Stability AI Publishes Transparency Report Detailing AI Safety Practices

Stability AI has released a transparency report covering April 2024 through April 2025, laying out in detail how it develops, tests, and monitors its generative models for video, image, 3D, and audio.

The report describes a multi-layered safety stack. Before training, the company filters datasets using in-house NSFW classifiers and CSAM hashlists from organizations like Thorn and the Internet Watch Foundation. At inference time, prompt filters and image classifiers block policy-violating inputs and outputs on the platform API. The company says it also uses fine-tuning and safety LoRAs at the model level, informed by structured red teaming exercises. Notably, Stability AI collaborated with the UK's Online CSEA Covert Intelligence Team to red-team Stable Diffusion 3 prior to release and reported that no CSAM was able to be generated.

On provenance, Stability AI implements the C2PA standard across its API, tagging generated images, video, and audio with cryptographic metadata including model name and version. However, the company acknowledges this has not yet been extended to its openly released models, calling it "an area that requires further work." It also noted that some alternative watermarking approaches it explored degraded image quality.

The company reports all detected CSAM to the National Center for Missing and Exploited Children and says no users submitted CSAM-related reports during the period. All Integrity team staff receive training on CSAM identification and reporting obligations. Stability AI also highlighted partnerships with Thorn, All Tech Is Human, the Internet Watch Foundation, and the Tech Coalition's Pathways program, all announced in 2024.

For builders and API customers, the report signals that Stability AI is investing heavily in safety tooling and regulatory positioning, though the gap between API-level protections (C2PA tagging, automated filtering) and open-model safeguards remains a live tension the company has yet to fully resolve.