What's changed: Added per-section figures (cert-figure-retrofit). New AI-901 Chapter 1 (Domain 1 "AI concepts and responsibilities": workloads = generative/agentic/text analysis/speech/CV/information extraction, Microsoft Foundry overview, six Responsible AI principles plus genAI/agent-specific risks and mitigations)
1.2Principles of Responsible AI
Understand Microsoft’s six Responsible AI principles—fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability—and how to address risks specific to generative AI and agents (hallucination, harmful content, prompt injection).
AI is powerful but carries risks like bias, errors, privacy harm, and misuse. Microsoft defines six Responsible AI principles as guidance. These are not just ideals—Foundry and Azure ship concrete tools supporting each. For AI-901, know each principle’s meaning and especially how to address risks specific to generative AI and agents.
1.2.1The six principles
- Fairness: avoid unfair bias against particular groups.
- Reliability & safety: operate consistently and safely, even with unexpected input.
- Privacy & security: protect personal data and handle it securely.
- Inclusiveness: ensure everyone (regardless of ability) can benefit.
- Transparency: make how the AI works, its capabilities, and limits understandable.
- Accountability: have governance so people/organizations are responsible for outcomes.
| Principle | Keyword | Example |
|---|---|---|
| Fairness | Avoid bias | Hiring AI not biased by gender |
| Reliability & safety | Safe, consistent | No runaway on odd input |
| Privacy & security | Protect data | Anonymize/protect PII |
| Inclusiveness | Everyone | Accessible via speech/captions |
| Transparency | Explainable | Explain why a decision was made |
| Accountability | Human responsibility | Org owns the outcomes |
1.2.2Risks and mitigations specific to generative AI and agents
Generative AI and agents carry unique risks. Key ones: hallucination (plausible but untrue output), harmful content (violent, hateful, sexual, or self-harm outputs), and prompt injection (malicious input that hijacks the model’s instructions). Because agents execute tools autonomously, wrong actions can cause real harm. Mitigations in Foundry combine Azure AI Content Safety to detect/block harmful content, grounding (RAG) to improve factuality, human-in-the-loop review, and limiting an agent’s permissions and tools.
| Risk | What it is | Main mitigation |
|---|---|---|
| Hallucination | Confident but false output | Grounding (RAG) / cite sources / human review |
| Harmful content | Violence, hate, sexual, self-harm | Detect/block with Azure AI Content Safety |
| Prompt injection | Malicious input hijacks instructions | Input filters / protect system prompt / limit privileges |
| Agent misbehavior | Autonomous actions cause harm | Least-privilege tools / human-in-the-loop |
Q. Transparency vs accountability? Transparency = the workings/reasons are understandable; accountability = people/orgs take responsibility and govern outcomes. Q. Is hallucination the same as harmful content? No—hallucination is an untruth problem (fix: grounding); harmful content is an unsafe/inappropriate output problem (fix: Content Safety).
Watch the mix-ups: (1) fairness (avoid bias) vs inclusiveness (accessibility). (2) transparency (understandable) vs accountability (responsibility/governance). (3) hallucination (factuality → grounding) vs harmful content (safety → Content Safety).
Questions often map principle/risk → mitigation. E.g., fairness (no gender bias), transparency (explain decisions), accountability (humans own outcomes), reduce untrue output = grounding (RAG), block inappropriate output = Azure AI Content Safety, hijacking via malicious input = prompt injection.
1.2.3Section summary
- Six Responsible AI principles: fairness / reliability & safety / privacy & security / inclusiveness / transparency / accountability
- GenAI/agent risks: hallucination (→ grounding) / harmful content (→ Content Safety) / prompt injection / agent misbehavior (→ least privilege, human-in-the-loop)
- The exam centers on principle/risk → mitigation
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. What is it called when generative AI outputs "plausible but untrue" content?
Q2. Which Azure service detects and blocks inappropriate outputs such as violence, hate, or sexual content?
Q3. What is the attack that tries to hijack a model’s instructions via malicious input called?
Q4. Ensuring a hiring AI is not biased against a gender relates most to which principle?
Q5. Which principle states that people/organizations are ultimately responsible for AI outcomes?
Q6. Which best mitigates real-world harm from an autonomous tool-using agent misbehaving?