1.2 Responsible AI Principles
Key Takeaways
- Microsoft's six responsible AI principles for AI-901 are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.
- The exam usually tests responsible AI through scenarios: identify the risk, choose the principle, then select a practical mitigation.
- Fairness is about equitable treatment across people and groups; high average accuracy does not prove a model is fair.
- Reliability and safety, privacy and security, and accountability become especially important when Foundry apps use tools, private data, or human-impacting decisions.
- Transparency and inclusiveness require clear limitations, accessible design, and user-facing communication, not just backend model settings.
The Six Principles You Must Own
AI-901 expects you to recognize Microsoft's responsible AI principles and apply them to Azure AI solution scenarios. Memorizing the list is necessary, but not enough. The useful exam skill is to read a short business scenario and decide whether the main risk is unfair outcomes, unsafe behavior, data exposure, inaccessible design, unclear system behavior, or missing human ownership.
The six principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. They often overlap in real systems, but Microsoft exam questions usually have one strongest principle or one best mitigation.
Principle Map For AI-901 Scenarios
| Principle | What it asks in plain language | Exam trigger | Practical mitigation |
|---|---|---|---|
| Fairness | Are people treated equitably? | One group gets worse recommendations, approvals, or quality of service | Measure outcomes by group, test for bias, adjust data or process |
| Reliability and safety | Does the system behave consistently and avoid harm? | The app gives risky advice, fails under edge cases, or cannot be trusted in production | Test, monitor, add fallbacks, filter content, use human review |
| Privacy and security | Is sensitive data protected? | Prompts, outputs, documents, keys, or tool calls expose private information | Use access control, encryption, data minimization, secret management |
| Inclusiveness | Can diverse users benefit? | Users with disabilities, language needs, or different contexts cannot use the app | Accessibility, localization, inclusive testing, multiple input modes |
| Transparency | Do people understand capabilities and limits? | Users cannot tell when AI is used or why a result was produced | Explain limitations, cite sources, document model behavior |
| Accountability | Are people responsible and in control? | No owner, review path, escalation, or audit trail exists | Assign owners, define approvals, log decisions, create governance |
How To Answer Scenario Questions
Use a four-step process:
- Identify who can be affected. Is the affected party a user, customer, employee, applicant, patient, developer, or downstream decision-maker?
- Name the failure mode. Bias, hallucination, harmful content, data leakage, inaccessible design, missing disclosure, and missing review all point to different principles.
- Choose the strongest principle. If a resume tool screens out a demographic group, fairness is stronger than transparency. If a support bot leaks account numbers, privacy and security is stronger than inclusiveness.
- Pick an implementable control. Responsible AI answers should lead to measurement, documentation, access control, content filtering, human review, or governance. A slogan is rarely enough.
Fairness Is Not The Same As Accuracy
A model can be accurate on average and still fail fairness. Suppose a model predicts loan risk correctly for most applicants but performs worse for a protected group because that group is underrepresented in training data. Overall accuracy may look acceptable while real people receive unequal outcomes.
For AI-901, connect fairness to testing and measurement. The practical answer is not merely "use AI ethically." It is to evaluate outcomes across relevant groups, look for skewed error rates, and change the data, features, thresholds, or workflow when a group is systematically harmed.
Reliability, Safety, And Human Review
Reliability and safety are central when generative AI is used in open-ended contexts. A Foundry chat client may produce a fluent answer that is unsupported, unsafe, or outside the system's intended use. Controls such as grounding, content filters, prompt constraints, fallback responses, and human approval reduce risk.
Human review is especially important when output affects legal, financial, health, employment, or safety decisions. A person remains responsible for approving or rejecting high-impact output; the model is not the accountable actor.
Privacy, Security, And Connected Tools
Foundry applications often use connections, data sources, tools, and agents. Every extra connection increases the need to protect secrets, restrict access, and prevent oversharing. A prompt that includes customer records, a tool call that sends too much context, or an output that reveals personally identifiable information can fail privacy and security even when the model answer is useful.
Transparency, Inclusiveness, And Accountability In Product Design
Transparency means users should understand they are interacting with AI, what the system can and cannot do, and when answers are grounded in sources. Inclusiveness means the design works for people with different abilities, devices, languages, and contexts. Accountability means owners define acceptable use, monitoring, escalation, and correction.
On AI-901, look for answers that make these principles concrete: publish limitations, provide accessible interaction options, keep audit records, and assign a team or role to review incidents. Responsible AI is an engineering practice, not a decorative statement at the end of a project.
A company tests an AI resume-screening model and finds that overall accuracy is high, but qualified applicants from one demographic group are rejected at a much higher rate than others. Which principle is most directly implicated?
A clinic uses a generative AI assistant to draft patient instructions, but a nurse must approve high-risk responses and the app logs who approved each release. Which responsible AI idea is best represented by that review workflow?