Chinese AI Developers Shared Safety Tests for Just 3.6% of Model Releases, Report Finds
According to a SemiAnalysis report, leading Chinese artificial intelligence companies have publicly disclosed safety testing results for only a tiny fraction of the models they released between 2021 and September 2026.
Leading Chinese artificial intelligence developers publicly shared safety test results for only 3.6% of the model versions they launched between 2021 and September 15, 2026, according to a new report published by research firm SemiAnalysis.
Details of the SemiAnalysis Review
California-based research firm SemiAnalysis examined 857 models released by nine leading Chinese artificial intelligence companies between 2021 and September 15, 2026. These companies include Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax, and StepFun.
Rates of Safety Test Results
The report revealed that out of the 857 model versions examined, only 31—or 3.6%—had publicly available safety evaluation results that could be mapped to a specific model. The number of models with such results at or before launch stood at just 9 versions, representing a 1.1% level.
Confidential Tests and Evaluation Criteria
Noting the possibility that companies may have conducted tests privately, researchers found no safety disclosures for 813 versions. SemiAnalysis defined disclosures as specific results such as harmful output tests, jailbreak resistance, toxicity, privacy, refusal behavior, or hazardous capabilities, while excluding general safety training claims from this scope.
Beijing's Regulations and Legal Framework
The report also stated that Beijing's binding rules primarily regulate applications and user interactions rather than mandating that leading developers conduct or publish risk assessments based on model capabilities.