# Enterprise-Grade Safety and Transparency by Design

The Bria platform is built for zero-risk AI implementation, with enterprise-grade safety, compliance, and content integrity embedded across every layer. Our proprietary foundation models are trained exclusively on licensed, safe-for-commercial-use data—avoiding scraped, harmful, or infringing content by design. While safety starts at the model training stage, Bria’s APIs offer developers built-in controls in-generation and post-generation through a multi-layered architecture. These include prompt and content moderation, flagging of IP-related prompts, and post-generation compliance features—helping teams meet legal, brand, and platform requirements by default.

# Safety Architecture

Bria’s enterprise-grade safety framework spans three layers:

## 1. Pre-Training Layer – Data Integrity

- Models are trained on **100% licensed data**
- **No scraped internet content o**r unauthorized material
- **No public figures, fictional characters, biometric data, NSFW, or violent content**
- Balanced, diverse, and **inclusive dataset representation**


## 2. In-Generation Layer – Real-Time Controls

Bria provides two configurable, opt-in runtime safety features: prompt content moderation and visual content moderation.

### Prompt Moderation

- Enabled via the `prompt_content_moderation` parameter
- Scans textual prompts for NSFW or restricted concepts before generation
- Uses non-AI-based blocklist filtering


#### Handling IP-Related Prompts

Prompts that reference public figures, brands, or other protected content are not blocked, but the models are not trained on this type of data.
As a result, the output may be unrelated or differ significantly from what you intended.
When an IP-related reference is detected in the prompt, the following warning will appear in the API response:


```text
  This prompt may contain intellectual property (IP)-protected content.
  To ensure compliance and safety, certain elements may be omitted or altered.
  As a result, the output may not fully meet your request.
```

**Example:**

- **Prompt:** "a Nike sneaker on a reflective white surface"


**Bria Output:**


**Outputs from Other Providers:**
The following images show how different visual generation providers handled the same prompt.

img
img
img
img
img
### Input & Output Visual Moderation

- Enabled via the `content_moderation` parameter
- Scans both input and output visuals


#### Moderated Categories

- **Explicit Content**: Nudity, sexual activity, sex toys
- **Non-Explicit Nudity**: Implied nudity, kissing
- **Swimwear/Underwear**
- **Violence**: Weapons, blood, self-harm, gore
- **Visually Disturbing Content**: Crashes, corpses, emaciated bodies
- **Substances**: Alcohol, pills, smoking
- **Offensive Gestures**
- **Gambling**
- **Hate Symbols**


## 3. Post-Generation Layer – Data Traceability & Compliance

- **C2PA Image Marking** adds metadata for content authenticity and traceability
- **Attribution Engine Layer** enables revenue-sharing with original data owners and provides transparency to Bria customers into the data  used to train the models


## Indemnity Guarantee

Bria provides **full indemnity against copyright infringement** for all outputs generated by its models, **for enterprise customers only**. This assurance is made possible through our use of 100% licensed, safe-for-commercial-use training data — ensuring that every visual generated with Bria’s platform is compliant by design.