Stability AI’s Annual Integrity Transparency Report
02:00 · September 17, 2025 · Stability AI News
At Stability AI, we are committed to building and deploying generative AI responsibly, and we believe that transparency is foundational to safe and ethical AI.
Summary
Stability AI has published its Annual Integrity Transparency Report covering April 2024 to April 2025, setting out how the company applies safety-by-design principles across its image, video, 3D and audio models, including those offered through its API. The document describes measures taken at each stage of the development and deployment lifecycle to reduce the risk of harmful outputs, with particular emphasis on preventing the generation or distribution of child sexual exploitation material.
Training data is assembled from publicly available internet sources and licensed third-party datasets; no material is drawn from the dark web or adult sites, and paywalled content is excluded. In-house and open-source NSFW classifiers, together with hash lists supplied by Thorn and the Internet Watch Foundation, are applied to remove prohibited material. During the reporting period no instances of CSAM or CSEM were identified in the datasets used for current models.
Before release, every model undergoes structured red-teaming exercises conducted by internal and external experts. These evaluations use prompts depicting adult nudity and sexual activity as proxies to test for CSAM or CSEM generation capabilities. All models released in the period were subjected to this process, and none were found to produce such content. Where risks are identified, additional fine-tuning or safety LoRAs are applied. At the API level, real-time prompt filters, NSFW image classifiers and CSAM hash matching block violative inputs and outputs before generation occurs.
Content provenance is addressed through C2PA metadata embedded in all media produced via the API. The metadata records the model name and version and is cryptographically signed with a Stability AI certificate. Openly released models do not yet carry this metadata, an area the company states requires further development. Automated detection is supplemented by human review of flagged content, and any confirmed CSAM is reported to the National Center for Missing and Exploited Children; thirteen such reports were filed during the period. Users must be at least eighteen years old and accept the Acceptable Use Policy before accessing the technology.
The company reports partnerships with Thorn, the Internet Watch Foundation and the Tech Coalition’s Pathways program, and states that it continues to monitor regulatory developments and refine its risk-management processes.
Why it matters
This report is highly relevant for Dutch product teams and builders as it details the safety mechanisms, API filters, and C2PA provenance standards implemented in Stability AI models. Understanding these safeguards is crucial for building compliant, ethical AI applications that align with stringent EU and Dutch regulations.
