Evaluate the system only with synthetic data.<\/span><\/li>\n<\/ol>\nCorrect Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Evaluating AI outcomes across demographic groups helps identify whether a system performs differently for particular populations. Overall accuracy can hide important disparities because a model may perform well for the majority population while producing substantially worse results for a smaller group. A governance-focused evaluation should therefore examine appropriate performance measures by subgroup, considering the use case and the potential impact of errors. Removing demographic information is not necessarily appropriate because demographic information may be needed for fairness testing and monitoring, even when it should not be used directly by the model. Synthetic data can support testing but does not automatically represent real-world population characteristics. Subgroup evaluation provides evidence that can inform mitigation, oversight, and deployment decisions.<\/span><\/p>\nQuestion 342<\/b><\/p>\n
What is a key limitation of using a single fairness metric to evaluate an AI system?<\/b><\/p>\n\n- A single metric always prevents model deployment.<\/span><\/li>\n
- Fairness metrics can never be calculated using real-world data.<\/span><\/li>\n
- One metric may capture one dimension of fairness while overlooking other relevant concerns.<\/span><\/li>\n
- Fairness metrics are only useful for security testing.<\/span><\/li>\n<\/ol>\n
Correct Answer: 3<\/b><\/p>\n
Explanation<\/b><\/p>\n
Fairness is multidimensional, and no single metric necessarily captures every concern relevant to an AI system. For example, a model may appear balanced according to one statistical measure while still producing unequal error rates, accessibility problems, or other harmful outcomes for particular groups. The appropriate evaluation depends on the system’s purpose, affected populations, decision context, and applicable requirements. Governance teams should therefore select measures that are relevant to the use case and interpret them alongside qualitative evidence, operational context, and human review. Treating one metric as definitive can create false confidence. A broader assessment provides a more realistic understanding of potential disparate outcomes and helps organizations choose appropriate mitigation strategies.<\/span><\/p>\nQuestion 343<\/b><\/p>\n
What is the primary governance concern when a variable acts as a proxy for a sensitive attribute?<\/b><\/p>\n\n- The model will always become less accurate.<\/span><\/li>\n
- The proxy can contribute to discriminatory outcomes even if the sensitive attribute is excluded directly.<\/span><\/li>\n
- The model cannot process numerical information.<\/span><\/li>\n
- The system will automatically stop making predictions.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Excluding a sensitive attribute from an AI model does not necessarily eliminate fairness risks. Other variables may correlate strongly with that attribute and effectively serve as proxies. For example, geographic information, purchasing patterns, or educational history could sometimes correlate with characteristics that an organization is trying to protect from inappropriate use. A governance process should therefore examine relevant features and relationships rather than assuming that removing one field solves the problem. Proxy analysis can be part of model testing, especially for high-impact use cases. If a proxy contributes to inappropriate disparities, organizations may need to modify the model, restrict certain features, introduce additional controls, or increase human oversight before deployment.<\/span><\/p>\nQuestion 344<\/b><\/p>\n
Why is intersectional analysis useful when evaluating AI fairness?<\/b><\/p>\n\n- It examines outcomes for combinations of demographic characteristics rather than only one characteristic at a time.<\/span><\/li>\n
- It guarantees that every demographic group receives identical outcomes.<\/span><\/li>\n
- It eliminates the need for model monitoring.<\/span><\/li>\n
- It replaces all quantitative fairness testing.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Intersectional analysis recognizes that individuals may belong to multiple demographic groups simultaneously and that risks may appear specifically at those intersections. An evaluation that examines each characteristic separately may show acceptable results while masking a significant disparity affecting a smaller combined group. For example, performance could appear reasonable when evaluating age and gender independently but differ substantially for a particular age-and-gender combination. Intersectional testing can therefore provide a more detailed understanding of potential unequal outcomes. It does not guarantee identical results and does not replace broader monitoring or quantitative analysis. Instead, it complements other fairness assessments and can help governance teams identify populations that require additional investigation, mitigation, or oversight.<\/span><\/p>\nQuestion 345<\/b><\/p>\n
