Responsible AI: Microsoft's 6 Principles Explained With Examples (AI-901)

AI systems make or influence decisions about loans, jobs, medical care and what people see online. Microsoft's responsible AI approach rests on six principles, and AI-901 expects you to recognise each one from a short scenario.

The six principles are:

1. Fairness
2. Reliability and safety
3. Privacy and security
4. Inclusiveness
5. Transparency
6. Accountability

A handy way to remember them: "Fair, Reliable, Private, Inclusive, Transparent, Accountable." Transparency and accountability are sometimes described as the foundation the other four rest on.

1. Fairness

AI systems should treat all people fairly. Similar people should get similar outcomes, whatever their gender, ethnicity, age, location or other characteristics.

Unfairness usually comes from the data. If a hiring model is trained on ten years of decisions that favoured one group, it learns to repeat that pattern.

What helps:

• Training data that represents everyone the system will serve
• Measuring results for different groups, not just overall accuracy
• Removing features that act as stand-ins for sensitive attributes (a postcode can stand in for ethnicity)
• Tools such as fairness assessments in Azure Machine Learning's Responsible AI dashboard

Exam clue: a model approves, rejects, ranks or prices groups of people differently.

2. Reliability and safety

AI systems should perform reliably and safely, including in situations they weren't designed for.

A self-driving feature that works in sunshine must also behave safely in fog. A medical chatbot must not give dangerous advice when a question is phrased oddly.

What helps:

• Rigorous testing, including rare edge cases and adversarial inputs
• Clear operating limits, with a safe fallback when the system is unsure
• Monitoring after release, and a way to roll back
• For generative AI: content filters, red-teaming and evaluations of groundedness and harmful content

Exam clue: testing, edge cases, unexpected conditions, harm to people, monitoring.

3. Privacy and security

AI systems should be secure and respect privacy. AI often needs large amounts of personal data, so it must be collected, stored and used responsibly.

What helps:

• Collecting only the data you need, and getting consent
• Encrypting data and restricting who can access it
• Removing or masking personal information, for example with PII detection
• Protecting against attacks such as prompt injection and attempts to extract training data
• Knowing where data is processed, which matters for laws like GDPR

Exam clue: personal data, consent, encryption, access, data leaks, attacks.

4. Inclusiveness

AI systems should empower everyone and engage people. Systems should work for people of all abilities, languages and backgrounds.

Examples:

• Speech recognition that works with different accents
• Apps that support screen readers, captions and voice input
• Interfaces in several languages
• Testing with people with disabilities during design, not after launch

Exam clue: disabilities, accessibility, accents, languages, "everyone can use it".

Fairness and inclusiveness are easy to confuse. Fairness is about outcomes (are people treated equally?), while inclusiveness is about access (can everyone use it?).

5. Transparency

AI systems should be understandable. People should know when they're dealing with AI, what it can and can't do, and roughly why it produced a result.

What helps:

• Telling users they're talking to an AI, and labelling AI-generated content
• Explaining which factors influenced a decision ("your application was declined mainly because…")
• Publishing documentation about intended uses and limitations, such as Microsoft's transparency notes
• Interpretability tools that show feature importance

Exam clue: explain, disclose, understand, limitations, "why did it decide that?".

6. Accountability

People should be accountable for AI systems. The AI is never responsible on its own: the designers, developers and the organisation deploying it are.

What helps:

• Clear ownership of each AI system
• Governance: review boards, approval processes and impact assessments before release
• Human oversight for high-impact decisions, with the ability to override the system
• Following laws, industry standards and internal policies

Exam clue: responsibility, governance, oversight, policies, "who answers for it?".

Scenario practice

Try these before checking the answers:

Responsible generative AI in Foundry

For generative AI, Microsoft describes a practical cycle: identify potential harms, measure them, mitigate them in layers, and operate the solution responsibly. In Microsoft Foundry, mitigations include:

• Content filters (guardrails) on prompts and responses for categories such as hate, sexual content, violence and self-harm, plus protections like prompt shields against jailbreak attempts
• System prompts that set boundaries for the model
• Grounding in trusted data to reduce made-up answers
• Evaluations that measure quality and safety before and after release

You'll see these again in the prompt engineering and agent lessons.

Key takeaways

• Learn the six principle names exactly. They're asked directly.
• Match keywords: outcomes → fairness, edge cases → reliability and safety, personal data → privacy, accessibility → inclusiveness, explanations → transparency, governance → accountability.
• Responsible AI isn't a single feature. It's a process across design, testing, release and monitoring.

Next lesson: how generative AI models work.