Answer: Detect imbalances or disparities in the data., Evaluate the model's behavior so that the company can provide transparency to stakeholders.

## Detailed Explanation To build and use an AI model responsibly to minimize bias in a loan discount application scenario, the company should focus on two key areas: **data quality assessment** and **model behavior evaluation for transparency**. ### **Correct Answers: A and C** **A: Detect imbalances or disparities in the data** - **Why this is optimal**: Bias in AI models often originates from the training data itself. If certain demographic groups (e.g., based on age, gender, ethnicity, or geographic location) are underrepresented or overrepresented in the dataset, the model may learn patterns that unfairly disadvantage those groups. For a loan company, this could lead to discriminatory lending practices that violate fair lending regulations like the Equal Credit Opportunity Act (ECOA). - **Implementation approach**: The company should conduct thorough data analysis to identify statistical disparities across protected classes, use techniques like demographic parity analysis, and potentially apply data augmentation or re-sampling methods to create a more balanced dataset before model training. - **Connection to responsible AI**: This addresses bias at its source, aligning with AWS's responsible AI principles of fairness and transparency by ensuring the model isn't perpetuating historical biases present in the data. **C: Evaluate the model's behavior so that the company can provide transparency to stakeholders** - **Why this is optimal**: Even with balanced data, models can develop complex decision patterns that may inadvertently discriminate. Regular evaluation using fairness metrics (like disparate impact analysis, equal opportunity difference) across different customer segments is essential to detect such issues. - **Stakeholder transparency**: For a financial institution, regulators, customers, and internal auditors require explanations of how AI-driven decisions are made. Techniques like SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) can help explain individual predictions, while overall model behavior documentation supports regulatory compliance and builds trust. - **Connection to responsible AI**: This aligns with AWS's emphasis on model interpretability and accountability, ensuring the company can justify decisions and demonstrate compliance with financial regulations. ### **Why Other Options Are Less Suitable** **B: Ensure that the model runs frequently** - While regular model retraining can be important for maintaining accuracy with changing data patterns, frequency of execution doesn't directly address bias minimization. A biased model running frequently would simply perpetuate unfair outcomes more often. **D: Use the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) technique to ensure that the model is 100% accurate** - ROUGE is specifically designed for evaluating text summarization models by comparing generated summaries to reference summaries. It's completely inappropriate for a loan discount prediction task. Additionally, aiming for 100% accuracy is unrealistic in real-world AI applications and doesn't address fairness concerns. **E: Ensure that the model's inference time is within the accepted limits** - Inference time optimization is important for user experience and operational efficiency, but it's a performance consideration rather than a bias mitigation strategy. A fast but biased model would still negatively impact customers unfairly. ### **Best Practices Integration** The combination of options A and C represents a comprehensive approach: 1. **Proactive bias prevention** through data analysis (A) 2. **Ongoing monitoring and accountability** through model evaluation (C) This two-pronged strategy addresses both the input (data) and output (model decisions) aspects of responsible AI development, which is particularly critical in regulated industries like financial services where algorithmic decisions have significant real-world consequences for customers.