An organization discovers that an AI hiring tool has substantially higher false-negative rates for one applicant group. What should it do first?<\/b><\/p>\n\n- Ignore the result if overall accuracy is high.<\/span><\/li>\n
- Immediately delete all historical hiring records.<\/span><\/li>\n
- Investigate the cause and assess the potential impact before deciding on remediation.<\/span><\/li>\n
- Increase the model’s automation level.<\/span><\/li>\n<\/ol>\n
Correct Answer: 3<\/b><\/p>\n
Explanation<\/b><\/p>\n
A substantial difference in false-negative rates may indicate that the system disadvantages a particular applicant group. The appropriate response is to investigate the finding, understand its cause, and assess the potential impact before selecting corrective measures. The organization should examine data quality, feature selection, labeling practices, model behavior, evaluation methodology, and the business process surrounding the AI tool. Overall accuracy alone may not reveal the problem. Deleting historical records could destroy useful evidence and would not necessarily solve the underlying issue. Increasing automation would also be inappropriate while a material fairness concern remains unresolved. Depending on the findings, remediation could involve changing the model, data, thresholds, workflow, or human review process.<\/span><\/p>\nQuestion 346<\/b><\/p>\n
Which control is most useful for an AI system used to support employment decisions?<\/b><\/p>\n\n- Allowing the system to make all decisions without review.<\/span><\/li>\n
- Providing meaningful human oversight for appropriate decisions and exceptions.<\/span><\/li>\n
- Disabling all performance monitoring after deployment.<\/span><\/li>\n
- Preventing users from documenting decisions.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Employment-related AI systems can affect individuals’ opportunities and livelihoods, making appropriate oversight particularly important. Meaningful human review can provide an opportunity to identify inappropriate recommendations, unusual cases, data problems, or circumstances that the model cannot adequately evaluate. Human oversight should be substantive rather than merely a formal approval step. Reviewers should have sufficient authority, information, training, and time to challenge or override AI outputs when appropriate. Organizations should also define escalation procedures and maintain relevant documentation. Fully automated decision-making without safeguards can increase the consequences of model errors or bias. Effective governance combines model testing with operational controls, human accountability, monitoring, and clear responsibility for final decisions.<\/span><\/p>\nQuestion 347<\/b><\/p>\n
What is an important consideration when using AI to monitor employee performance?<\/b><\/p>\n\n- The organization should define appropriate purposes, transparency, and safeguards for monitoring.<\/span><\/li>\n
- Employees should never be informed that AI is being used.<\/span><\/li>\n
- Any collected information can automatically be reused for unrelated purposes.<\/span><\/li>\n
- Monitoring data should be retained indefinitely.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
AI-based employee monitoring can create significant privacy, fairness, and workplace governance concerns. Organizations should clearly define the purpose of monitoring and establish appropriate boundaries around what information is collected, how it is analyzed, who can access it, and how long it is retained. Depending on the context and applicable requirements, employees may also need appropriate transparency about the use of monitoring systems. Information collected for one purpose should not automatically be repurposed without considering authorization, necessity, and governance requirements. Indefinite retention can increase privacy and security risks. A responsible approach combines purpose limitation, proportionality, access controls, retention rules, oversight, and mechanisms for addressing concerns or disputed outcomes.<\/span><\/p>\nQuestion 348<\/b><\/p>\n
What is the best reason to maintain documentation about an AI dataset’s provenance?<\/b><\/p>\n\n- To guarantee that the model will never make an error.<\/span><\/li>\n
- To make the model execute faster.<\/span><\/li>\n
- To understand where the data originated and how it was collected or transformed.<\/span><\/li>\n
- To eliminate the need for security controls.<\/span><\/li>\n<\/ol>\n
Correct Answer: 3<\/b><\/p>\n
Explanation<\/b><\/p>\n
