When an AI system used for safety-critical quality control produces inaccurate responses that could harm users, the most immediate and effective safeguard is inserting human judgment into the decision process before unsafe outputs can reach production or end users.

Why A is Correct: The ISACA AAIR human oversight guidance identifies human-in-the-loop reviews as the most effective mitigation when AI outputs pose safety risks. In safety-critical applications like manufacturing quality control, human reviewers can catch and correct AI errors before they result in unsafe products reaching consumers. This control is immediately implementable, does not require model retraining, and directly addresses the safety risk. It is the appropriate response when AI accuracy cannot be fully trusted.

Why B is Wrong: Bias and fairness testing is a model evaluation activity that assesses whether outputs are systematically skewed. While useful for improving the model, it does not provide immediate protection against the safety risk of current inaccurate outputs.

Why C is Wrong: Synthetic data augmentation may improve model quality over time but requires model retraining and does not prevent currently inaccurate outputs from causing harm in the interim.

Why D is Wrong: Prompt engineering training improves how users interact with AI systems to elicit better outputs. It is useful for generative AI applications but does not directly address safety risks from QC system inaccuracies, which require operational oversight rather than improved prompting.

Training program effectiveness KPIs must measure behavioral change—whether training has actually altered how participants act in their work environment—rather than knowledge acquisition or adoption rates. The most meaningful behavioral signal for AI risk awareness training is whether trained users report suspicious activity.

Why B is Correct: According to ISACA AAIR training effectiveness measurement guidance, the number of AI irregularities and potential tampering incidents reported by users is the most appropriate behavioral KPI for risk awareness training effectiveness. When users report unusual AI behavior, this demonstrates they have internalized training concepts well enough to recognize anomalies and understand their obligation to report them. This behavioral change—from passive observation to active reporting—is the intended outcome of awareness training.

Why A is Wrong: Risk rating changes measure risk management process adjustments rather than individual behavioral change from training. Risk ratings reflect aggregate organizational risk, not the specific behavioral impact of an awareness training program.

Why C is Wrong: Ability to identify financial impacts is a knowledge assessment metric—it measures what users know rather than what they do differently as a result of training. Awareness training aims to change behavior, not just inform.

Why D is Wrong: AI adoption rates measure technology uptake, not risk awareness. Higher adoption could indicate confidence in AI systems but does not measure whether employees recognize and report AI risks—the specific objective of risk awareness training.

When multiple risk sources are present in a critical AI deployment, the risk practitioner must apply a prioritization framework that focuses resources on the risks with the greatest potential for organizational harm. This risk-based prioritization is more effective than comprehensive but undifferentiated risk cataloging.

Why A is Correct: The ISACA AAIR risk assessment methodology prioritizes risk factors based on potential harm severity as the most appropriate approach for critical AI systems. Focusing on risks most likely to generate substantial harm ensures that the organization ' s risk management resources are directed toward the exposures that matter most—protecting the critical functions that the AI system supports and preventing the most consequential adverse outcomes.

Why B is Wrong: Quantifying competitors ' risk events provides external benchmarking data but cannot accurately characterize the organization ' s specific risk profile. Competitor risk events may involve different AI architectures, use cases, and organizational contexts that make direct comparison unreliable.

Why C is Wrong: Rating each risk factor independently without integration produces a fragmented view that misses risk correlations, cascade effects, and the compounding nature of multiple simultaneous risk factors. Independent ratings also do not inherently lead to the harm-based prioritization needed for critical systems.

Why D is Wrong: Documenting technical limitations is a useful input to risk identification but represents a technical inventory activity rather than a comprehensive risk assessment methodology. Technical limitations are one category of risk factor among many—operational, governance, data quality, and third-party risks also require assessment.