regulation
Chinese Military Uses OpenAI and Claude Models for Training
Chinese military researchers have employed responses from OpenAI and Anthropic's Claude models to train their own AI systems through a process called distillation, raising concerns about the effectiveness of export restrictions.
AS1 News
Analysis of over 80 scientific papers and patents reveals that Chinese military efforts involve using AI responses for training in areas such as military code processing, social media monitoring, target recognition, and navigation. In one instance, GPT-3.5 was used to generate descriptions of programming code, while Claude was employed to create synthetic data. These practices demonstrate that export controls on chips and models do not fully prevent the transfer of AI capabilities to military applications via APIs or synthetic data. Despite restrictions, such methods can facilitate the development of local systems without direct access to the original models.
The use of commercial AI models for military training indicates potential circumvention of export controls, impacting international security and AI regulation policies.