AI-901 Identify AI concepts and capabilities: 437 practice questions

12 of the 437 Identify AI concepts and capabilities questions in the Certsqill AI-901 bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AI-901? Take the free 5-min readiness check →

1. Compare subgroup outcomes on representative validation: What should the team do next?

1. Compare subgroup outcomes on representative validation data. ✓

   Subgroup comparison can reveal unequal errors or outcomes that overall accuracy hides.
2. Require identical predictions for every applicant.

   Fairness does not require ignoring legitimate differences in application evidence.
3. Use confidence scores to identify fair decisions automatically.

   Confidence measures model certainty, not equitable treatment across groups.
4. Remove demographic fields before retraining the model.

   Other inputs may act as proxies for demographic attributes and preserve disparities.

Compare outcomes across representative demographic subgroups to assess fairness.

2. Analyze proxy attributes and compare subgroup outcomes: What should the team do next?

1. Use overall accuracy as evidence that predictions are fair.

   Overall accuracy can conceal different error rates or outcomes among groups.
2. Increase the confidence threshold for every decision.

   A threshold changes acceptance criteria but does not diagnose proxy-based disparities.
3. Delete every variable correlated with a business outcome.

   This would discard useful information without specifically identifying harmful proxy effects.
4. Analyze proxy attributes and compare subgroup outcomes. ✓

   This investigation can identify indirect signals and reveal whether remaining disparities are unjustified.

Investigate proxy attributes and compare subgroup outcomes after protected-field removal.

3. Test representative languages: What is the best next action to support fairness?

1. Test representative languages, regions, and users. ✓

   Broad representative testing can reveal uneven quality across affected populations.
2. Reuse identical cultural references in every regional test.

   References familiar in one region may be inappropriate or unclear in another.
3. Translate one prompt and compare response length.

   One prompt and response length provide an incomplete measure of usefulness and accuracy.
4. Test only the most common language.

   This can hide poor performance for smaller language communities.

Use representative tests across languages, regions, users, and request types.

4. Test noisy inputs and add controlled fallback paths: What should the team do next?

1. Disable input validation for uninterrupted processing.

   Removing validation increases exposure to unexpected inputs without making behavior dependable.
2. Increase temperature for more flexible translations.

   Temperature changes response variability but does not fix noisy-input recognition failures.
3. Report only average quality across all requests.

   An average can hide severe failures concentrated in particular input types.
4. Test noisy inputs and add controlled fallback paths. ✓

   Testing realistic variations supports validation, fallback handling, or human review when translation is unreliable.

Test noisy inputs and provide validation, fallback, or review for failures.

5. Route high-risk answers to qualified reviewers: What should the club do next?

1. Publish every answer generated from a medical prompt.

   A specialized prompt does not guarantee accurate or safe recommendations.
2. Use confidence scores alone to approve injury guidance.

   Confidence is not proof that health-related advice is correct or safe.
3. Exclude injury questions from testing.

   Avoiding high-risk tests conceals failure modes instead of establishing safe behavior.
4. Route high-risk answers to qualified reviewers. ✓

   Qualified human review provides safer judgment for uncertain, consequential guidance.

Escalate uncertain or high-risk injury guidance to qualified human reviewers.

6. Run adversarial tests for unsafe: What should it do next?

1. Ask the assistant to confirm that its own safeguards are working correctly.

   Self-assessment cannot replace independent testing with challenging inputs and observed results.
2. Test routine maintenance questions that represent the primary workload.

   Routine testing may not reveal unsafe responses, ambiguity, or safeguard bypasses.
3. Run adversarial tests for unsafe, ambiguous, and bypass requests. ✓

   Adversarial testing checks whether safeguards respond appropriately to challenging inputs and attempted bypasses.
4. Increase response creativity to handle dangerous requests more naturally.

   Creativity increases variation but does not demonstrate safe handling of dangerous requests.

Test unsafe, ambiguous, and adversarial inputs before deployment.

7. Send only the feedback content needed for summarization: What should the publisher do next?

1. Use a larger model that claims to protect personal data automatically.

   Model size does not replace data-minimization controls or appropriate personal-data handling.
2. Send only the feedback content needed for summarization. ✓

   Removing unrelated personal data minimizes exposure while preserving the information needed for the task.
3. Tell the model to ignore names and email addresses after sending them unchanged.

   A prompt instruction does not prevent unnecessary personal data from being transmitted or processed.
4. Send every available field so the model has maximum context.

   Unnecessary personal data increases privacy exposure without supporting the stated task.

Remove or redact unnecessary personal data before summarization.

8. Define retention and deletion rules: What should it do next?

1. Publish a general statement that the assistant is secure without changing transcript handling.

   A broad security statement does not define retention, access, deletion, or minimization practices for personal data.
2. Retain every transcript indefinitely so future model improvements have maximum training data.

   Indefinite retention increases privacy exposure and may exceed the information needed for support or improvement.
3. Define retention and deletion rules, restrict transcript access, and minimize stored personal information. ✓

   Clear lifecycle and access controls reduce unnecessary exposure while supporting legitimate support and governance needs.
4. Allow all support staff to search transcripts because they work for the marketplace.

   Employment alone does not justify unrestricted access to customer conversations containing personal information.

Protect retained transcripts through defined retention, deletion, limited access, and minimization of personal information.

9. Provide accessible output and test the complete experience: Which action best addresses inclusiveness?

1. Offer the same visual interface to every visitor without alternative interaction methods.

   A single visual interface may prevent access for visitors who rely on assistive technologies or different interaction methods.
2. Provide accessible output and test the complete experience with screen readers and keyboard navigation. ✓

   Accessible design and testing with assistive technologies help ensure visitors with disabilities can use the AI guide effectively.
3. Use smaller text and more visual effects to make the guide feel modern.

   Visual styling does not ensure compatibility with screen readers or keyboard-based interaction.
4. Ask visitors whether they like the guide but exclude disability-related feedback from the evaluation.

   Excluding affected users removes essential evidence about whether the experience is inclusive.

Inclusiveness requires accessible outputs and testing with assistive technologies, including screen readers and keyboard navigation.

10. Add language support and test accessibility with diverse: What should the team do first?

1. Use one interaction design for every customer to maintain consistency.

   A uniform design may exclude customers whose languages or assistive technologies require adaptations.
2. Remove language and demographic fields from customer records.

   Removing fields does not provide language access or compatibility with assistive technologies.
3. Publish response-confidence information for each supported language.

   Confidence reporting supports transparency but does not make the assistant accessible to diverse customers.
4. Add language support and test accessibility with diverse users. ✓

   Language support and testing with screen readers and diverse users directly address inclusive access.

Provide language coverage and test the assistant with diverse users and assistive technologies.

11. Evaluate varied speech and provide an alternative input: What is the best next action?

1. Publish a notice that speech recognition can make mistakes.

   A limitation notice supports transparency but does not provide an accessible way to complete the task.
2. Increase the confidence threshold so uncertain transcripts are rejected.

   A stricter threshold may reject errors but does not help people whose speech is poorly recognized.
3. Evaluate varied speech and provide an alternative input method. ✓

   Representative testing and an alternative input path address access for employees with different communication needs.
4. Require every employee to use the same approved microphone and settings.

   Standard equipment may improve consistency but cannot address varied speech or alternative communication methods.

Test varied speech and provide an alternative input method.

12. Explain the AI use and limitations: What should the team do to address this claim?

1. Explain the AI use and limitations, while retaining human review for consequential decisions. ✓

   This corrects the misconception: transparency communicates use and limitations, while human review supports accountability for consequential decisions.
2. Document model limitations and then guarantee accurate estimates.

   Documentation can support transparency, but a guarantee of correctness is not justified by transparency or documented limitations.
3. Test estimates across routine and unusual delivery conditions before release.

   Representative testing supports reliability, but it does not by itself explain the system or establish human accountability.
4. Restrict estimate access to authorized staff and retain delivery records securely.

   These controls support privacy and security, but they do not address whether transparency guarantees correctness.

Transparency explains AI use and limitations; it does not guarantee correctness, so consequential decisions still need human oversight.

Other AI-901 domains

• Implement AI solutions by using Microsoft Foundry — 603 questions →
• All 1040 AI-901 questions →