Dataset provenance provides information about where data originated, how it was obtained, and what transformations or processing occurred before the data was used. This information supports governance because organizations may need to assess whether data was appropriately sourced, whether relevant permissions or restrictions apply, and whether the dataset is suitable for its intended purpose. Provenance can also assist with investigating data-quality issues, fairness concerns, unexpected model behavior, or compliance questions. Documentation does not guarantee error-free models and does not replace security controls. Instead, it creates traceability that helps teams understand the data lifecycle and make informed decisions about whether a dataset should be used, modified, retained, or removed.<\/span><\/p>\nQuestion 349<\/b><\/p>\n
What does data lineage primarily help an organization understand?<\/b><\/p>\n\n- The physical size of an AI model.<\/span><\/li>\n
- How data moves, changes, and is used across systems and processes.<\/span><\/li>\n
- The exact future predictions of a model.<\/span><\/li>\n
- Whether every user agrees with an AI decision.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Data lineage describes the movement and transformation of data through systems, processes, and stages of use. In an AI environment, lineage can show where data originated, which processing steps were applied, where the information was stored, and which models or applications consumed it. This visibility supports troubleshooting, governance, auditability, and impact analysis. For example, if an upstream dataset changes, lineage can help identify which models or reports may be affected. Data lineage does not predict model outputs or measure user agreement. It is primarily a traceability mechanism that helps organizations understand dependencies and data flows. Strong lineage practices can make investigations and change-impact assessments more efficient.<\/span><\/p>\nQuestion 350<\/b><\/p>\n
Why is label quality important when training a supervised AI model?<\/b><\/p>\n\n- Poor labels can cause the model to learn incorrect relationships or patterns.<\/span><\/li>\n
- Labels have no effect on supervised learning.<\/span><\/li>\n
- Labels only determine the model’s storage requirements.<\/span><\/li>\n
- Poor labels automatically improve generalization.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
In supervised learning, labels provide the target information that the model uses to learn relationships between inputs and desired outputs. If labels are incorrect, inconsistent, incomplete, or systematically biased, the model may learn patterns that do not accurately represent the intended task. This can reduce performance and may create unequal outcomes if labeling errors disproportionately affect certain groups. Organizations should therefore establish appropriate labeling standards, quality checks, documentation, and review procedures. Where human annotators are involved, training and consistency assessments can be useful. Improving label quality does not guarantee perfect model performance, but it reduces one important source of training-data problems and supports more reliable evaluation and governance.<\/span><\/p>\nQuestion 351<\/b><\/p>\n
Which practice can help assess consistency among multiple human data annotators?<\/b><\/p>\n\n- Removing all annotation guidelines.<\/span><\/li>\n
- Measuring agreement between annotators on overlapping samples.<\/span><\/li>\n
- Allowing each annotator to use unrelated definitions.<\/span><\/li>\n
- Evaluating only the final model’s processing speed.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
When human annotators create labels for AI training data, differences in interpretation can introduce noise or systematic bias. Measuring agreement on overlapping samples can help organizations determine whether annotators are applying the labeling rules consistently. Low agreement may indicate ambiguous instructions, inadequate training, difficult examples, or weaknesses in the annotation process. Governance teams can respond by refining definitions, providing additional guidance, conducting adjudication, or reviewing problematic examples. Agreement measures should be interpreted in context because disagreement can sometimes reflect genuinely ambiguous cases rather than simple annotator error. Nevertheless, monitoring annotation consistency provides useful evidence about dataset quality and can improve confidence in the labels used to train and evaluate AI systems.<\/span><\/p>\nQuestion 352<\/b><\/p>\n
What is a key governance benefit of maintaining dataset documentation?<\/b><\/p>\n\n- It eliminates the need for testing.<\/span><\/li>\n
- It guarantees legal compliance in every jurisdiction.<\/span><\/li>\n
- It records important characteristics, intended uses, limitations, and relevant data information.<\/span><\/li>\n
- It ensures the model will always be unbiased.<\/span><\/li>\n<\/ol>\n
Correct Answer: 3<\/b><\/p>\n
Explanation<\/b><\/p>\n
