AI-901 lesson 2: responsible AI principles
Microsoft’s responsible AI framework has six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. An exam question rarely asks for the list. It describes a problem, such as one group being approved less often or a user not knowing why a decision was made, and asks which principle is at stake. Learning each principle as a type of problem makes the answer quick to spot.
The six principles as six kinds of problem
The topic page on responsible AI principles gives the formal list. Here each principle is paired with the failure it prevents.
| Principle | What it asks | A failure that violates it |
|---|---|---|
| Fairness | Are comparable people treated comparably? | A hiring model favours candidates of one gender with equal qualifications |
| Reliability and safety | Does it work consistently and fail safely, including in unusual conditions? | A medical triage assistant gives wildly different advice for near-identical input, or fails badly on rare cases |
| Privacy and security | Is personal data protected and access controlled? | A chatbot reveals one customer’s details to another, or training data is stored without protection |
| Inclusiveness | Can people with different abilities, languages and backgrounds use it? | A voice assistant works poorly for people with speech impairments or certain accents |
| Transparency | Can people understand how and why the system behaves as it does, and that they are dealing with AI? | A loan applicant is refused with no explanation of the main factors |
| Accountability | Is it clear who is answerable, and is there human oversight? | A system makes harmful decisions and no team owns monitoring or correction |
Reading a scenario for the primary principle
Several principles can apply to the same situation, so the skill is finding the primary harm. Three questions help.
- Who is affected, and how? Different outcomes between groups point to fairness. Some people unable to use the product at all point to inclusiveness.
- Is the problem about data or about behaviour? Exposure or misuse of personal data is privacy and security. Erratic, unsafe or untested behaviour is reliability and safety.
- Is the problem about understanding or about ownership? Not knowing why points to transparency. Not knowing who answers for it, or having no human able to step in, points to accountability.
The description usually contains one telling phrase: “approved less often” (fairness), “cannot be used by” (inclusiveness), “produces inconsistent results” (reliability), “exposed customer records” (privacy), “cannot explain” (transparency), “nobody is responsible” (accountability).
Worked example: a 300-person insurer launches a claims assistant
A 300-person insurer introduces an AI assistant that pre-screens claims. Three issues emerge in testing.
- Claims from customers in two rural districts are routed to manual review far more often than similar claims elsewhere. This is fairness: comparable claims receive different treatment based on a group trait. The remedy is to examine the data and the outcomes by group and correct the imbalance.
- Customers who dictate claims by phone complain that the assistant often misunderstands them, especially customers with speech differences. This is inclusiveness: the design did not account for the full range of users. The remedy is to test with those users and offer alternative ways to submit a claim.
- Staff cannot tell customers why a claim was flagged. This is transparency: the reasoning is hidden. The remedy is to surface the main factors and to state clearly that an AI model assisted the decision.
After launch the insurer also names a team that owns monitoring, defines when a person must review a case, and keeps a route for appeals. That is accountability. Securing claim records with access controls and avoiding storing more personal data than needed covers privacy and security, and regular testing against unusual claims covers reliability and safety.
Easy to confuse
| Often mixed up | How to separate them |
|---|---|
| Fairness vs inclusiveness | Fairness is about equal treatment of comparable people in outcomes; inclusiveness is about whether everyone can access and use the system at all |
| Transparency vs accountability | Transparency is understanding how the system works and why it decided; accountability is who is answerable and whether humans oversee it |
| Privacy and security vs reliability and safety | Privacy and security protect data and access; reliability and safety concern correct, consistent, harm-avoiding behaviour |
| Fairness vs transparency | An unequal result is fairness; an unexplained result is transparency, even if the result was equal |
An exam trap is a scenario where a fix sounds like another principle. If the question asks what the problem is, answer with the principle of the problem, not of the fix.
Key points
- The six principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.
- Fairness concerns equal treatment; inclusiveness concerns whether people with different abilities and backgrounds can use the system.
- Transparency is about understanding and disclosure; accountability is about ownership and human oversight.
- Privacy and security protect data; reliability and safety concern consistent, safe behaviour.
- Find the primary harm in the scenario and match it to one principle; the key phrase usually gives it away.
Check yourself
Question 1. A photo-tagging feature recognises faces well for lighter skin tones but regularly fails for darker skin tones. Which principle is most directly affected?
- A. Transparency
- B. Fairness
- C. Accountability
- D. Privacy and security
Show answer
Answer: B
Unequal performance for different groups of people is a fairness problem. Transparency concerns explaining behaviour, accountability concerns ownership, and privacy concerns the handling of personal data; none of these is the described issue.
Want more questions like this? Full AI-901 practice tests →
Question 2. A public-service chatbot only accepts typed input and cannot be used with a screen reader, so some residents cannot use it. Which principle does this primarily concern?
- A. Fairness
- B. Reliability and safety
- C. Transparency
- D. Inclusiveness
Show answer
Answer: D
The problem is that people with certain abilities cannot use the system, which is inclusiveness. Fairness is about unequal outcomes for people who can use the system, reliability is about consistent behaviour, and transparency is about explaining decisions.
Want more questions like this? Full AI-901 practice tests →
Question 3. After a harmful decision by an AI system, an organisation finds that no person or team was assigned to monitor it or to step in when it went wrong. Which principle was neglected?
- A. Accountability
- B. Transparency
- C. Privacy and security
- D. Inclusiveness
Show answer
Answer: A
Accountability means people are answerable for the system and can oversee it. Transparency would apply if the issue were a lack of explanation, privacy and security if data were exposed, and inclusiveness if some people could not use the system.
Want more questions like this? Full AI-901 practice tests →
This lesson belongs to the first exam domain, Identify AI concepts and capabilities, which carries 40 to 45% of AI-901, and the recap page AI concepts explained keeps the six principles in a single table for revision.