Stability AI Annual Integrity Transparency Report

Stability AI released its Annual Integrity Transparency Report, built on Safety-by-Design principles, to responsibly build and deploy generative AI. The report aims to build trust with users, developers, and policymakers by transparently disclosing the design, testing, monitoring, and misuse-prevention framework of its models.

Training data draws on publicly available internet data, third-party partner data, and Synthetic Data generated by researchers. Sources where harmful content circulates, such as the dark web or adult websites, are excluded, and data is filtered through in-house-developed and open-source NSFW classifiers. To date, CSAM (Child Sexual Abuse Material) has been confirmed at 0% within the training datasets.

The following multi-layered mitigation measures are applied to ensure the safety of models and platforms:

• API level: Real-time content filters and classifiers detect policy-violating inputs and outputs, and a CSAM hashing system is integrated to block and report harmful material.
• Model level: Models are deployed using Fine-tuning and Safety LoRA techniques based on insights gained from Red Teaming.

Red Teaming is a key process for identifying serious risks in models. It involves both internal and external experts, and includes ongoing risk assessments such as collaborating with OCCIT, a UK law enforcement agency, to verify the safety of the Stable Diffusion 3 model.