Moderation tools are the components that check text going into and coming out of an AI application against a policy. The goal is not to make the model smarter, it is to catch content that falls outside the rules of the product. The check runs both ways, over the prompt a user writes and over the response the model returns.
Most providers offer this as a separate classification service. You send text, the service returns scores for categories such as hate, harassment, self-harm, violence and sexual content. The application reads those scores and blocks the request, hides the response, or routes the item to human review. Rule-based filters, blocklists and rate limits usually sit alongside it.
The tools matter most on surfaces open to the public. A general chat assistant, a flow that publishes user-facing copy, a service summarising community content, or anything aimed at children all belong in that group.
A worked example: an assistant that summarises new forum posts quotes an insult verbatim in its output. Output moderation flags that response, and the summary gets regenerated with the offending sentence removed.
It helps to be realistic about the limits. Classifiers do not always read context correctly and can push medical or legal text into the wrong bucket. Thresholds get tuned per product, and an appeal path stays open.