SemiAnalysis: Only 3.6% of Chinese AI Releases Disclose Safety Results

A new analysis by SemiAnalysis reveals that transparency regarding safety testing remains extremely low among major Chinese artificial intelligence developers. The firm reviewed 857 model releases produced by nine distinct Chinese AI labs over a period spanning from 2021 through September 2026. The findings indicate that only 3.6% of these releases included any safety results disclosed by the developers, with a mere 1.1% providing such data at the moment of launch.

The data highlights a significant gap in the current AI safety landscape, particularly regarding how organizations communicate risk and mitigation efforts to the public and technical community. According to SemiAnalysis, the overwhelming majority of models were released without accompanying safety documentation or test results. This lack of disclosure makes it difficult for external researchers, regulators, and enterprise adopters to assess the potential risks associated with deploying these systems. The analysis covers a five-and-a-half-year period, suggesting that this trend has persisted rather than improved as the technology has matured.

Context and Industry Response

The release of these findings coincides with broader industry debates about the pace of AI development. On September 12, 2026, Dario Amodei, CEO of Anthropic, published an essay arguing that frontier AI laboratories must deliberately slow the rate at which they improve model capabilities. Amodei’s position emphasizes the need for caution and rigorous safety measures before advancing to more powerful systems. While Amodei’s comments focus on the global frontier lab community, the SemiAnalysis report provides specific data points that illustrate the current state of safety reporting in the Chinese AI sector.

The contrast between the call for deliberate slowing and the reality of rapid, opaque releases underscores the tension between competitive pressure and safety compliance. With only a tiny fraction of releases including safety metrics at launch, the industry appears to be prioritizing speed and market entry over transparent risk communication. This environment complicates efforts to establish standardized safety benchmarks or regulatory frameworks that rely on developer-provided data.

Who Is Affected

The lack of safety disclosures primarily impacts organizations and individuals who integrate third-party AI models into their operations. Enterprises relying on these Chinese-developed models for critical functions may face unknown risks due to the absence of verified safety testing data. Security professionals tasked with evaluating the security posture of AI integrations are left without the necessary documentation to perform thorough risk assessments.

Regulators and policy makers are also affected, as the scarcity of safety data hinders the development of evidence-based regulations. Without consistent reporting from developers, it is challenging to determine the prevalence of specific failure modes or safety vulnerabilities across the ecosystem. This opacity forces downstream users to rely on independent testing, which may not be feasible for all organizations due to resource constraints.

What should you do?

Given the low rate of safety disclosures, organizations should adopt a verification-first approach when integrating new AI models. Relying solely on developer claims is insufficient when historical data shows that most releases lack accompanying safety evidence. Teams should prioritize models that provide independent audit results or undergo rigorous internal testing before deployment. Monitoring industry developments and regulatory updates is also essential, as the landscape for AI safety reporting continues to evolve.

What to do and how to stay safe: SemiAnalysis

• Conduct independent security and safety testing of any third-party AI models before integrating them into production environments.
• Review vendor documentation carefully to identify any missing safety metrics or undisclosed limitations in model capabilities.
• Implement robust monitoring and logging for AI outputs to detect anomalous behavior or safety failures in real time.
• Stay informed about emerging regulatory requirements and industry standards for AI safety transparency and reporting.

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How many Chinese AI releases included safety results?

According to SemiAnalysis, only 3.6% of the 857 releases analyzed included safety results from the developers.

When were these AI releases published?

The analyzed releases were published between 2021 and September 2026.

Did Dario Amodei comment on this specific report?

No, Dario Amodei published an essay on September 12, 2026, arguing for slower development, but the source material does not link his comments directly to the SemiAnalysis report.