How to Develop and Deploy LLMs Responsibly?

An accounting firm wants to implement a large language model (LLM) to automate document processing. The firm must proceed responsibly to avoid potential harms. What should the firm do when developing and deploying the LLM? (Choose two.)

Include fairness metrics for model evaluation. Source Reference Answer
Adjust the temperature parameter of the model.
Modify the training data to mitigate bias. Source Reference Answer
Avoid overfitting on the training data.
Apply prompt engineering techniques.

Community Votes

100%

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

Community Insight

This question tests the core pillars of responsible AI development—fairness evaluation and bias mitigation at the data level—rather than general ML tuning techniques.

When deploying LLMs for sensitive tasks like document processing in accounting, firms must include fairness metrics and modify training data to mitigate bias. These practices ensure equitable, ethical AI outcomes and prevent harmful discrimination.

Candidates often choose options like adjusting temperature or applying prompt engineering, confusing general LLM tuning with responsible AI practices that specifically address fairness and bias.

Community Discussion (6 comments)

Rcosmos 👍 1

Explicação: Desenvolver e implantar LLMs com responsabilidade envolve garantir que o modelo seja justo, ético, e evite causar danos involuntários. As duas práticas mais alinhadas com esse objetivo são:A. Incluir métricas de imparcialidade para avaliação do modelo.Avaliar a imparcialidade (fairness) garante que o modelo não gere resultados discriminatórios ou enviesados, especialmente importante em áreas sensíveis como contabilidade e finanças. C. Modificar os dados de treinamento para mitigar o viés.O viés muitas vezes está nos dados. Corrigir ou balancear o conjunto de dados pode ajudar a reduzir decisões injustas ou distorcidas do modelo.

kopper2019 👍 1

A. Include fairness metrics for model evaluation. C. Modify the training data to mitigate bias.

Jessiii 👍 1 Selected: AC

A. Include fairness metrics for model evaluation: Fairness metrics help evaluate whether the model treats all groups fairly and does not introduce harmful bias. When developing an LLM for document processing, it's essential to assess and mitigate any potential biases in the model's outputs to ensure that it operates responsibly and equitably, especially when the model interacts with sensitive data. C. Modify the training data to mitigate bias: Bias in AI models often originates from biased training data. Modifying the training data to ensure it is representative and free from harmful biases is a crucial step in reducing the risk of the model producing unfair or harmful outcomes. This proactive approach is essential for responsible deployment.

dspd 👍 1 Selected: AC

A: Include fairness metrics for model evaluation Critical for responsible AI implementation Helps identify discriminatory patterns Ensures equitable treatment across different groups Allows for continuous monitoring of fairness Essential for an accounting firm handling sensitive financial data C: Modify the training data to mitigate bias Addresses bias at the source Ensures representative training data Helps prevent discriminatory outcomes Critical for fair treatment of all clients Fundamental to responsible AI development

KawtarZ 👍 1 Selected: BE

A. no need for fairness metrics as the use case is for document processing C. modifying the training data means there is a re-training of the model. not needed for this use case D. there is no re-training needed for this case. avoiding overfitting is also not needed

jove 👍 3 Selected: AC

A. Include fairness metrics for model evaluation: Fairness metrics help ensure that the model is not biased against any particular group. This is especially important in fields like accounting, where any biases in automated decisions could lead to unethical outcomes. Fairness metrics provide insight into how well the model treats all data groups equally. C. Modify the training data to mitigate bias: Adjusting the training data to address any identified biases is crucial for developing responsible AI applications. This can involve balancing the dataset or removing biased samples, ensuring the model generalizes fairly across different data types and groups.

Comments & Corrections

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

Expert Analysis

Understanding Responsible AI in LLM Deployment

The question asks what an accounting firm should do to develop and deploy an LLM responsibly to avoid potential harms. Responsible AI encompasses principles such as fairness, transparency, accountability, and bias mitigation.

Why Option A is Correct

Including fairness metrics for model evaluation is a foundational practice in responsible AI. Fairness metrics allow organizations to quantitatively assess whether the LLM produces equitable outcomes across different demographic groups, document types, or data distributions. In an accounting context—where the LLM may process financial documents, tax records, or client data—biased outputs could lead to discriminatory decisions or regulatory violations. Community members like dspd and Jessiii correctly emphasize that fairness metrics are essential for identifying discriminatory patterns and ensuring continuous monitoring.

Why Option C is Correct

Modifying the training data to mitigate bias addresses the root cause of many AI harms. If the training data contains historical biases—such as underrepresentation of certain groups or skewed financial patterns—the model will perpetuate or amplify those biases. By curating, balancing, or augmenting the training dataset, the firm directly tackles bias at its source. As jove points out, this is especially critical in fields like accounting where biased automated decisions could lead to unethical or legally problematic outcomes.

Why the Other Options Are Incorrect

Option B (Adjust the temperature parameter): Temperature controls the randomness of the model's outputs. While useful for tuning creativity vs. determinism, it is a general inference parameter and does not address fairness, bias, or responsible AI concerns.
Option D (Avoid overfitting on the training data): Overfitting is a standard machine learning concern related to model generalization. While important for model performance, it is not specifically tied to responsible AI or harm prevention.
Option E (Apply prompt engineering techniques): Prompt engineering optimizes how the model is queried but does not inherently address systemic bias or fairness. It is a usability technique, not a responsible AI safeguard.

Community consensus strongly supports A and C, with 71% of voters selecting this combination. A minority (e.g., KawtarZ) argued against retraining, but modifying training data is a recognized best practice in responsible AI frameworks.

Official Reference

Exam Strategy

When a question emphasizes 'responsible' AI deployment, focus on options that directly address fairness, bias mitigation, transparency, and ethical safeguards. Ignore general ML tuning or inference parameters unless they explicitly relate to harm prevention.

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