The Pulse
Chinese AI Labs Published Safety Results for 3.6% of Releases
SemiAnalysis found model-specific safety results for 31 of 857 releases from nine leading Chinese AI developers. Only nine releases had results available at launch, while the report says missing public results do not establish that companie
AI.info Team · Written with AI ·
China’s latest AI safety framework warns that models could deceive evaluators, conceal capabilities and bypass safeguards. Yet a SemiAnalysis review of nine leading Chinese developers found a public, model-specific safety result for just 31 of 857 releases—3.6%—and a result available by launch for only nine. The report, published October 8, measures what companies disclosed, not whether they conducted private tests.
Framework 3.0 warns of risks; release records show few results
The review covers releases from 2021 through September 15, 2026, by Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Zhipu (Z.ai), MiniMax and StepFun. Its dataset includes 741 product models and 116 research models. The authors checked developers’ model cards, release notes and technical reports for results they could connect to a named release.
SemiAnalysis counted quantitative or substantive findings on harmful output, jailbreaks, toxicity, privacy, refusal behavior or dangerous capabilities. Broad claims that a model was “safety-trained” or “evaluated” did not qualify; nor did a test result for one model automatically count for another size or version. The authors also caution that “not found” applies only to the materials they checked and does not mean a model was not tested.
Nine releases had results by launch
Of the 31 releases with a published result, nine had one available at or before release, equal to 1.1% of the full dataset. Another 16 results appeared afterward, with a median delay of 42 days and a maximum of 349 days. For six releases, the researchers found results but could not establish when they appeared or match them confidently to the specific model.
The other 826 releases divide into three groups in the report: 10 with a claim of evaluation but no figures; three whose evaluation was known only through press or investor accounts, without developer documentation the researchers could retrieve; and 813 with no safety disclosure in the checked materials. The last group represents 94.9% of all releases in the dataset.
Startups disclosed more often than China’s tech giants
The startups published results for 20 of 317 releases, or 6.3%, compared with 11 of 540, or 2%, for the four larger technology companies. The gap does not make disclosure routine: SemiAnalysis found no company did so consistently. Alibaba had results for seven of its 238 releases; Tencent had one of 133, ByteDance two of 120 and Baidu one of 49.
Zhipu was the only developer in the sample with a result in every year from 2022 onward, the report says. Across the dataset, releases with any result never exceeded seven in a quarter, even as the review records a rise from three releases in the first quarter of 2023 to 101 in the third quarter of 2025. SemiAnalysis also found 93% of reasoning models had no published safety result.
Disclosure is not the same as testing
The report sets its release census against China’s AI Safety Governance Framework 3.0, issued by the standards committee TC260 under the Cyberspace Administration of China on September 14. The framework identifies model risks including unauthorized access to system resources, deception of evaluators and hidden capabilities. SemiAnalysis argues that China’s binding rules mainly govern AI applications and their effects on users, rather than requiring frontier-model developers to publish risk assessments tied to model capabilities.
“However, we think China’s real approach to AI safety is speed-based, not safety-based,” the authors write. That is their interpretation of the policy and release evidence, not a finding that the unreported models went untested. The report provides no comparable count for U.S. developers, so its 3.6% figure describes the nine Chinese companies and the materials SemiAnalysis examined—not a direct comparison between countries.