Responsible AI Principles

What are responsible AI principles?

The six principles

Microsoft's six responsible AI principles

Feature	What It Means	How It Applies to Copilot
🎯 Fairness	AI should treat all people equitably	Copilot shouldn't produce biased recommendations based on gender, race, or other protected characteristics
🛡️ Reliability & Safety	AI should work correctly and not cause harm	Copilot should produce accurate responses; hallucinations are monitored and mitigated
🔒 Privacy & Security	AI should protect data and respect privacy	Copilot respects M365 permissions; data isn't used to train models; customer data stays in the tenant boundary
🔍 Transparency	AI should be understandable and explainable	Users know when they're interacting with AI; Copilot shows which sources it used
♿ Inclusiveness	AI should be accessible and work for everyone	Copilot supports accessibility features, multiple languages, and diverse user needs
📋 Accountability	People should be answerable for AI systems	Microsoft publishes impact assessments; organisations should have AI governance policies

Exam tip: The exam usually presents a scenario and asks “which responsible AI principle is being demonstrated?” Focus on the keywords: bias → Fairness, accuracy → Reliability, data handling → Privacy, explainability → Transparency, accessibility → Inclusiveness, oversight → Accountability.

Key Copilot design decisions driven by responsible AI

Decision	Principle
Copilot doesn’t use your data to train AI models	Privacy & Security
Copilot shows citations (sources it used)	Transparency
Copilot responses can be reviewed in audit logs	Accountability
Content filters prevent harmful/toxic outputs	Reliability & Safety
Copilot respects existing M365 permissions	Privacy & Security
Human review is recommended for important decisions	Accountability

💡 Scenario: Northwave's AI governance policy

After deploying Copilot, Alex (CEO) asks Maya to create an AI governance policy. Here’s what they develop, mapped to responsible AI:

Policy Rule	Principle
”Copilot outputs must be reviewed by a human before sending to customers”	Accountability
”We will audit Copilot usage quarterly for signs of bias”	Fairness
”Copilot web grounding is disabled — only internal data”	Privacy & Security
”All teams must be trained on Copilot’s limitations”	Transparency
”Copilot must be usable by employees with disabilities”	Inclusiveness
”We test Copilot responses in critical workflows before relying on them”	Reliability & Safety

What the exam specifically tests

The exam focuses on practical implications, not just definitions:

1. Copilot doesn’t train on your data — your organisational data stays private and is not used to improve the AI model
2. Copilot can hallucinate — it may produce inaccurate information, especially when data is incomplete. Human review is essential.
3. Copilot shows its work — citations link back to source documents so users can verify
4. Admin oversight exists — audit logs, usage reports, and DLP policies allow governance
5. Content safety filters — Copilot blocks harmful, toxic, or inappropriate outputs

🎬 Video walkthrough

Flashcards

Question

What are Microsoft's six responsible AI principles?

Click or press Enter to reveal answer

Answer

1) Fairness — equitable treatment. 2) Reliability & Safety — accurate, no harm. 3) Privacy & Security — protect data. 4) Transparency — explainable and understandable. 5) Inclusiveness — accessible to all. 6) Accountability — humans are answerable.

Click to flip back

Question

Does Microsoft use your organisational data to train Copilot's AI models?

Click or press Enter to reveal answer

Answer

No. Your data is NOT used to train, retrain, or improve the large language models. Copilot reads your data at query time through Microsoft Graph but doesn't learn from it. Prompts and responses are subject to your organisation's Microsoft 365 compliance, audit, and eDiscovery controls.

Click to flip back

Show hint

Question

Why does Copilot show citations in its responses?

Click or press Enter to reveal answer

Answer

Transparency — users can verify where the information came from by clicking the citation links. This allows them to check accuracy and confirm the sources, reducing blind trust in AI-generated content.

Click to flip back

Knowledge Check

Knowledge Check

Copilot generates a summary of a customer meeting that contains an inaccurate detail about pricing. Maya wants to prevent this from happening again. Which responsible AI principle is most relevant?

AFairness — Copilot is treating customers unfairlyBPrivacy & Security — Copilot is exposing sensitive pricing dataCReliability & Safety — Copilot should produce accurate information, and human review should catch errorsDInclusiveness — Copilot should be accessible to the pricing team

Check Answer

---

Next up: Compliance Manager & eDiscovery — measuring your compliance posture and searching for content during investigations.