Chinese AI Firms Show Only 3.6% Transparency in Safety Testing, Raising Risks

Against the backdrop of growing global emphasis on the safety and ethics of artificial intelligence (AI) systems, a recent report from the technology research firm SemiAnalysis indicates that only 3.6% of the publicly released models from major Chinese AI developers have disclosed specific safety test results. This figure not only highlights a transparency gap among Chinese AI companies but has also raised concerns about their risk management capabilities.

SemiAnalysis’s investigation covered nearly 100 AI companies based in China or with their primary research and development hubs in the country, reviewing the safety test reports for their publicly released models. The report notes that while most companies mention in technical white papers or press releases that certain forms of testing have been completed, only a tiny fraction actually publish specific test data, methodologies, or results. This low disclosure rate stands in stark contrast to the practices of international firms such as OpenAI in the United States and DeepMind in the United Kingdom, which openly publish their safety test reports.

Safety testing typically encompasses multiple dimensions, including adversarial robustness, bias detection, data privacy protection, and interpretability assessment. These tests help developers identify and correct potential risks in advance and provide users with a reference for evaluating model reliability. Without transparent and public test reports, users find it difficult to assess the safety risks of models in real-world applications, which may lead to the misuse or inappropriate use of the technology, potentially causing unforeseen negative impacts on society.

Several factors contribute to the low disclosure rate among Chinese AI enterprises. First, the domestic regulatory environment is relatively conservative, leading many companies to keep test results confidential to avoid touching upon policy-sensitive areas. Second, commercial competitive pressure leads companies to view test data as a core competitive advantage, fearing that public disclosure could weaken their market position. Furthermore, the lack of unified national-level safety testing standards and review mechanisms has resulted in a lack of consensus among companies regarding testing methodologies and report formats, making it difficult to produce standardized documents for external review.

Globally, AI safety and transparency have become issues of joint concern for policymakers and the industry. The draft European Union Artificial Intelligence Act explicitly requires AI developers to conduct rigorous risk assessments and testing for high-risk systems and to publish test reports. The United States is also promoting AI ethics regulations in multiple states and encouraging companies to adopt open-source safety tools. In comparison, progress in this area in China appears slower, which may have long-term implications for its reputation and cooperation opportunities in global AI competition.

In response to this challenge, the industry and regulatory bodies need to jointly develop clearer and more actionable safety testing standards and encourage companies to disclose key testing methodologies and results while protecting trade secrets. Only with dual guarantees of transparency and safety can the Chinese AI industry maintain its competitiveness on the global technology stage while delivering safer and more reliable AI applications to society. (Source: CNA)