Skip to main content
Content moderation answers a narrow question: is this content allowed to enter or leave the system. It is about the content itself, not about who is trying to control the model or whether the request is within product scope. Anthale runs this guardrail when your policy enables it on the input or output path. That lets you apply moderation before model invocation, after generation, or both.

​
What moderation covers

Typical moderation coverage includes hate and harassment, sexual content, violence and self-harm, illicit guidance, and other categories your policy treats as disallowed for the product.

​
Common moderation categories

Moderation categories are never purely theoretical. They need to reflect your product, geography, and risk posture. A customer-support assistant, an internal operations assistant, and a moderation pipeline for user-generated content will not all draw the same line in the same place.

​
Where it helps most

This guardrail matters on both sides of the model. On the input path, it limits what the model is asked to process. On the output path, it limits what the application is willing to return or automate. It is especially important when your application rewrites, summarizes, translates, or otherwise transforms user-supplied material.

​
When to moderate input, output, or both

Moderate input when your application should refuse harmful requests before the model sees them. Moderate output when the model may transform allowed input into disallowed output. Many production systems need both.

​
What it does not replace

Content moderation does not replace Prompt Injection Protection, which deals with authority and control, or Topic Control, which deals with product scope. A request can be in scope and still disallowed, or harmless in wording while still trying to seize control of the model. Read Moderation Evasion when repeated abuse shows up in different wording.

​
Next steps

Next, read Moderation Evasion or go to Topic Control.
Last modified on March 19, 2026