Answer: C. Govern, Map, Measure, Manage. NIST's voluntary framework organizes risk work through Govern, Map, Measure, and Manage.
What combination reflects trustworthy AI characteristics?
A Fast, profitable, popular, complex
B Fair, safe, transparent, reliable
C Private, secret, untested, opaque
D Cheap, novel, autonomous, persuasive
Answer: B. Fair, safe, transparent, reliable. NIST describes trustworthy characteristics including fairness, safety, transparency, and validity and reliability.
Why does validity matter in an AI system?
A It performs the intended task
B It guarantees zero errors
C It removes human judgment
D It prevents all misuse
Answer: A. It performs the intended task. Validity concerns whether a system performs the task it was designed for.
What does reliability describe over repeated use?
A A colorful user interface
B A large training dataset
C A popular product name
D Consistent expected performance
Answer: D. Consistent expected performance. Reliability concerns dependable performance under expected conditions and over time.
Which practice supports AI safety before release?
A Assume low-risk behavior
B Skip edge-case evaluation
C Test foreseeable failure conditions
D Hide known limitations
Answer: C. Test foreseeable failure conditions. Pre-deployment testing can reveal harmful failures before users encounter them.
How does security differ from safety?
A Safety addresses malicious threats
B Security addresses malicious threats
C Both terms mean uptime
D Security means user consent
Answer: B. Security addresses malicious threats. Security counters attacks and unauthorized access; safety focuses on preventing harm in operation.
What does transparency help people understand?
A System purpose and limitations
B Every internal model calculation
C A guarantee of fairness
D All private training records
Answer: A. System purpose and limitations. Transparency communicates relevant system purpose, operation, limits, and use context.
Which action strengthens accountability for an AI decision?
A Remove decision records
B Blame the model alone
C Avoid escalation procedures
D Assign a responsible owner
Answer: D. Assign a responsible owner. Accountability needs identifiable people or teams responsible for decisions and outcomes.
When is an explanation especially useful to affected people?
A When changing button colors
B When sorting public weather
C When decisions affect important interests
D When formatting a title
Answer: C. When decisions affect important interests. Explanations can help people understand and contest consequential AI-assisted decisions.
How should tradeoffs among trustworthy properties be handled?
A Maximize speed every time
B Use context-specific judgment
C Ignore affected communities
D Treat one metric as proof
Answer: B. Use context-specific judgment. NIST emphasizes balancing trustworthy characteristics for the system's context of use.
What should AI governance establish clearly?
A Roles, policies, and oversight
B A single model score
C Marketing slogans and logos
D Unreviewed vendor claims
Answer: A. Roles, policies, and oversight. Governance sets responsibilities, policies, and oversight for managing AI risks.
Which activity belongs early in an AI impact assessment?
A Choose a launch slogan
B Publish untested predictions
C Rank teams by model use
D Identify affected stakeholders
Answer: D. Identify affected stakeholders. Early assessment maps who may be affected and the system's potential impacts.
Why involve people affected by an AI system?
A They can guarantee accuracy
B They replace technical testing
C Their context reveals impacts
D They remove every bias
Answer: C. Their context reveals impacts. Stakeholder input can reveal impacts and lived context that technical teams might miss.
What should a team consider before adopting AI?
A Only vendor popularity
B A suitable non-AI alternative
C Maximum automation possible
D Removing appeal channels
Answer: B. A suitable non-AI alternative. NIST's Map function suggests considering non-AI alternatives for trustworthy outcomes.
Which record helps explain a model's intended use?
A A model card
B A sales brochure
C A blank checklist
D A launch announcement
Answer: A. A model card. A model card can document intended uses, limits, evaluation, and relevant context.
What data quality concern can undermine fairness?
A Consistent data labels
B Documented collection methods
C Relevant examples
D Poor group representation
Answer: D. Poor group representation. Underrepresented populations may receive poorer performance or greater harm.
How should training and evaluation data be handled?
A Mix all records randomly
B Reuse test answers in training
C Keep evaluation data distinct
D Hide data provenance
Answer: C. Keep evaluation data distinct. Separating evaluation data helps measure generalization instead of memorization or leakage.
What should happen before a system is deployed?
A Trust vendor claims alone
B Evaluate it for its context
C Skip documentation
D Use accuracy alone
Answer: B. Evaluate it for its context. Pre-deployment evaluation should match the intended context, users, and risk.
Why monitor an AI system after launch?
A Risks and behavior can change
B Deployment freezes all inputs
C Testing guarantees future performance
D Updates never alter systems
Answer: A. Risks and behavior can change. New data, users, integrations, or updates can change system behavior and risk.
What is model drift in practice?
A The model's name changes
B Users switch browsers
C A dashboard changes color
D Performance changes over time
Answer: D. Performance changes over time. Drift means a model's data patterns or performance shift from earlier conditions.
Which response is appropriate after a harmful AI incident?
A Delete evidence immediately
B Continue without review
C Contain, investigate, and remediate
D Conceal the incident internally
Answer: C. Contain, investigate, and remediate. Incident response should limit ongoing harm, preserve evidence, investigate causes, and correct issues.
How should system limitations be communicated?
A Only in source code
B In clear user-facing language
C Through unexplained scores
D After every user complaint
Answer: B. In clear user-facing language. Clear communication helps users interpret outputs and avoid unsupported assumptions.
When should human oversight increase?
A As consequences become more serious
B Whenever text has punctuation
C Only for simple suggestions
D When the model is popular
Answer: A. As consequences become more serious. Higher-impact uses often warrant stronger review and human decision authority.
What does meaningful human oversight require?
A A reviewer name only
B Automatic approval of outputs
C A hidden override button
D Authority, information, and time
Answer: D. Authority, information, and time. Oversight needs a capable reviewer with sufficient context, time, and power to intervene.
How should an affected person challenge an AI-assisted decision?
A Repeat the same automated score
B Contact an unrelated vendor
C Provide an accessible review path
D Accept every result silently
Answer: C. Provide an accessible review path. A clear review path lets people raise concerns and seek accountable reconsideration.
Why disclose AI assistance in an appropriate context?
A Disclosure proves every answer true
B People can interpret work accurately
C It removes all bias
D It guarantees user consent
Answer: B. People can interpret work accurately. Disclosure helps audiences understand how content or decisions were produced.
What should a user do with an uncertain AI answer?
A Verify it using reliable sources
B Share it as confirmed fact
C Assume fluency means accuracy
D Remove all supporting evidence
Answer: A. Verify it using reliable sources. AI outputs can be wrong; check important claims against dependable sources.
Which sign may indicate a hallucinated answer?
A A short response
B Correctly cited evidence
C A cautious limitation
D Unsupported specific details
Answer: D. Unsupported specific details. Confident wording does not establish accuracy; unsupported details need verification.
What makes a generated citation trustworthy?
A It looks professionally formatted
B It names a famous journal
C It resolves to supporting evidence
D It includes many numbers
Answer: C. It resolves to supporting evidence. Check that a citation exists and supports the claim it is attached to.
What is a responsible approach to copyrighted material?
A Assume all outputs are free
B Follow applicable rights and terms
C Remove creator attribution
D Upload any private work
Answer: B. Follow applicable rights and terms. Users should respect relevant licenses, ownership, and service terms when using AI outputs.
Which step protects confidential workplace information?
A Use approved tools and policies
B Paste it into any chatbot
C Share it in public prompts
D Include credentials for context
Answer: A. Use approved tools and policies. Use only approved workflows and avoid exposing confidential data without authorization.
When should a use case receive stronger review?
A When the prompt is short
B When a demo looks polished
C When answers are entertaining
D When potential harm is high
Answer: D. When potential harm is high. Potential severity and likelihood of harm should guide the depth of review.
How can teams test for uneven performance?
A Measure only overall accuracy
B Choose the best single case
C Compare relevant user groups
D Ignore missing data
Answer: C. Compare relevant user groups. Group-level evaluation can reveal disparities obscured by an aggregate score.
What supports accessible AI experiences?
