Responsible AI Principles — Azure AI Fundamentals AI-901 practice questions

Which offering scores content for hate, violence, sexual and self-harm categories?

1. Azure AI Content Safety category scoring
2. Microsoft Entra ID conditional access
3. Azure Monitor metrics and log queries
4. Azure Functions event-driven compute

Answer: A — Azure AI Content Safety category scoring

A) Correct — Azure AI Content Safety analyzes text and images and returns severity scores across those harm categories. B) Governs sign-in and access. C) Collects telemetry. D) Runs application code on demand.

An AI system that drives a car must behave predictably even in rare edge cases. This is most directly which principle?

1. Fairness, so the system treats comparable groups equitably
2. Inclusiveness, so people of all abilities can use the system
3. Transparency, so people can understand how a decision was made
4. Reliability and safety, so behavior stays predictable under stress

Answer: D — Reliability and safety, so behavior stays predictable under stress

D) Right answer - reliability and safety covers consistent, expected behavior, and it carries the most weight in safety-critical settings such as driving. A) Fairness concerns unequal outcomes across groups. B) Inclusiveness concerns accessibility of the experience. C) Transparency concerns explaining behavior, not guaranteeing it.

A doctor needs to understand WHY an AI model predicted a disease for a patient. Which Responsible AI principle most directly addresses this need?

1. Inclusiveness
2. Privacy and security
3. Transparency
4. Reliability and safety

Answer: C — Transparency

A) Inclusiveness is about accessibility. B) Privacy is about data protection. C) Correct — transparency means models are interpretable; users can understand decisions. D) Reliability is consistent performance.

Which feature filters harmful content from model inputs and outputs by default?

1. Content filters run by Azure AI Content Safety
2. Dedicated compute clusters in this subscription
3. Reserved capacity purchased for the workload
4. Resource groups holding the deployed resources

Answer: A — Content filters run by Azure AI Content Safety

A) Correct — deployments apply content filters, integrated with Azure AI Content Safety, to both prompts and completions. B and C) Affect capacity and cost, not safety. D) Is an organizational container for resources.

Which is a responsible-AI concern specific to generative AI?

1. Network latency between client and region
2. Generating false or harmful content at scale
3. The compute cost of running the workload
4. Disk capacity consumed by the stored data

Answer: B — Generating false or harmful content at scale

B) Correct — generative output can be untrue, unsafe or infringing and is produced at volume, which is why content filters and grounding exist. A, C and D) Are ordinary operational concerns that apply to any cloud workload.

Which statement does NOT describe one of Microsoft's six responsible AI principles?

1. Systems should treat comparable groups of people equitably
2. Systems should perform dependably and safely under expected conditions
3. People should be able to understand how and why a system behaves
4. Systems should minimize the cloud spend incurred by every inference call

Answer: D — Systems should minimize the cloud spend incurred by every inference call

D) Correct — controlling inference cost is an engineering and financial concern, not one of the six responsible AI principles. A) Tempts as a possible omission, but this describes fairness, which is a listed principle. B) Tempts likewise, but this describes reliability and safety. C) Tempts likewise, but this describes transparency.

Which is NOT one of Microsoft's Responsible AI principles?

1. Fairness
2. Privacy & Security
3. Profitability
4. Transparency

Answer: C — Profitability

A/B/D) All are Responsible AI principles. C) Correct — profitability is a business concern, not an RAI principle. The 6 are: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability.

A bank requires that every automated credit decision have a named business owner, a documented human review path, and a channel through which a customer can contest the outcome. This governance MOST directly implements which principle?

1. Transparency, because the review path is documented and published
2. Accountability, because people remain answerable and in control of outcomes
3. Fairness, because contested decisions can be corrected on appeal
4. Reliability and safety, because human review catches malfunctioning decisions

Answer: B — Accountability, because people remain answerable and in control of outcomes

B) Correct — accountability holds that the people who design and deploy AI remain responsible for how it operates, with human oversight and redress built into the process. A) Tempts because documentation is present, but transparency concerns comprehensibility of the system, not who answers for it. C) Tempts because appeals can correct bias, but fairness is about equitable treatment across groups. D) Tempts because oversight aids safety, but the defining feature here is human responsibility and redress.

An audit finds that recorded customer support calls were used to ground a generative assistant, and that transcripts containing payment details are retrievable by any employee. Which principle is MOST directly violated?

1. Privacy and security, because personal data must be protected and access controlled
2. Transparency, because customers were not told their calls train the assistant
3. Accountability, because the grounding pipeline had no documented approver
4. Reliability and safety, because exposed payment details could produce incorrect answers

Answer: A — Privacy and security, because personal data must be protected and access controlled

A) Correct — privacy and security requires that personal information be safeguarded, access be restricted, and data be used consistently with consent and regulation. B) Tempts because notice is also missing, but the concrete harm described is uncontrolled access to sensitive data. C) Tempts because governance gaps often accompany such findings, but accountability is not the principle the exposure breaches. D) Tempts by chaining to output quality, but exposure of payment data is a confidentiality issue, not a correctness one.

A hiring tool favours male applicants because it was built mainly from resumes submitted by men. The root issue is:

1. A hardware fault in the cluster that produced the model
2. Network latency between the client and the model endpoint
3. Skewed source data, which breaches the fairness principle
4. Insufficient transparency about how the tool makes choices

Answer: C — Skewed source data, which breaches the fairness principle

C) Right answer - unrepresentative source data carries its imbalance straight into the model's behavior, which is precisely the fairness principle failing. A) Nothing here points at faulty hardware. B) Latency affects speed, never who gets favoured. D) Better explanation would expose the bias but would not remove it.

A chat assistant repeats a customer's account number back in a summary that is emailed to a shared inbox. Which principle is engaged?

Privacy and security. Personal data entered in a prompt must not be propagated to a wider audience than the person who supplied it; the fix is to redact sensitive entities before the summary leaves the trusted boundary.

Responsible AI — name 3 principles.

Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability.

Name the six Microsoft responsible AI principles.

Fairness • Reliability and safety • Privacy and security • Inclusiveness • Transparency • Accountability. The first four are operational qualities of the system; transparency and accountability are governance qualities that sit over the whole lifecycle.

What does Azure AI Content Safety do, and where does it sit relative to a model?

It screens text and images for harmful content — hate, violence, sexual and self-harm categories — plus jailbreak and protected-material detection. It sits on both sides of the model: filtering the user prompt before generation and the model's completion before it reaches the user.