How Moderation Works

Introduction

Most services describe moderation as something that happens to content after it exists. Ours is mostly designed so that prohibited content does not get made in the first place. This page explains the difference, because it is the part worth understanding.

There are four layers. Three are automatic and one is a person.

1. Messages are classified before the model sees them

When you send a message to a companion, it is assessed first. Our filters read what you have written and decide whether it breaks our Blocked Content Policy.

If it does, the message is refused and never reaches the model. There is no generated reply to moderate, because nothing was generated. This matters: a service that filters output has already produced the thing it did not want to produce.

2. Images cannot be asked for in free text

You do not write a prompt for an image. You build a request by choosing from a fixed list we maintain — the scene, the outfit, the pose, the activity — and the request is assembled from those choices.

The practical effect is that a large class of prohibited imagery cannot be requested at all. Not “is refused”: cannot be expressed. There is no field in which to ask. The search boxes on the generation page filter that list; they are not a way to write your own instruction.

The one piece of free text that reaches generation is the name you give a character you create.

3. Nothing reaches a public surface unreviewed

Every generated image and video carries two separate states: whether it is visible, and whether it has been reviewed. Both must be satisfied before it can appear anywhere other people can see it — a character profile, a public gallery, a preview, an album cover.

Content awaiting review is not hidden from the person who made it; it is hidden from everyone else. A piece of media that has been hidden stays hidden regardless of who generated it or what they paid.

4. People review what the filters flag

Our support team reviews flagged cases daily — the cases our filters have raised, and anything reported to us by a user.

To be precise about what that is and is not: it is a daily human pass over flagged cases. It is not a person reading every conversation, and we are not going to claim otherwise. Conversations with a companion are private between you and the service, subject to the controls described here and to our Privacy Notice.

5. When we get it wrong

Automated classification is not perfect in either direction, and we would rather say so than pretend.

It sometimes refuses a message that was entirely innocent — an unlucky combination of ordinary words. If that happens to you, please tell us at support@bonza.chat rather than trying to rephrase your way past it. A false refusal is a bug in our filter and we want to fix it; deliberately working around moderation is a breach of our Community Guidelines, and we cannot always tell the two apart from the outside.

A refused message is not, by itself, a penalty. It simply does not go through.

6. What happens to accounts

Repeated or serious breaches are recorded and reviewed. Depending on what we find, we may remove content, restrict features, suspend an account or terminate it. Content involving minors results in termination without warning, and is reported where we are required or permitted to do so.

If you believe a decision about your account is wrong, you can appeal by writing to support@bonza.chat. We will look again. Where a breach is confirmed on review, the decision stands. The full position is in our Community Guidelines.

Related

• Safety at Bonza.Chat — the overview
• Blocked Content Policy — what is not permitted
• Community Guidelines — expectations and enforcement
• Content Removal Policy — requesting a takedown