> ## Documentation Index
> Fetch the complete documentation index at: https://anthale.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Anthale Content Moderation on Input and Output Paths

> Understand Anthale content moderation, which categories it covers, and where to enforce it on input and output paths in your application.

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

| Category                       | What it usually covers                                                          | Why teams enable it                                                      |
| ------------------------------ | ------------------------------------------------------------------------------- | ------------------------------------------------------------------------ |
| Hate and harassment            | Abuse, threats, or demeaning language aimed at protected groups or individuals  | Public-facing assistants usually need clear boundaries here              |
| Sexual content                 | Explicit sexual content, solicitation, and adult material                       | Many products need strong output limits even when input is user-supplied |
| Violence and self-harm         | Graphic violence, threats, instructions for harm, or self-harm encouragement    | These categories carry direct safety and trust risk                      |
| Crime and illegal activity     | Advice that facilitates fraud, theft, weapon creation, evasion, or other crimes | The risk is not only reputational. It can become operational and legal   |
| Profanity or brand sensitivity | Coarse language or content your product does not want to surface                | This is usually product-specific rather than universal                   |

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](/docs/learn/guardrails/prompt-injection), which deals with authority and control, or [Topic Control](/docs/learn/guardrails/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.

## Related pages

Read [Moderation Evasion](/docs/learn/guardrails/content-moderation/moderation-evasion) when repeated abuse shows up in different wording.

## Next steps

Next, read [Moderation Evasion](/docs/learn/guardrails/content-moderation/moderation-evasion) or go to [Topic Control](/docs/learn/guardrails/topic-control).