Dataset documentation helps organizations understand what a dataset contains and how it should be used. Useful documentation may describe the dataset’s purpose, sources, collection methods, population characteristics, known limitations, labeling processes, quality issues, and restrictions. This information supports responsible development because teams can determine whether the dataset is appropriate for a particular application and identify risks that should be addressed during testing. Documentation does not automatically guarantee legal compliance or eliminate bias. It also does not replace validation. Instead, it provides a durable record that improves transparency, reproducibility, review, and governance throughout the AI lifecycle. Well-maintained documentation can also make future investigations and audits more efficient.<\/span><\/p>\nQuestion 353<\/b><\/p>\n
Why should an organization evaluate whether training data remains representative over time?<\/b><\/p>\n\n- Because population behavior and operating conditions can change.<\/span><\/li>\n
- Because models never require monitoring after deployment.<\/span><\/li>\n
- Because old data is always more accurate than new data.<\/span><\/li>\n
- Because representative data guarantees zero model risk.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
The characteristics of a population or operating environment can change over time. Consumer behavior, market conditions, language, technology use, organizational processes, and other factors may shift after a model is trained. As a result, data that was representative when collected may become less representative later. Monitoring relevant changes can help organizations determine whether retraining, recalibration, additional evaluation, or other controls are necessary. Representativeness is not a permanent property of a dataset. It should be considered in relation to the model’s intended use and operating environment. Even highly representative data cannot eliminate every AI risk, so ongoing monitoring remains an important part of governance and lifecycle management.<\/span><\/p>\nQuestion 354<\/b><\/p>\n
What is the main purpose of a data retention schedule for an AI system?<\/b><\/p>\n\n- To keep every dataset forever.<\/span><\/li>\n
- To define how long relevant information should be retained and when it should be deleted or reviewed.<\/span><\/li>\n
- To prevent authorized users from accessing data.<\/span><\/li>\n
- To eliminate the need for data classification.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
A data retention schedule establishes expectations for how long information should be kept and what should happen when the retention period ends. In AI environments, this may apply to training data, evaluation datasets, prompts, outputs, logs, personal information, and other records. Retention should generally reflect legitimate business, legal, operational, security, and governance requirements rather than an assumption that more data is always better. Appropriate deletion or review procedures can reduce unnecessary privacy and security exposure. Retention schedules also help organizations apply consistent practices across systems. They do not replace data classification or access controls, but they provide an important lifecycle control for managing information responsibly.<\/span><\/p>\nQuestion 355<\/b><\/p>\n
Why is synthetic data not automatically considered risk-free for AI development?<\/b><\/p>\n\n- Synthetic data can still contain biases, privacy risks, or unrealistic patterns depending on how it is generated.<\/span><\/li>\n
- Synthetic data cannot be used for any AI purpose.<\/span><\/li>\n
- Synthetic data always contains real people’s names.<\/span><\/li>\n
- Synthetic data guarantees representative outcomes.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Synthetic data can reduce reliance on certain real-world datasets, but it is not inherently free from risk. The generation process may reproduce biases present in source data, create unrealistic distributions, or introduce artifacts that do not reflect actual operating conditions. If synthetic data is generated from sensitive information, privacy considerations may also remain relevant depending on the generation method and resulting risk. Organizations should therefore evaluate synthetic datasets for quality, representativeness, privacy, and suitability for the intended purpose. Synthetic data can be valuable for testing and development, but it should not automatically be treated as equivalent to real-world data. Appropriate validation is necessary before relying on it for important AI decisions.<\/span><\/p>\nQuestion 356<\/b><\/p>\n
What should an organization consider when using an AI model trained on data obtained from a third party?<\/b><\/p>\n\n- Only the model’s inference speed.<\/span><\/li>\n
- Whether the data source, usage rights, restrictions, quality, and provenance are appropriate.<\/span><\/li>\n
- Whether the provider has a colorful website.<\/span><\/li>\n
