-discovery
  17. https://learn.microsoft.com/en-us/purview/sensitivity-labels
  18. https://learn.microsoft.com/en-us/training/modules/get-ready-work-microsoft-365-copilot/3-data-security-copilot
  19. https://learn.microsoft.com/en-us/agents/design-guidelines/responsible-ai