SemiAnalysis finds public safety results for 3.6% of Chinese AI releases

SemiAnalysis found published safety evaluations for just 31 of 857 releases from nine Chinese AI developers, equivalent to 3.6% of its sample, Reuters reported.

The study covers models released from 2021 through September 15, 2026, by Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun. Only nine releases, or 1.1%, had evaluation results available before or at launch.

Researchers found no public safety disclosures for 813 releases. Missing disclosures do not establish that testing never took place, as companies may have assessed their systems internally.

SemiAnalysis counted findings linked to a specific model, including evaluations of harmful responses, resistance to safeguards being bypassed, toxicity, privacy and dangerous capabilities. General assurances that a model had undergone safety training or assessment did not meet that standard.

The analysts also argue that Chinese regulation focuses primarily on applications and their effects on users. In their assessment, recognition of loss-of-control risks has yet to produce mandatory duties for frontier developers triggered by their systems’ capabilities.