Before a model does anything in this course, there is the question of how it should behave. Microsoft answers it with six principles for responsible AI. In Microsoft's own words, they form the framework of the "Responsible AI Standard" that the company's development teams follow. The AI-901 exam asks about each of these principles individually. This lesson explains them at exam level and ties them to the example that runs through the whole course.

The example is the made-up Nordlicht Reisen GmbH, a group of travel agencies in northern Germany with 25 agencies and a booking portal. Customers mostly write in German, and a few messages come in English. In three months a pilot project is to start: small AI helpers for customer service. They are meant to support the advisers, not replace them. The head of customer service, Ms. Berg (berg@example.com), leads the project. All names and numbers in the course are assumptions made for the exercises.

The six principles

The order is the one in which Microsoft lists them:

No. Principle In short
1 Fairness Similar people and cases are treated similarly.
2 Reliability and safety The system reliably does what it is meant to do, even under unexpected conditions.
3 Privacy and security Personal and business data is protected.
4 Inclusiveness The system includes everyone and fits different needs.
5 Transparency People understand what the system can do, what it cannot do and how it arrives at results.
6 Accountability People are responsible for the system and stay in control.

Two notes on the names. The names above are the ones used in the exam and on Learn. In addition, one Learn page combines two principles in one section ("Fairness and inclusiveness"). There are still six separate principles.

What each principle means

Fairness. AI systems should treat all people fairly and should not affect similar groups differently. As an example, Learn names advice on treatments, loan applications or job applications: people with similar qualifications should receive similar recommendations. A model learns from data. If that data reflects old prejudices, the output can reflect them too.

Nordlicht: A helper suggests replies to complaints. In the same situation, it should not answer more warmly or more tersely just because a name sounds unfamiliar or the message contains mistakes.

Reliability and safety. The system should work as designed, respond safely to unexpected conditions and not allow itself to be manipulated in harmful ways. How well it copes with different situations depends on what the developers considered when designing and testing it. The consequence: test, including with rare and difficult inputs.

Nordlicht: The helper receives a message in English, a photo instead of text or an empty input. It should not crash and should not make anything up, but respond safely, for example by pointing out that a person will take over.

Privacy and security. AI systems need data to work. That is why the protection and security of personal and business information matter. Learn points to data protection laws that require transparency about collection, use and storage and give people control. Technically, this includes access restriction, encryption and monitored configuration.

Nordlicht: The hotline recordings and invoices contain names, addresses and booking data. Who may see them and where they are processed is a design question from the start. Lesson 1.3 takes up the topic of data location.

Inclusiveness. AI systems should address and include everyone. A Learn page about agents names accessibility, language variants and different levels of technical background as examples.

Nordlicht: The customers are not uniform: some write at length, others dictate, some need plain language. A helper that copes with only one kind of input excludes people.

Transparency. AI systems should make their capabilities and limits clear. People should understand what the system can do, how information is used and how outputs come about. Interpretability matters here: useful explanations of why a system behaves the way it does. This applies especially when decisions affect people's lives.

Nordlicht: Whoever writes with the helper learns that an AI system is answering and that a person reviews the answer. The principle does not demand explaining every calculation of the model. It demands that nobody is left unsure about the nature of the party on the other end.

Accountability. The people who design and deploy AI systems are responsible for how those systems work. Learn puts it this way: AI systems should not be the final authority on decisions that affect people's lives, and people keep meaningful control over highly autonomous systems. Another Learn page about agents is even more concrete: a named person is responsible for the behavior and results of every agent. It also says that people stay involved in consequential actions and that agents facing outward should be clearly identifiable as AI.

Nordlicht: That is Ms. Berg. She is the named person responsible for the pilot project. If the helper says something wrong, "the model did it" is not an answer. Ms. Berg must be able to say who set it up, what it is allowed to do and who reviews its results.

The limits of the pilot are the principles as rules

That is why the project sets two limits from the start. No model decides on refunds, cancellations or contract matters. And: A person reviews every answer of the agent before it goes to customers. Both rules follow from accountability (people stay in control), from fairness (no automatic rejection without review) and from reliability (a mistake does not reach customers unchecked). Later lessons build the helpers within these limits.

What Microsoft offers beyond that

In "Responsible AI for Microsoft Foundry", Microsoft describes a process in three stages: discover risks (for example by testing an agent with hostile inputs), protect (content filters and safeguards so that harmful content does not reach users) and govern (logging and ongoing monitoring). The AI track covers content filters in its module on safety (AI track, T6, lesson T6.3); here it is enough to know that filters help but do not replace thinking about the six principles. How principles and law relate is covered in the AI track, T1.4.

Exercise 1

Match each situation to the principle it touches most strongly (Fairness, Reliability and safety, Privacy and security, Inclusiveness, Transparency, Accountability):

  1. Only the customer service team may listen to the hotline recordings, and the files are stored encrypted.
  2. The pilot project has a responsible person who answers to the management.
  3. The helper gives useful answers even to messages with typos and in plain language.
  4. The page says: "This answer was prepared by an AI system and reviewed by a person."
  5. Before launch, the helper is tested with empty, very long and foreign-language messages.
  6. Two requests with the same content but different names in the sender field get equally good answers.

Solution 1
  1. Privacy and security. It is about access to and protection of personal data.
  2. Accountability. People stand behind the system.
  3. Inclusiveness. The system should also cope with different ways of expressing things. The typo alone would also be a question of reliability; the plain language tips the balance toward inclusiveness.
  4. Transparency. People find out what they are dealing with and how the answer comes about.
  5. Reliability and safety. The test shows how the system reacts under unexpected conditions.
  6. Fairness. Similar cases are treated similarly.

Some situations touch several principles. What matters is that you can name the strongest one and justify it in one sentence.

Further reading