Free AI-901 practice questions

Fairness: AI systems should treat people fairly and give similar outcomes to people in similar circumstances.

The model reproduces a pattern from historical decisions, so applicants with the same grades are ranked differently by neighbourhood. That is the classic fairness problem: bias in the training data leads to discriminatory outputs. Inclusiveness is the tempting near-twin, but it concerns whether people can use and benefit from the system at all, not whether equally qualified people get equal results.

1. Spot the harm: equally qualified people receive different outcomes.
2. Trace the cause: bias in the historical training data.
3. Separate fairness (equal treatment of similar people) from inclusiveness (nobody excluded from using the system).
4. Choose fairness.

Exam tip: Similar people, different outcomes because of group membership: fairness.

System message vs user prompt in a chat application.

Standing rules that apply to every turn, such as the assistant's role, tone and what it must decline, belong in the system message, which the app sends with each request and the model treats as high-priority guidance. User prompts carry the customer's specific request, so they cannot enforce consistent behaviour. Content filters screen harmful content but do not shape tone or topic, and assistant messages are conversation history.

1. Notice that the rules apply to every reply, whatever the customer types.
2. Standing behaviour (role, tone, boundaries) is the job of the system message.
3. User prompts hold the per-turn request; assistant messages hold previous model replies.
4. Content filters handle harmful content, not tone or topic scope.

Exam tip: Rules for every turn go in the system message; the request of the moment goes in the user prompt.

Assessing fairness by comparing model outcomes and errors across sensitive groups.

The request is for evidence that similar applicants get similar decisions across groups, which is a fairness question. The way to get it is to evaluate the model per group, comparing approval rates and error rates among applicants with similar profiles (the fairness assessment in Microsoft's Responsible AI dashboard does exactly this). The other actions are useful but serve transparency, privacy and security, or reliability.

1. Identify the principle behind the request: similar people should get similar decisions.
2. Ask which action measures outcomes per group.
3. Reject actions that serve other principles (disclosure, encryption, human review of uncertain cases).
4. Choose a per-group comparison of approval and error rates.

Exam tip: Fairness is checked by measuring outcomes and errors per group, not assumed.

Components of an effective system message.

Microsoft's guidance lists the key components of a system message as the role and task, the audience and tone, scope and boundaries (including what to do when the assistant cannot comply), safety guidelines and, optionally, the tools and data it can use. The employee's question is the user prompt, earlier model replies are assistant messages, and the deployment name is an API parameter rather than prompt content.

1. Ask which items stay the same across every conversation.
2. Role, audience and boundaries are stable instructions, so they belong in the system message.
3. The current question is the user prompt; previous replies are assistant messages.
4. Request parameters such as the deployment name are not part of any prompt.

Exam tip: System message = who the assistant is, who it serves, and where its limits are.

Reducing bias: representative training data plus measuring outcomes per group (proxy features defeat simply removing a sensitive column).

Bias here comes from historical data in which one group dominates. Two things reduce the risk: making the training data more representative, and measuring the model's outcomes per group so any remaining bias is visible. Removing the gender column is the tempting answer, but correlated features act as proxies and let the model rebuild the same pattern, so on its own it neither removes the bias nor proves it is gone. Explanations (transparency) and content filters (safety) address different principles.

1. Identify the source of risk: skewed historical training data.
2. Recall that sensitive attributes leak through proxy features, so deleting a column is not a fix.
3. Pick the data-side mitigation (rebalance or reweight).
4. Pick the measurement step (compare outcomes and errors per group).
5. Reject actions that serve transparency or content safety.

Exam tip: Reduce bias at the data and prove it with per-group testing; deleting the sensitive column is not enough.