IBM A1000-047 - Assessment: Foundations of AI Advanced Practice Exam: Hard Questions 2025

A financial services company is implementing an AI system to detect fraudulent transactions. During testing, they discover that the model has 98% accuracy overall but fails to identify 60% of actual fraud cases while maintaining low false positives. The training dataset contains 100,000 legitimate transactions and 500 fraudulent ones. What is the PRIMARY issue causing this performance problem?

An enterprise is designing an AI governance framework and needs to implement model lineage tracking across multiple teams using different ML platforms (IBM Watson Studio, open-source frameworks, and third-party tools). They require full traceability from raw data sources through feature engineering, model training, deployment, and inference. Which architectural approach BEST addresses this complex multi-platform requirement?

A Watson Natural Language Understanding service is processing customer feedback and consistently misclassifies industry-specific terminology. The default model identifies 'churn rate optimization' as negative sentiment when it's actually a positive business outcome in their context. The company has 5,000 labeled domain-specific documents. What is the MOST effective strategy to address this issue while maintaining cost-efficiency?

A healthcare AI system uses ensemble learning combining decision trees, neural networks, and support vector machines to predict patient readmission risk. During validation, you discover that while the ensemble achieves 89% accuracy, the individual models show 87%, 84%, and 86% accuracy respectively. However, when analyzing error patterns, you find that all three models make similar mistakes on the same patient subgroups. What does this indicate and what should be the NEXT course of action?

An organization is implementing a conversational AI assistant using IBM Watson Assistant that needs to handle multi-turn conversations involving complex financial transactions with strict regulatory requirements. Users often provide incomplete information across multiple turns, and the system must maintain context while ensuring all required information is collected before executing transactions. Which architectural pattern BEST addresses these requirements?

A company deploying a computer vision model for quality control in manufacturing discovers that the model performs excellently (95% accuracy) on the test set but poorly (73% accuracy) in production. Investigation reveals that the production environment has different lighting conditions and camera angles than the training data. The model uses transfer learning from a pre-trained ImageNet model. What combination of strategies would MOST effectively address this issue?

An AI ethics committee is evaluating an algorithmic decision system that will be used for loan approvals. During fairness testing, they discover that while the model shows demographic parity (equal approval rates across protected groups), there is significant disparity in true positive rates—qualified applicants from minority groups are rejected at higher rates than equally qualified majority group applicants. What does this situation represent and what is the appropriate response?

A global enterprise is implementing an AI strategy involving multiple use cases across different business units. They need to decide between building custom models, using pre-trained foundation models with prompt engineering, fine-tuning foundation models, or using specialized AI services like Watson. For a use case involving extracting structured information from highly specialized legal contracts with unique clauses specific to their industry, which approach offers the BEST balance of performance, cost, and time-to-value?

A data science team is debugging a deep learning model that shows training accuracy increasing steadily to 95% while validation accuracy plateaus at 78% after the first few epochs. They've already implemented dropout, L2 regularization, and early stopping. Analysis reveals that validation loss starts increasing after epoch 10 while training loss continues decreasing. What is the MOST likely root cause and appropriate next step?

An organization is deploying an AI-powered medical diagnostic assistant that provides treatment recommendations to healthcare providers. From an AI governance and ethics perspective, which architectural approach BEST addresses the unique responsibilities and risks associated with high-stakes medical AI applications?