IAPP Artificial Intelligence Governance Professional (AIGP) — Question 26
CASE STUDY -
Please use the following to answer the next question:
ABC Corp. is a leading insurance provider offering a range of coverage options to individuals. ABC has decided to utilize artificial intelligence to streamline and improve its customer acquisition and underwriting process, including the accuracy and efficiency of pricing policies.
ABC has engaged a cloud provider to utilize and fine-tune its pre-trained, general purpose large language model (“LLM”). In particular, ABC intends to use its historical customer data – including applications, policies, and claims – and proprietary pricing and risk strategies to provide an initial qualification assessment of potential customers, which would then be routed to a human underwriter for final review.
ABC and the cloud provider have completed training and testing the LLM, performed a readiness assessment, and made the decision to deploy the LLM into production. ABC has designated an internal compliance team to monitor the model during the first month, specifically to evaluate the accuracy, fairness, and reliability of its output. After the first month in production, ABC realizes that the LLM declines a higher percentage of women’s loan applications due primarily to women historically receiving lower salaries than men.
During the first month when ABC monitors the model for bias, it is most important to:
Answer options
- A. Continue disparity testing.
- B. Analyze the quality of the training and testing data.
- C. Compare the results to human decisions prior to deployment.
- D. Seek approval from management for any changes to the model.
Correct answer: A
Explanation
The correct answer is A, as continuing disparity testing is crucial to identify and address any biases in the model's decisions, especially regarding the higher decline rates for women's applications. While analyzing data quality (B), comparing results to human decisions (C), and seeking management approval (D) are important steps, they do not directly address the immediate need to monitor and rectify bias in the model's outputs.