How to Fix LLM Bias in Loan Approvals Most Cost-Effectively?

A 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?

  1. Include more diverse training data. Fine-tune the model again by using the new data. Source Reference Answer
  2. Use Retrieval Augmented Generation (RAG) with the fine-tuned model.
  3. Use AWS Trusted Advisor checks to eliminate bias.
  4. Pre-train a new LLM with more diverse training data.

Community Votes

A
100%

100% of anonymous learners picked answer A. Votes are pick records left by other test-takers — they are not the verified answer.

Community Insight

The exam tests understanding that bias originates from unrepresentative training data, and fine-tuning with diverse data is the targeted, low-cost fix—whereas pre-training a new model is expensive and RAG does not change model behavior.

AIF-C01 candidates should know that the most cost-effective way to address demographic bias in a fine-tuned LLM is to include more diverse training data and fine-tune again. Community consensus strongly supports option A over alternatives like RAG, Trusted Advisor, or pre-training a new model.

Choosing option D (pre-train a new LLM) is the most common mistake because it seems comprehensive, but pre-training from scratch is far more expensive and unnecessary when fine-tuning with diverse data can correct the bias.

Community Discussion (3 comments)

Jessiii 👍 1 Selected: A
The model's bias likely stems from unrepresentative training data. Adding more diverse data and fine-tuning the model is the most cost-effective solution to address bias.
may2021_r 👍 1 Selected: A
The correct answer is A. Fine-tuning with more diverse data is the most cost-effective bias mitigation approach.
aws_Tamilan 👍 1 Selected: A
A. Include more diverse training data. Fine-tune the model again by using the new data. Explanation: The issue of bias in the loan approval model likely arises from the model being trained on data that does not sufficiently represent all demographics. To address this, the bank should augment the training dataset with more diverse data to ensure that the model can learn to make fair and equitable decisions across different demographics. After incorporating the more diverse training data, the bank can fine-tune the model again to adjust its behavior and reduce any biases identified during the audit.

Comments & Corrections

No comments yet — spotted an error or have a note? Share it below.

Log in to comment, report an error, or add a note about this question.

Submitted for moderation before publishing. Keep it helpful and respectful.

Expert Analysis

Why the Answer Is Correct

Option A is correct because the bias in the loan approval model is almost certainly a result of the model being trained on data that underrepresents certain demographics. Adding more diverse training data and fine-tuning the same model directly addresses the root cause without the need to rebuild the model infrastructure. This is the most cost-effective approach because fine-tuning is significantly cheaper than pre-training a new LLM, and it leverages the existing model's learned knowledge.

Why the Other Options Are Wrong

Option B (RAG) is incorrect because RAG retrieves external information at inference time and does not alter the model's learned biases; it is not designed to fix training-data imbalance. Option C (AWS Trusted Advisor) is irrelevant because Trusted Advisor provides infrastructure and security recommendations, not bias detection or correction in ML models. Option D (pre-train a new LLM) is technically possible but extremely expensive and wasteful; it would require vast computational resources and new base data, whereas fine-tuning is sufficient and far cheaper.

Community Comment Notes

Community comments consistently support option A, noting that "the model's bias likely stems from unrepresentative training data" and that "fine-tuning with more diverse data is the most cost-effective bias mitigation approach." One comment explicitly emphasizes that the issue arises from data that does not sufficiently represent all demographics, reinforcing that the correct action is to augment the training dataset and fine-tune. No comments advocate for the other options, indicating strong consensus on A.

Official Reference

Exam Strategy

When a question asks for the most cost-effective bias fix in an LLM, always ask whether the issue can be solved by fine-tuning existing model weights with better data. Avoid choosing expensive retraining or irrelevant AWS services unless the question explicitly requires them.

Related Analysis

Practice All AIF-C01 Questions

Access 100 questions with complete answers and detailed explanations.

View Full AIF-C01 Practice Test →

← Back to AIF-C01 Study Guide