Task 1.1: Describe principles of responsible AI

• AThe inclusiveness principle

  Incorrect

  Inclusiveness is about designing so that nobody is shut out of using the system, for example people with disabilities. Every applicant here can use the system; the problem is the outcome it produces.

• BThe fairness principle

  Correct

  Fairness means people in similar situations get similar outcomes. Similar applicants receiving different results because of a factor that tracks a demographic group is exactly the harm this principle targets.

• CThe transparency principle

  Incorrect

  Transparency is about people understanding how a system reaches its decisions. Explaining the screening model wouldn't stop it from treating similar applicants differently.

• DThe reliability and safety principle

  Incorrect

  Reliability and safety is about consistent behavior and avoiding harm under expected and unexpected conditions. The model is behaving consistently; it is consistently unequal across groups.

Microsoft describes fairness as AI systems treating everyone fairly and not affecting similarly situated groups differently, with lending, employment and similar decisions as typical examples. An input that acts as a stand-in for a demographic group can create that harm even when the group itself is never used as an input.

• AFairness

  Incorrect

  Fairness concerns how outcomes compare between groups of people. The threshold guards against any wrong reading, whoever the patient is.

• BTransparency

  Incorrect

  Transparency would call for explaining to patients and staff how the system works and where it can fail. Holding uncertain orders is a safety control, not an explanation.

• CPrivacy and security

  Incorrect

  Privacy and security is about protecting the personal data on the prescriptions. Holding uncertain orders doesn't change how that data is stored or who can see it.

• DReliability and safety

  Correct

  Models are probabilistic and can be wrong. Refusing to act on low-confidence results in a setting where a mistake could harm a patient is a direct way to make the system behave safely.

AI systems are based on probabilistic models and aren't infallible. Using a confidence value to decide whether the system acts at all, and avoiding action below a threshold, is a standard way to keep a system reliable and safe when the cost of an error is high.

• ADetect and redact personal information in each transcript before it's stored for analysis

  Correct

  Removing names, phone numbers and addresses before the data is kept protects the people in the transcripts while leaving the complaint content available to analysts.

• BLabel chat replies that a model generated so customers know they came from AI

  Incorrect

  Labeling AI-generated replies supports transparency. It does nothing to protect the personal details stored in the transcripts.

• CRun the transcripts through a content filter that flags abusive language before analysts read them

  Incorrect

  Filtering abusive language is a safety measure. It leaves the names, phone numbers and addresses in place.

• DOffer the support chat in a mode that works with screen readers and keyboard navigation for customers who rely on them

  Incorrect

  Accessible design supports inclusiveness. It doesn't change what personal information ends up in the stored transcripts.

Privacy and security is about protecting personal and business data that AI systems collect and use. Detecting and redacting personally identifiable information before data is shared or analyzed is a common way to meet privacy policies and regulations.

• ATransparency

  Incorrect

  Transparency is about making people aware of how the system works and what its limits are. Offering other ways to ask and read answers doesn't explain the system.

• BPrivacy and security

  Incorrect

  Privacy and security covers protecting the data visitors provide. The new input and output options don't change how data is handled.

• CInclusiveness

  Correct

  A speech-only kiosk would shut out visitors who are deaf, hard of hearing or unable to speak. Adding typed input and text output lets everyone use it.

• DAccountability

  Incorrect

  Accountability concerns human responsibility for the system. The change here is about who can use the kiosk.

Inclusiveness asks that the benefits of AI be available to everyone, so solutions shouldn't exclude groups of users. Offering more than one way to give input and receive output is a common way to apply it to speech experiences.

• AStore a recording of each call with the caller's number, which serves accountability

  Incorrect

  Keeping recordings with phone numbers adds personal data to protect, and it doesn't tell callers that the voice is synthetic.

• BTell callers at the start that they're talking to a synthetic voice, which serves transparency

  Correct

  Disclosure lets people know they're listening to a computer-generated voice. Microsoft's guidance for synthetic voices treats it as the core of being transparent with users.

• CRoute complicated requests to a human operator, which serves accountability

  Incorrect

  Escalating to people is useful, but callers would still assume the synthetic voice is a person.

• DUse a faster speaking rate so calls finish sooner, which serves reliability and safety

  Incorrect

  Speaking rate is a delivery setting. It isn't a responsible AI control, and the synthetic nature of the voice stays hidden.

As synthetic voices become hard to tell apart from human ones, Microsoft recommends disclosure: letting people know they're interacting with a synthetic voice. Making users aware of how a system works is the heart of the transparency principle.

• AFairness

  Incorrect

  Fairness would require checking that similar applicants get similar rankings. A sign-off step assigns responsibility; it doesn't measure or correct unequal treatment.

• BTransparency

  Incorrect

  Transparency concerns openness with the people affected. The policy assigns responsibility instead.