A Use one input method
B Test with diverse users
C Avoid captions
D Rely on color alone
Answer: B. Test with diverse users. Inclusive testing helps identify barriers for users with different abilities and needs.
Why document data provenance?
A To trace origins and permissions
B To increase model speed
C To hide source limitations
D To avoid quality checks
Answer: A. To trace origins and permissions. Provenance supports traceability, quality assessment, and review of data permissions.
What does privacy-enhancing practice contribute to responsible AI?
A Automatic correctness
B Guaranteed fairness
C Unlimited retention
D Reduced unnecessary data exposure
Answer: D. Reduced unnecessary data exposure. Privacy controls reduce unnecessary collection and exposure but do not guarantee other properties.
Which security test helps uncover misuse paths?
A Logo review
B Font comparison
C Red-team testing
D Sales forecasting
Answer: C. Red-team testing. Red teaming probes systems for weaknesses and plausible misuse before and after release.
How should AI tools receive external permissions?
A Enable every available tool
B Grant only task-required access
C Share administrator credentials
D Allow unrestricted file access
Answer: B. Grant only task-required access. Limited permissions reduce the consequences of errors or malicious instructions.
What is a risk of relying too heavily on AI suggestions?
A People may overlook errors
B Models become transparent
C Data quality improves automatically
D Oversight becomes unnecessary
Answer: A. People may overlook errors. Overreliance can weaken independent judgment and let incorrect outputs pass unchecked.
What does automation bias mean?
A Rejecting every computer output
B Measuring model speed
C Removing manual data entry
D Overtrusting automated recommendations
Answer: D. Overtrusting automated recommendations. Automation bias occurs when people give undue weight to automated advice.
How can education improve responsible AI use?
A Promise perfect outputs
B Discourage all questions
C Teach limits and verification
D Replace policy training
Answer: C. Teach limits and verification. Practical training helps users recognize limits, protect data, and check important outputs.
Which employee guidance best reduces avoidable AI risk?
A One universal prompt template
B Approved uses and escalation steps
C A ban on reporting issues
D Informal verbal promises
Answer: B. Approved uses and escalation steps. Clear approved-use rules and escalation routes support safer day-to-day decisions.
Why measure more than overall accuracy?
A Impacts vary across contexts
B Accuracy is always misleading
C Metrics slow all systems
D Users dislike evaluation
Answer: A. Impacts vary across contexts. No single aggregate metric captures every relevant impact or group outcome.
What should an independent reviewer examine?
A Only the launch date
B A vendor's popularity
C The user interface color
D Methods, evidence, and conflicts
Answer: D. Methods, evidence, and conflicts. Independent review can improve testing and reduce conflicts of interest.
When should an organization reassess an AI system?
A Only at initial purchase
B When its logo changes
C After material changes
D Never after deployment
Answer: C. After material changes. Changes to data, purpose, model, or integration can alter risk and merit reassessment.
Which documentation makes AI decisions more traceable?
A Unrelated meeting slogans
B Inputs, versions, and outcomes
C Passwords and secret keys
D Unverified promotional claims
Answer: B. Inputs, versions, and outcomes. Appropriate records help reconstruct system versions, relevant inputs, decisions, and outcomes.
Which principle helps people retain control over consequential uses?
A Human authority and recourse
B Hidden automation
C Unreviewable scoring
D Forced data sharing
Answer: A. Human authority and recourse. Human authority and recourse help affected people question and challenge important outcomes.
What is a responsible response to a failed evaluation?
A Publish results without changes
B Delete the evaluation
C Blame test participants
D Fix, restrict, or defer release
Answer: D. Fix, restrict, or defer release. Teams should address failure before release, including mitigation or limiting the system's use.
Why keep a record of AI-assisted decisions?
A Prove the model is perfect
B Replace consent requirements
C Support review and accountability
D Prevent every future error
Answer: C. Support review and accountability. Appropriate records support auditing, incident investigation, and accountable review.
Which approach supports continual improvement?
A Treat deployment as final
B Monitor, learn, and update controls
C Ignore user feedback
D Freeze risk assessments forever
Answer: B. Monitor, learn, and update controls. Responsible AI risk work continues as evidence, technology, and context change.