Chinese AI developers are extremely reluctant to share information about testing their models for safety of use, research company SemiAnalysis has found. This comes as concerns grow around the world about the risks associated with the rapid development of advanced AI systems.
Image Source: Steve A Johnson/unsplash.com
According to SemiAnalysis’ analysis of 857 models released between 2021 and September 15 this year by nine leading Chinese artificial intelligence companies – Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun – only 31, or 6.3%, published security assessment results that could be compared to a specific model. Only nine releases, or 1.1% of all releases, had such results available at or before the model’s release. The researchers found no safety information for the 813 models, although companies could conduct tests privately.
SemiAnalysis explained that information disclosure refers to the publication of results related to a specific model (including tests for the generation of malicious content, resistance to bypassing security restrictions, toxicity, privacy, behavior when refusing to complete a request, or the presence of dangerous capabilities). This does not include general statements that the model has undergone safety training or testing.
High-profile incidents involving AI agents acting on their own have intensified global debate over whether AI companies should slow down the development of advanced models in order to create more effective controls over them.
Last week, media reports emerged that Chinese AI agents, just like American ones, can deceive users, bypass restrictions and hide failures in tests.
China’s latest AI safety management system can identify risks, including models obtaining system permissions or external resources without permission, deceiving evaluators, hiding capabilities, and bypassing security controls. But, according to SemiAnalysis, it does not impose any obligations on developers related to the capabilities of the models. The rules primarily govern applications and their impact on users, without making any requirement for advanced model developers to conduct or publish risk assessments based on model capabilities.
According to SemiAnalysis, no major Chinese developer has released an advanced LLM with test results for dangerous capabilities made public.
If you notice an error, select it with the mouse and press CTRL+ENTER.
