Question 970A model application is asked to include private customer data in its answer despite authorization boundaries. What controlled validation most directly confirms the approach worked? Inject synthetic secrets and verify redaction, log access denials and retention configuration. Compare recent labeled metrics and calibration to a frozen prior period under comparable evaluation rules. Run allowed and denied access tests using the actual serving identity and audit role grants. Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable. ExplanationCorrect Answer D:Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable. This tests the underlying failure: Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. Other Answers A:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. That does not establish the required outcome; verify it by checking: Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable. B:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. That does not establish the required outcome; verify it by checking: Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable. C:IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. That does not establish the required outcome; verify it by checking: Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable.
Question 971A model application is asked to include private customer data in its answer despite authorization boundaries. What governance or operational safeguard should accompany the implementation? Version explainability methodology and require explanation quality reviews when features or architectures change. Require preprocessing digest parity and contract tests for every model and feature pipeline update. Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. ExplanationCorrect Answer D:Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. This protects the selected approach because Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The immediate solution is: Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. Other Answers A:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The continuing safeguard required here is: Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. B:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. The continuing safeguard required here is: Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. C:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The continuing safeguard required here is: Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources.
Question 972A bank uses an automated model that produces different error rates across protected customer segments. Which proposed approach best resolves the described model engineering problem? Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. Monitor labeled performance and investigate concept drift using refreshed ground truth instead of input drift alone. Use supported model-explainability methods on Agent Platform and validate attribution against domain expectations. ExplanationCorrect Answer A:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. To validate the correction: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Other Answers B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. C:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. D:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations.
Question 973A bank uses an automated model that produces different error rates across protected customer segments. Which underlying ML principle makes the proposed approach effective? Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. ExplanationCorrect Answer D:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. This supports the specific implementation: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. A suitable independent check is: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Other Answers A:Data drift is a change in input distributions and can occur even when prediction-outcome labels are delayed. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. B:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. C:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions.
Question 974A bank uses an automated model that produces different error rates across protected customer segments. Which measurement or test should the team perform after implementation? Inject synthetic secrets and verify redaction, log access denials and retention configuration. Test a suite of adversarial and unauthorized prompts and verify sensitive output is blocked and auditable. Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Compare recent labeled metrics and calibration to a frozen prior period under comparable evaluation rules. ExplanationCorrect Answer C:Compare group metrics and representative false negatives on consented and legally appropriate audit samples. This tests the underlying failure: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. Other Answers A:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. B:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. D:Concept drift changes the relationship between features and targets rather than necessarily changing feature marginals. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples.
Question 975A bank uses an automated model that produces different error rates across protected customer segments. What preventive release checklist item best protects the intended outcome? Apply data-classification and logging reviews before enabling verbose AI application diagnostics. Define attribution-change alert policies and require domain analysis before model replacement. Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. Require periodic IAM entitlement recertification and automatic policy checks before model deployment. ExplanationCorrect Answer C:Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. This protects the selected approach because Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The immediate solution is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. Other Answers A:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. B:Attribution drift can signal changes in how a model uses information that overall aggregate metrics may miss. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. D:IAM controls cloud API permissions independently of whether a model produces accurate or safe predictions. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions.
Question 976A recruitment tool is suspected of systematically rejecting qualified applicants from a minority group. What should the team change to address the specific technical requirement? Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. Redact or minimize sensitive prompt and response fields and restrict monitoring access with a documented retention policy. Use supported safety filters and explicit application-level response validation rather than relying on prompts alone. ExplanationCorrect Answer B:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. To validate the correction: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Other Answers A:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. C:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. D:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations.
Question 977A recruitment tool is suspected of systematically rejecting qualified applicants from a minority group. What Google Cloud or model behavior explains the best solution? Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. Training-serving skew can degrade predictions without any change to the underlying fitted model weights. Observability data can itself expose confidential training or inference content if logs are collected without safeguards. ExplanationCorrect Answer A:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. This supports the specific implementation: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. A suitable independent check is: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Other Answers B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. C:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. D:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. The cause in the stated scenario is different: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions.
Question 978A recruitment tool is suspected of systematically rejecting qualified applicants from a minority group. What is the most relevant acceptance check for the proposed ML change? Replay identical raw cases in both environments and compare feature vectors and model outputs. Submit defined unsafe test inputs and inspect filter decisions, false positives and user-safe fallback responses. Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Inject synthetic secrets and verify redaction, log access denials and retention configuration. ExplanationCorrect Answer C:Compare group metrics and representative false negatives on consented and legally appropriate audit samples. This tests the underlying failure: Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. Other Answers A:Training-serving skew can degrade predictions without any change to the underlying fitted model weights. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. B:Model instructions are probabilistic and safety guardrails need separate policy checks at input and output boundaries. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. D:Observability data can itself expose confidential training or inference content if logs are collected without safeguards. That does not establish the required outcome; verify it by checking: Compare group metrics and representative false negatives on consented and legally appropriate audit samples.
Question 979A recruitment tool is suspected of systematically rejecting qualified applicants from a minority group. What preventive release checklist item best protects the intended outcome? Maintain prompt-injection regression tests, least-privilege retrieval and incident review for new knowledge sources. Version explainability methodology and require explanation quality reviews when features or architectures change. Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. Version retrieval indexes and evaluation prompts and gate releases on grounded-answer regression tests. ExplanationCorrect Answer C:Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. This protects the selected approach because Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. The immediate solution is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. Other Answers A:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. B:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions. D:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. The continuing safeguard required here is: Require fairness monitoring, documented mitigation ownership and human escalation for high-impact decisions.
Question 980A public services classifier needs evidence of fairness and disparate-impact monitoring before expanding use. Which Google Cloud machine-learning design choice best satisfies the stated requirement? Enforce retrieval access controls and layered safety inspection such as Model Armor for malicious prompts and sensitive outputs. Use supported model-explainability methods on Agent Platform and validate attribution against domain expectations. Run continuous groundedness, citation relevance and task-completion evaluations on a held-out generative AI dataset. Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. ExplanationCorrect Answer D:Aggregate accuracy can obscure uneven subgroup errors and biased training labels or sampling distributions. To validate the correction: Compare group metrics and representative false negatives on consented and legally appropriate audit samples. Other Answers A:Prompt injection is untrusted input attempting to override application instructions and can cause data leakage without isolation. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. B:Feature attribution describes how inputs relate to predictions, but explanations are not necessarily causal proof. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations. C:Generative AI answer quality can regress through retrieval or prompt changes even when infrastructure is healthy. It does not resolve the specific failure; the targeted correction is: Evaluate model outcomes and error rates across relevant groups and apply documented responsible-AI mitigations.