IBM Assessment: Foundations of AI Advanced Practice Exam: Hard Questions 2025

A financial services company is deploying an AI system that uses ensemble learning combining gradient boosting and neural networks for credit risk assessment. During validation, they notice that while individual model predictions show 85% accuracy, the ensemble system produces inconsistent results when the models disagree significantly. The neural network shows higher confidence on edge cases while gradient boosting performs better on typical cases. What is the most appropriate strategy to handle this model disagreement in production?

An AI system trained on customer service interactions from 2018-2020 is now performing poorly in 2023, showing significant accuracy degradation. Analysis reveals that customer language patterns, product terminology, and communication channels have evolved substantially. The model uses transfer learning from a pre-trained language model. Which combination of strategies would most effectively address this concept drift while minimizing computational costs and maintaining service continuity?

A healthcare AI system designed to assist radiologists in detecting lung abnormalities consistently shows high overall accuracy (92%) but demonstrates a 15% lower sensitivity for detecting abnormalities in patients from underrepresented demographic groups. The training dataset had proportional representation across demographics. Post-deployment analysis reveals the issue. What is the most likely root cause and appropriate remediation strategy?

A manufacturing company is implementing a predictive maintenance AI system using IoT sensor data from machinery. The system must operate in a hybrid architecture where some processing happens at the edge (on-premises) and some in the cloud. Latency requirements mandate that critical failure predictions occur within 100ms, while the system also needs to perform complex pattern analysis that requires significant computational resources. How should the AI processing be architected to meet these conflicting requirements?

An organization is developing a conversational AI agent that needs to handle multi-turn dialogues while maintaining context, handle ambiguous queries, and integrate with multiple backend systems. Initial testing shows that the agent loses context after 3-4 turns and struggles when users reference previous statements implicitly. The system uses a transformer-based language model with intent classification. What architectural enhancement would most effectively address these context management challenges?

A global e-commerce company's product recommendation AI system is being audited for GDPR compliance. The system uses collaborative filtering and deep learning, combining user behavior data, purchase history, and demographic information. Users are requesting data deletion under "right to be forgotten" provisions, but the data science team explains that removing individual user data from trained neural networks is technically infeasible without complete retraining. What is the most appropriate technical and governance approach to achieve compliance?

A research team is developing an AI system that combines reinforcement learning with computer vision for autonomous warehouse robots. During simulation testing, the robots achieve 95% task completion rates, but in the real warehouse environment, performance drops to 68%. Analysis shows the robots struggle with lighting variations, unexpected obstacles, and floor surface changes not present in the simulation. The deployment deadline is approaching. What is the most effective strategy to bridge this simulation-to-reality gap?

An AI-powered hiring system uses natural language processing to screen resumes and predict candidate success based on historical hiring data from 2010-2023. External auditors have identified that the system shows bias against career gaps in resumes, disproportionately affecting candidates who took parental leave. The correlation between employment gaps and job performance in the training data is weak (0.12), yet the model weights this feature significantly. What explains this behavior and what is the appropriate intervention?

A financial institution is deploying a fraud detection AI system that processes millions of transactions daily. The system uses an ensemble of models including isolation forests, autoencoders, and gradient boosting. In production, they observe that 2% of flagged transactions are false positives, but these represent 40% of their high-value customer transactions, creating significant customer service issues. The precision-recall tradeoff is already optimized for their target operating point. What systemic issue is most likely causing this pattern and how should it be addressed?

An organization is implementing an explainable AI (XAI) framework for a critical decision-support system used in medical diagnosis. They need to provide explanations to three different stakeholder groups: patients (non-technical), physicians (domain experts), and regulators (compliance focus). The underlying model is a deep neural network ensemble. Initial attempts using SHAP values have confused patients and frustrated physicians who find them inconsistent with medical reasoning. What comprehensive explainability strategy should be implemented?