A bank fine-tuned a large language model to speed up loan approvals. An external audit found the model approves loans more quickly for one demographic than for others. What is the most cost-effective fix?

The most cost-effective fix is to include more diverse training data and fine-tune the existing model. This directly addresses the root cause of the bias, which is likely an imbalance or lack of representation in the original training data. Fine-tuning is less resource-intensive than pre-training a new model from scratch. Retrieval Augmented Generation (RAG) enhances factual accuracy and reduces hallucinations but does not directly mitigate bias originating from the model's core training. AWS Trusted Advisor provides cost optimization, security, and performance recommendations for AWS services, but it does not directly analyze or eliminate bias within an AI model's training data or outputs. Pre-training a new LLM is a significantly more expensive and time-consuming process, requiring vast computational resources and data, making it not the most cost-effective solution for an existing, fine-tuned model.