- Whether the model has the largest possible number of parameters.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Third-party data introduces additional governance questions because the organization may not control how the information was originally collected, labeled, transformed, or licensed. Before using such data, teams should consider its provenance, quality, intended purpose, applicable usage restrictions, contractual terms, and any relevant privacy or intellectual-property considerations. The organization should also understand what responsibilities remain with the provider and what responsibilities belong to the organization using the data. Model size and inference speed do not answer these governance questions. Appropriate due diligence helps reduce the risk of building an AI system on data that is unsuitable, improperly sourced, insufficiently documented, or restricted from the intended use.<\/span><\/p>\nQuestion 357<\/b><\/p>\n
What is an important purpose of an AI transparency notice?<\/b><\/p>\n\n- To hide the use of AI from affected individuals.<\/span><\/li>\n
- To communicate relevant information about how AI is being used in a context where transparency is appropriate.<\/span><\/li>\n
- To guarantee that every AI output is correct.<\/span><\/li>\n
- To replace all security documentation.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
An AI transparency notice can help people understand that an AI system is being used and, where appropriate, provide relevant information about its role, purpose, or potential effects. The appropriate level of transparency depends on the use case, audience, risks, and applicable requirements. Effective transparency should be meaningful rather than simply providing technical information that ordinary users cannot understand. A notice does not guarantee correct outputs and cannot replace security or governance documentation. Instead, transparency supports informed interaction and can help affected individuals understand when AI contributes to a process. For higher-impact systems, organizations may also need additional information about human review, contestability, or available mechanisms for raising concerns.<\/span><\/p>\nQuestion 358<\/b><\/p>\n
What does contestability mean in the context of AI-supported decisions?<\/b><\/p>\n\n- Individuals have an appropriate opportunity to question or challenge relevant AI-supported outcomes.<\/span><\/li>\n
- The AI system can change its own source code.<\/span><\/li>\n
- The model is allowed to operate without oversight.<\/span><\/li>\n
- Users must accept every AI recommendation.<\/span><\/li>\n<\/ol>\n
Correct Answer: 1<\/b><\/p>\n
Explanation<\/b><\/p>\n
Contestability refers to providing an appropriate mechanism through which affected individuals can question, challenge, or seek review of an AI-supported outcome. This is particularly important when AI contributes to decisions that can materially affect people. A useful process may include information about the decision, a way to raise concerns, access to an appropriate human review process, and a mechanism for correcting errors when warranted. Contestability does not mean that every decision must be reversed when challenged. Instead, it creates accountability and an opportunity to identify mistakes, inappropriate data, or other problems. Effective governance should define who handles challenges, how they are documented, and how recurring issues are escalated.<\/span><\/p>\nQuestion 359<\/b><\/p>\n
Which practice best supports accessibility when deploying an AI-enabled user interface?<\/b><\/p>\n\n- Designing only for users with identical abilities and devices.<\/span><\/li>\n
- Testing the interface with diverse users and considering applicable accessibility requirements.<\/span><\/li>\n
- Removing all human support options.<\/span><\/li>\n
- Requiring users to interact only through one communication method.<\/span><\/li>\n<\/ol>\n
Correct Answer: 2<\/b><\/p>\n
Explanation<\/b><\/p>\n
Accessibility should be considered throughout the design, testing, and deployment of AI-enabled interfaces. Users may have different physical, sensory, cognitive, language, or technological needs, and an interface that works well for one group may create barriers for another. Testing with diverse users can reveal problems that standard functional testing may miss. Organizations should also consider applicable accessibility standards and provide appropriate alternatives where needed. Removing human support or forcing everyone to use one interaction method can increase exclusion. Accessibility is not simply a technical feature; it is also a governance concern because an AI system that cannot be meaningfully accessed by affected users may produce unequal practical outcomes even when its underlying model performs well.<\/span><\/p>\nQuestion 360<\/b><\/p>\n
What is the most appropriate response when fairness testing reveals a significant and unexplained disparity before an AI system is deployed?<\/b><\/p>\n\n