Stability AI's first transparency report counted 13 CSAM cases. The real story is what it doesn't say.

Stability AI published its first annual Integrity Transparency Report in July 2025, covering the period from April 2024 to April 2025. The document, released on the company's Child Safety page, discloses the number of CSAM reports sent to NCMEC, the metrics of its red-teaming program, and the safeguards built into its image, video, 3D, and audio models.

The headline figure is 13 reports to NCMEC. The company says multiple reports may have been submitted for the same user when more than one image upload attempt was detected. No CSAM was found in the company's training datasets during the reporting period, nor did any of its generative AI models exhibit CSAM or CSEM generation capabilities after stress testing. The report also notes that no user-submitted CSAM reports reached Stability AI.

A layered safety stack

Stability AI's safety approach rests on three layers: filtering training data, preventing misuse at the model and platform level, and enforcing its Acceptable Use Policy. Training data is drawn from open datasets, publicly available websites, third-party partners, and synthetic data generated by researchers. The company says it uses in-house and open-source NSFW classifiers, plus industry CSAM hashlists from Thorn's Safer and the Internet Watch Foundation, to pre-filter its training data. No CSAM was detected.

At the platform API level, Stability AI runs real-time content filters and classifiers that block policy-violating inputs and outputs, and integrates CSAM hashing from Thorn to detect and block known CSAM. At the model level, the company applies fine-tuning and safety LoRAs informed by structured red teaming. Red teaming covered 100% of the company's generative AI models. The company also collaborated with the Online CSEA Covert Intelligence Team (OCCIT), a UK law enforcement unit, to test its Stable Diffusion 3 model before release, and found no CSAM generation capabilities. The approach mirrors broader industry trends in aligning models with ethical constraints, similar to how uncensored models wrestle with refusal and safety trade-offs.

Provenance gaps

The report acknowledges that content provenance via C2PA metadata is implemented only for API-generated images, video, and audio (the wav. format used for sound effects and instrument riffs, which the company says carries no CSEM risk). Openly released models do not include C2PA tags. The company says it has found challenges with non-C2PA watermarking solutions that degraded image output quality, and continues to explore more effective approaches. This limitation is part of a larger conversation about keeping generative media traceable, as explored in the game engine rule that generative models keep forgetting.

"Content provenance has not been implemented during the content generation process for our openly released models. These are areas that require further work to strengthen provenance and traceability across our systems."

Partnerships and looking ahead

Beyond the transparency report, Stability AI has been expanding its commercial footprint. The company launched its Image Services suite on Amazon Bedrock, bringing tools like Inpaint, Erase, Remove Background, and Style Transfer to AWS infrastructure. It also announced a strategic partnership with Electronic Arts to co-develop generative AI models and workflows for game development, starting with accelerating the creation of PBR materials and pre-visualizing 3D environments from prompts. That kind of enterprise push suggests the company is betting on business-to-business relationships as its main growth vector.

The transparency report itself says the company is monitoring regulatory developments and plans to align its practices with emerging responsible AI frameworks. Whether those frameworks will keep pace with fast-moving generative tools remains an open question, as highlighted in DeepMind's framework on conceptual, epistemic, and operational rigor.