Microsoft AB-100 Exam - Topic 3 Question 4 Discussion

The scenario describes a deployed AI solution using Azure OpenAI that exhibits bias (creating disparate outcomes based on customer traits). This directly impacts the Fairness principle of Microsoft's Responsible AI framework.

Why 'Modify the system instructions' is the Correct Strategy:

Direct Control via System Metaprompts: In large language model (LLM) applications like those powered by Azure OpenAI, the system instructions (or system message) define the behavior, constraints, and tone of the model. By modifying these instructions, you can explicitly direct the model to treat all customer segments equitably and ignore specific sensitive traits when drafting marketing content.

Mitigation without Re-engineering: * Option B and D (Training/Retraining): Azure OpenAI models are foundation models. Most companies use them via API and do not have access to the original 'training dataset' to modify it. While fine-tuning is possible, it is significantly more expensive and complex than prompt engineering.

Option C (Randomization): Randomization does not solve bias; it creates inconsistency and potentially irrelevant content, violating the Reliability and Safety principle.

Alignment with Responsible AI: Microsoft's documentation on Fairness recommends 'Instructional Mitigation.' This involves adding specific rules to the system prompt, such as: 'You must ensure the tone and value proposition of the email remain consistent across all demographic groups' or 'Do not use customer traits such as age or gender to influence the core marketing message.'