An accounting firm wants to implement a large language model (LLM) to automate document processing. The firm must proceed responsibly to avoid…

Developing and deploying a large language model (LLM) responsibly, especially in a sensitive domain like accounting, requires a focus on fairness and bias mitigation to avoid potential harms. The two most fundamental actions are addressing bias at its source and continuously measuring for fairness. Modifying the training data is a proactive, pre-processing step to mitigate systemic biases the model might learn. Including fairness metrics in the evaluation process is a critical step to quantify and understand any remaining biases in the model's performance, ensuring it does not disproportionately harm specific groups. Together, these actions form a foundational approach to responsible AI development.

1. Google Cloud. (2023). Introduction to Responsible AI. Google Cloud Skills Boost. In the module on "Identifying and mitigating bias

" it is stated that "Mitigating bias can happen at three points in the modeling pipeline: in the training data (preprocessing)

in the model itself (in-processing)

and in the model's predictions (post-processing)." This supports option C. The module on "Evaluating models for fairness" emphasizes using fairness metrics

supporting option A.

2. Microsoft. (2022). Microsoft Responsible AI Standard

v2. Microsoft AI. Section 4.2

"Fairness

" outlines requirements to "Identify and mitigate fairness harms" (p. 11). It specifies the need for "fairness assessment and mitigation throughout the AI system lifecycle

" which includes both analyzing training data for bias (supporting C) and using "fairness metrics to assess the AI system" (supporting A).

3. Hovy

D.

& Prabhumoye

S. (2021). Five Sources of Bias in Natural Language Processing. ACM Computing Surveys

54(8)

1-30. https://doi.org/10.1145/3471455. This peer-reviewed paper discusses that a primary source of bias is the training data itself (Section 3

"Data"). It highlights data augmentation and re-sampling as key mitigation strategies

directly supporting option C.

4. Narayanan

A. (2023). Lecture 15: Fairness. Princeton University

COS 597G: Understanding Large Language Models. The lecture materials emphasize two main approaches for fairness: (1) pre-processing data to remove bias and (2) post-processing model outputs. It also covers the necessity of defining and measuring fairness using specific metrics

which directly supports options C and A as core strategies.