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

A financial services company is developing an AI system to automate loan approval decisions. During model validation, you discover that the model achieves 92% accuracy overall but has a 15% false negative rate for minority applicants compared to 6% for majority applicants, while false positive rates are similar across groups. The business stakeholder argues that 92% accuracy exceeds their target and wants to deploy immediately. As the AI ethics lead, what is the most appropriate recommendation?

An enterprise is building a conversational AI solution that must handle domain-specific queries about complex technical documentation. Initial testing with a pre-trained Watson Assistant shows poor understanding of technical terminology and inconsistent responses. The team has 5,000 labeled conversation logs and access to 200,000 pages of technical documentation. What architectural approach would most effectively improve performance?

A data science team trained a deep neural network for image classification that achieves 94% validation accuracy but only 78% accuracy on production data. After analysis, they discover the training images were collected in controlled lighting conditions while production images come from varied real-world environments. The model also shows high confidence scores (>0.95) even for incorrect predictions. What combination of techniques would most effectively address both issues?

An organization implements Watson Natural Language Understanding to extract entities and sentiment from customer service transcripts across multiple languages. They notice that sentiment scores for negative feedback in Japanese and Arabic are consistently less negative than equivalent English feedback, while entity extraction recall is 20% lower for non-English languages. What is the most likely root cause and appropriate solution?

A healthcare AI system uses ensemble learning combining three models: a deep learning model (85% accuracy, 0.88 AUC), a gradient boosting model (83% accuracy, 0.91 AUC), and a logistic regression model (79% accuracy, 0.86 AUC) for disease prediction. The current ensemble uses simple majority voting. Patient safety requires minimizing false negatives even at the cost of some false positives. How should the ensemble strategy be optimized?

A manufacturing company deploys a predictive maintenance AI model trained on sensor data from 2018-2021. Six months post-deployment, the model's precision drops from 82% to 64% while recall remains stable. Investigation reveals that new equipment firmware updates changed sensor calibration slightly, and the company has started predictive maintenance on younger equipment. What phenomenon is occurring and what is the most comprehensive remediation strategy?

An organization is designing an AI governance framework for multiple AI projects spanning computer vision, NLP, and predictive analytics. They need to balance innovation speed with risk management. Which governance approach best aligns with IBM's AI ethics principles while maintaining operational efficiency?

A global retailer uses Watson Assistant to handle customer inquiries. They observe that conversation abandonment rates spike when users switch topics mid-conversation or ask complex multi-part questions. Analytics show the assistant correctly identifies intents but provides incomplete responses. The team has implemented context variables and slots. What architectural enhancement would most effectively address this limitation?

A company develops an AI system to screen job applicants by analyzing resumes and predicting job performance. During testing, they discover that veterans receive lower scores due to employment gaps and non-traditional career paths, even though historical data shows veterans perform as well as non-veterans once hired. The model was trained on 10 years of hiring data and actual performance reviews. What is the primary AI ethics issue and most appropriate intervention?

An enterprise implements a multimodal AI system combining Watson Visual Recognition for product images, Watson Natural Language Understanding for descriptions, and a custom neural network for final classification. In production, the system shows 89% accuracy, but analysis reveals that when visual and text predictions conflict, the final classification is correct only 62% of the time. The visual model has 91% standalone accuracy and the text model has 87% standalone accuracy. What is the most likely cause and solution?