AIF-C01 Practice Question 2: Traditional ML vs Foundation Model
AWS Certified AI Practitioner (AIF-C01) · Test 1 · question 2 of 10
A bank is automating approval decisions for small-business loans using structured application data such as revenue, credit history, and loan amount. Regulators require the bank to give each rejected applicant the specific factors behind the decision, and the same application must always produce the same outcome. Which approach best fits these requirements?
Show answer and explanation
Answer:
Foundation models (FMs) aren't always the right tool. This is a prediction task on structured, tabular data with labeled history, and it comes with regulatory explainability and consistency requirements. A traditional supervised model, such as a classifier trained on past loan outcomes, fits that well. It returns the same output for the same input, and feature-attribution techniques can show which inputs (for example, credit history) pushed each decision toward rejection (option D).
- A is wrong because an LLM's free-text reasoning isn't a reliable, auditable account of how it decided. With sampling, its output can also vary between runs, and it's a costly way to score tabular data. - B is wrong because fine-tuning changes the model's behavior but doesn't make its decision process more explainable. Generating a letter doesn't answer the regulator's question of *why* the loan was rejected. - C is wrong because image generation doesn't make or explain a lending decision at all.
Reference: Model explainability