Question 38
Domain 4: Guidelines for Responsible AIA bank has fine-tuned a large language model (LLM) to expedite the loan approval process. During an external audit of the model, the company discovered that the model was approving loans at a faster pace for a specific demographic than for other demographics. How should the bank fix this issue MOST cost-effectively?
Correct answer: A
Explanation
The model shows “consistent errors for underrepresented groups,” which the source links to bias from insufficient representative data. AWS says minimizing bias requires “enough high-quality, representative training data,” so adding diverse data and fine-tuning again is the most cost-effective fix.
Why each option is right or wrong
A. Include more diverse training data. Fine-tune the model again by using the new data.
AWS ties consistent performance gaps for a specific demographic to bias caused by insufficiently representative training data, and the prescribed mitigation is to improve the dataset rather than change the deployment logic. Under the AWS SageMaker Clarify guidance, representative data is the lowest-cost fix: add diverse examples, then fine-tune again so the model learns the missing patterns and reduces the systematic error.
B. Use Retrieval Augmented Generation (RAG) with the fine-tuned model.
RAG improves retrieval-grounded generation, not demographic bias in model training data.
C. Use AWS Trusted Advisor checks to eliminate bias.
Trusted Advisor checks service configuration and cost, not model fairness or bias.
D. Pre-train a new LLM with more diverse training data.
Pre-training a new LLM is far more expensive than re-fine-tuning an existing model.