Plan and manage an Azure AI solution 1.4 Implement responsible AI across generative AI and agentic systems
You are customizing a content filter for a model deployment in Foundry.
Study guide skill: Configure safety filters, guardrails, risk detection, and content moderation
Question B1-009
You are customizing a content filter for a model deployment in Foundry. Which statement about the hate, sexual, violence and self-harm categories is correct?
Select one answer.
Answer: D
The harm categories use four severity levels (safe, low, medium and high). Low, medium and high are configurable for both input and output filters, and safe content is labeled in annotations but isn't filtered or configurable. Filters can be applied to prompts as well as completions.
Deep dive
How it works
Foundry's content filtering system, powered by Azure AI Content Safety, runs both the prompt and the completion through classification models. Its neural multiclass models cover four categories (hate, sexual, violence and self-harm) at four severity levels: safe, low, medium and high. Content rated safe is labeled in annotations, but it isn't filtered and isn't configurable.
You can adjust the settings separately for prompts and completions. For the four harm categories, three severity thresholds are configurable:
- Low, medium, high: the strictest setting. Content at low, medium and high is filtered.
- Medium, high: low isn't filtered, and medium and high are.
- High: only high-severity content is filtered.
Turning filters off entirely ("No filters") or using "Annotate only" requires approval for completions. Other optional models, such as jailbreak risk detection and protected material detection, are binary classifiers rather than severity scales. You can use any deployment with the default content filter, or create your own to make it stricter or more lenient. Keep in mind that the filter doesn't apply to embedding models or audio models such as Whisper.
The other options, and when you'd use them
- Severity levels numeric from 1 to 10, set per user: Foundry's harm filter uses the four named levels above. (The standalone Content Safety text API uses a 0-7 scale, which is a different surface.)
- Set the threshold to safe so every response is blocked: Safe content is labeled in annotations but isn't filtered, and safe isn't configurable.
- Categories apply only to outputs: The system detects and acts on harmful content in both input prompts and output completions, and you configure input filters and output filters separately.
Exam tip
Remember "four categories, four levels, three configurable thresholds". Safe is annotate-only, and turning filtering off for completions needs approval.
References
Last reviewed: