AI-900 Practice Question: Describe features of generative AI workloads on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Physical barriers in AI data centers for safety

  Why it's wrong here

  Physical barriers such as secure cages, biometric access, and multi-factor authentication in Azure datacenters protect hardware and data from unauthorized physical intrusion—they are part of the physical security perimeter and business-continuity design. They have no ability to evaluate or constrain the text, images, or code generated by an AI model. Calling these 'guardrails' confuses datacenter infrastructure-level security with the algorithmic/content-level safety controls that shape model behavior.

• ✓

  Controls and filters that prevent generative AI from producing harmful or inappropriate outputs

  Why this is correct

  Guardrails are the combination of configuration, software filters, and validation layers applied to a generative AI model so that its outputs stay within safe, responsible boundaries. In Azure AI this includes Azure AI Content Safety category filters (hate, sexual, self-harm, violence), prompt-shield/jailbreak detection, groundedness detection for RAG, and output moderation before a completion is returned to the user. These controls operate at inference time to block or blocklist/reword harmful or inappropriate content, which is exactly what makes them guardrails.

• ✗

  Rate limiting controls to prevent API overuse

  Why it's wrong here

  While rate limiting is an essential Azure API management control—capping requests per minute or per deployment to manage cost, throttling, and backend capacity—it does not inspect or alter the semantic content of model outputs. Azure AI Content Safety guardrails, on the other hand, evaluate prompts and completions against safety categories (hate, sexual, violence, self-harm) during inference. Thus rate limits address resource governance, not the responsible-AI objective of preventing harmful generation.

• ✗

  Version control systems for managing model updates

  Why it's wrong here

  Model versioning in Azure Machine Learning—using registered models, version numbers, and lineage metadata—helps teams reproduce experiments, compare performance, and roll back to earlier snapshots, but it is a lifecycle-management practice, not a runtime safety control. A versioned model can still emit harmful or toxic content if no safety filters are applied. Guardrails are behavioral constraints embedded in the model-serving pipeline (for example, content filters and output grounding), which is an entirely different layer from code/model version tracking.

Go deeper

Related to this question