To prevent hateful or harmful responses from being returned by a generative AI solution, the correct approach is to implement content filtering. According to the Microsoft Learn documentation for Azure OpenAI Service and the Responsible AI principles, content filtering is a built-in safety mechanism that automatically screens both user prompts (inputs) and model outputs (responses) for inappropriate, harmful, or policy-violating material.
Content filters are designed to detect and block content such as:
* Hate speech or harassment
* Sexual or explicit material
* Self-harm or violent content
* Personally identifiable information (PII) misuse
In Azure OpenAI, the content filtering system is part of Microsoft's Responsible AI standard and cannot be disabled. It ensures that generative AI models such as GPT-3.5 or GPT-4 operate safely and ethically, reducing the risk of producing offensive or discriminatory text. The filter evaluates model responses in real time and can modify, block, or flag inappropriate outputs before they reach the user.
Let's review the other options:
* A. Abuse monitoring tracks misuse after deployment but does not actively prevent hateful responses.
* C. Fine-tuning customizes a model's style or domain knowledge but does not guarantee filtering of offensive content.
* D. Prompt engineering helps steer model behavior but cannot fully prevent harmful outputs.
Therefore, to proactively prevent hateful, unsafe, or offensive responses in a generative AI system built on Azure OpenAI, the correct and Microsoft-verified approach is B. Content filtering.