• CPrivacy and security

  Incorrect

  Privacy and security is about protecting applicants' personal data. The policy decides who makes the decision, not how data is handled.

• DAccountability

  Correct

  People must stay answerable for decisions that affect others, and AI shouldn't be the final authority on them. A named manager owning each decision keeps humans in control.

Accountability means the people who design and deploy AI systems answer for how they operate. Microsoft's guidance calls for norms that keep AI from being the final authority on decisions that affect people's lives and keep humans in meaningful control.

• AProbe the assistant with simulated adversarial prompts before launch to find ways it can be pushed into harmful answers

  Correct

  Adversarial testing, such as automated scans with the AI Red Teaming Agent, finds safety weaknesses before real users do.

• BShow a note above the chat window explaining that answers are generated by AI

  Incorrect

  Telling users that answers come from AI supports transparency. It doesn't stop the assistant from producing a harmful answer.

• COffer a high-contrast, large-text version of the chat for customers with low vision

  Incorrect

  A high-contrast, large-text version helps more customers use the chat, which is inclusiveness rather than safety.

• DApply guardrails that scan user prompts and model outputs and block harmful content

  Correct

  Guardrails protect at runtime by detecting harmful content in inputs and outputs and blocking it before it reaches the customer.

• ELet customers download a copy of their own chat history from their account

  Incorrect

  Giving customers control over their own data is a privacy practice, not a safety control.

Microsoft's guidance for trustworthy generative AI asks teams to discover risks, for example by testing with adversarial prompts, and to protect against them with content filters and guardrails that block harmful outputs. Both keep the system behaving safely under misuse.

• ASome old records hold typing errors that weaken results, a reliability risk

  Incorrect

  Data errors can hurt reliability, but the stem describes a skew toward one group, not mistakes in the records.

• BAn attacker could alter the stored hiring records, a security risk

  Incorrect

  Protecting stored data matters, but tampering isn't something the skew creates. The concern here is how the model will treat applicants.

• CThe model may learn and repeat bias in past decisions, a fairness risk

  Correct

  Models learn from the data they're given. If that data reflects unconscious bias, the model can produce discriminatory outputs.

• DPast applicants didn't choose how their data is reused, a privacy risk

  Incorrect

  Control over data use is a privacy consideration, but it would apply whatever the makeup of the data. It isn't the risk the skew creates.

Training data is usually selected by people, and both the data and the selection criteria can reflect unconscious bias. A model trained on such data can reproduce it and treat similar applicants differently, so developers need to minimize bias in training data.

• APrompt Shields, which detect attacks hidden in documents, emails and prompts

  Correct

  Prompt Shields analyze text for user prompt attacks and for document attacks, where hidden instructions in content try to hijack the model.

• BPII detection in Azure Language, which finds personal information in the email text

  Incorrect

  PII detection finds personal details such as names and phone numbers. It doesn't detect instructions planted in an email.

• CLanguage detection in Azure Language, which identifies the language of each email

  Incorrect

  Knowing the email's language doesn't reveal hidden instructions in it.

• DThe analyze image API, which scans pictures and attachments for violence and hate

  Incorrect

  Image analysis scores harmful visual content; the attack here is text inside an email.

Indirect, or document, attacks hide instructions in content that a model processes, such as emails or web pages. Prompt Shields scan for both direct user prompt attacks and these document attacks, which protects the security of the solution.

• ACite the source documents behind each answer

  Correct

  Citations let users inspect the content an answer is based on and notice when a source is outdated or doesn't fit.

• BHave HR staff spot-check a sample of answers each month

  Incorrect

  Spot checks help catch wrong answers behind the scenes, but an employee reading an answer still can't see what it's based on.

• CKeep the assistant's conversation logs for 90 days

  Incorrect

  Retention is a data management choice. It tells employees nothing about where answers come from.

• DLimit the assistant to approved, current HR documents

  Incorrect

  Curating the sources improves what answers draw on, but employees still can't tell which document a given answer used.

Transparency helps people understand how a system reaches its outputs and what its limits are. For grounded answers, returning citations or links lets users verify claims against the sources.

• ATesting the agent with adversarial prompts before it is released

  Incorrect

  Adversarial testing finds risks before and after deployment, which is the Discover stage.

• BApplying guardrails that block harmful outputs before they reach users

  Incorrect

  Blocking harmful outputs at runtime is the Protect stage.

• CAssessing a planned use case for potential risks before building it

  Incorrect

  Identifying and assessing risks belongs to the Discover stage.

• DTracing and monitoring the agent in production to spot new risks

  Correct

  Govern covers tracing, monitoring and compliance integrations after deployment, so new risks surface while the agent is live.

Microsoft's guidance asks teams to discover risks, protect against them at the model and agent level, and govern agents through tracing, monitoring and compliance integrations once they're in production.