IAPP Artificial Intelligence Governance Professional (AIGP)

✅ Option B (Correct)
Reasoning: Continuous monitoring is crucial to track an AI tool's live performance, detect issues like model or data drift, bias, or performance degradation, and ensure it consistently meets its intended objectives. Adjustments (e.g., retraining, recalibration) are then made based on these insights to maintain efficacy and alignment.
❌ Why the other choices are incorrect:

• Option A is incorrect: Monitoring provides human oversight; its primary goal is not to enable independent algorithm evolution without intervention, but rather to ensure controlled and responsible operation.
• Option C is incorrect: While privacy is vital, restricting data access is a control measure, not the primary objective of continuously monitoring the tool’s performance against its goals.
• Option D is incorrect: Upgrading computational resources relates to infrastructure management. Continuous monitoring primarily focuses on the AI model's output, behavior, and effectiveness, not solely hardware efficiency.

Understanding the algorithmic methodology (how the AI model works) is primarily relevant during the model's development, deployment, and operational phases for risk assessment, performance, and explainability. During deactivation, the focus shifts to ensuring proper shutdown, managing data, and mitigating external impacts. The other options directly address critical aspects of safely and compliantly decommissioning an AI system.

✅

Option C (To establish accountability) (Correct)

Reasoning: Collaboration among cross-functional teams is crucial for AI governance to establish clear accountability. AI systems involve complex ethical, legal, technical, and business risks. Diverse perspectives ensure comprehensive risk identification and assignment of responsibilities for design, development, deployment, and monitoring, which is fundamental to responsible AI throughout its lifecycle.

Why the other choices are incorrect:

• Option A is incorrect: Collaboration among internal cross-functional teams does not inherently aim to minimize third-party involvement. External partnerships or vendors may still be essential for specialized services, data, or infrastructure, regardless of internal collaboration efforts.

• Option B is incorrect: While establishing a user-centric design is important and benefits from cross-functional input, it focuses on the user experience aspect. Establishing accountability is a broader and more fundamental governance concern, encompassing not just user experience but also ethical, legal, and societal impacts.

• Option D is incorrect: Minimizing potential liability to users is a significant goal, but it is an outcome or a component of broader risk management and accountability. Establishing clear accountability (Option C) provides the framework and responsibilities necessary to proactively identify, mitigate, and manage risks, thereby reducing potential liability.

The consulting firm is explicitly retained "to supplement its small data science team and help develop the algorithm." Selecting the specific algorithm type is a core technical task within algorithm development. Given the healthcare network's internal data science team is described as "small" and requires supplementation, the consulting firm provides the critical expertise and capacity for this highly technical decision. They are, therefore, most important in this specific technical selection.

Reference: N/A - General AIGP principles regarding project staffing and technical expertise.

✅ Option D (Correct) Reasoning: Offering students the choice to utilize generative AI in limited, defined circumstances directly supports learning how to use AI as a supportive educational tool. This enables students to understand AI's capabilities, limitations, and ethical integration into their learning processes, aligning with GVC's efforts to incorporate and monitor AI use for educational benefit. ❌ Why the other choices are incorrect:

• Option A is incorrect: While AI can sometimes aid efficiency, learning time management is not the primary or best reason for an educational institution to deliberately offer generative AI use.
• Option B is incorrect: Learning about performing research often involves critical source evaluation. Solely focusing on AI for research without emphasizing its supportive role risks misunderstanding its limitations and ethical considerations.
• Option C is incorrect: While related, "practical applications of AI" is broader. Option D, "AI as a supportive educational tool," more precisely frames the pedagogical value within an academic context, focusing on how AI enhances learning under controlled usage.

Reference: UNESCO. (2023). Guidance for generative AI in education and research. This guidance emphasizes the importance of AI literacy and understanding AI's role as a supportive tool in learning environments.

The scenario exemplifies 'hallucination,' where a generative AI produces plausible but factually incorrect or nonsensical information. The tool inventing a non-existent restaurant and providing false details about others directly aligns with this definition. Prompt injection involves manipulating AI behavior with specific instructions. Model collapse is a degradation in generative models due to training on synthetic data. Overfitting refers to a model performing poorly on new data after over-learning its training data.

✅ Option C (Correct)
Reasoning: High false positives mean the model is flagging legitimate transactions, causing customer inconvenience and operational strain. Deactivating it for assessment prevents further harm and allows for thorough investigation, recalibration, or retraining to restore accuracy.

❌ Why the other choices are incorrect:

• Option A is incorrect: While monitoring is crucial, simply dedicating more resources to it doesn't solve the underlying model inaccuracy causing the false positives; it only detects them more efficiently.
• Option B is incorrect: Maintaining records is valuable for post-mortem analysis and future improvements, but it does not immediately address or stop the ongoing problem of the model generating incorrect fraud alerts.
• Option D is incorrect: Training customer service helps manage the fallout from false positives but is a mitigation strategy for symptoms, not a direct solution to fix the inaccurate model itself or reduce its error rate.

✅ Option D (Correct)
Reasoning: Ethical impact assessments for AI should be initiated early in development to proactively identify and address potential risks. Continuing these assessments throughout the entire lifecycle ensures ongoing monitoring, adaptation to new contexts, and mitigation of emerging ethical challenges, aligning with responsible AI governance.

❌ Why the other choices are incorrect:

• Option A is incorrect: Waiting until production is too late. Ethical issues embedded in design or data are harder and more costly to fix at this stage, risking deployment of problematic systems.
• Option B is incorrect: Tying assessments to user thresholds is arbitrary and reactive. Ethical concerns can arise regardless of user count, necessitating early and continuous evaluation.
• Option C is incorrect: Performing assessments only after issues emerge is reactive. The goal is proactive prevention and mitigation of harm, not just post-incident response.

✅ Option B (Correct)

Reasoning: Cognitive learning is a concept primarily from human psychology and education, describing how humans acquire and process information. It is not a distinct algorithmic type or paradigm within machine learning, unlike supervised, unsupervised, or reinforcement learning.

❌ Why the other choices are incorrect:

• Option A is incorrect: Deep learning is a specialized and widely recognized subset of machine learning, using neural networks with multiple layers to learn representations from data.
• Option C is incorrect: Unsupervised learning is a fundamental machine learning paradigm where models discover patterns in unlabeled datasets, crucial for tasks like clustering and dimensionality reduction.
• Option D is incorrect: Reinforcement learning is a common machine learning approach where an agent learns to make optimal decisions through trial and error in an environment, maximizing rewards.

✅ Option D: Supervised learning (Correct)

Reasoning: Supervised learning is a machine learning paradigm where models are trained using labeled datasets. Each data point includes both input features and the desired output label, allowing the algorithm to learn the mapping between them.

❌ Why the other choices are incorrect:

• Option A is incorrect: Segmentation is a computer vision task (e.g., image segmentation), not a fundamental subcategory of AI/ML describing how algorithms are trained with labeled datasets.
• Option B is incorrect: Generative AI focuses on creating new content. While it can use labeled data, it's a type of AI output, not the general training paradigm described by using labeled datasets.
• Option C is incorrect: Expert systems are an older AI branch based on rules and knowledge bases, not primarily on statistical learning from labeled datasets like modern machine learning approaches.

Reference: https://www.ibm.com/topics/supervised-learning