Researchers linked to Chinese People’s Liberation Army (PLA) have used results generated by AI models developed by OpenAI y Anthropic to train proprietary systems intended for defense applications. The information comes from a Reuters investigation based on the analysis of dozens of academic articles and patents that document the use of these technologies in military projects.
AI models accelerate the development of military systems
The technique identified in the documents is called distillation of models this method allows the use of responses from a high-capacity model to train smaller, more specialized versions. This approach significantly reduces the computational resources needed to develop systems capable of running on local infrastructures.
Furthermore, the research indicates that this practice appears repeatedly in work carried out by institutions related to the Chinese military and security sectors. The researchers believe that US models represent a source of technical knowledge that can shorten the development time for new capabilities artificial intelligence.
Academic papers show applications in defense
The analysis reviewed more than 80 academic publications and patent records, as well as research compiled by the Jamestown Foundation. Among the cases identified was a study by researchers from PLA Unit 96941 describing the use of GPT-3.5 to process source code related to military applications.
According to Sunny Cheung, a researcher at the Jamestown Foundation, some scientific teams are seeking to transfer the reasoning processes of Western models to national systems geared towards surveillance, cyber warfare, and support for tactical decision-making.
Cheung explained that training a model to reproduce the reasoning behind a response is a much greater challenge than simply teaching it the final result. According to his analysis, several projects are precisely aiming to capture this ability and incorporate it into models that can be deployed under local control.
Intellectual property and technological controls under dispute
Model distillation is a practice used in various areas of the artificial intelligence industry, however, the international debate focuses on cases where results obtained from commercial models could be used without authorization to reproduce some of their capabilities.
US officials argue that certain Chinese entities use this strategy to reduce the impact of restrictions on access to advanced chips and other strategic technologies, as well as to raise potential implications for intellectual property rights.
For its part, China rejects these accusations and maintains that USA it tries to maintain a dominant image in the development of artificial intelligence. Companies from the Asian country have also argued that their technological advances are the result of their own developments. Among them, the startup Moonshot recently denied that its Kimi K3 model was built by distilling foreign systems.
AI governance is gaining traction on the international agenda
The revelations come as the United States and China hold talks on the governance and security of artificial intelligence. The development of increasingly sophisticated models for civilian and military applications has made AI a major area of technological competition between the two powers.
The evolution of techniques such as pattern distillation adds a new component to that scenario by combining issues of invention, national security, intellectual property, and control over technologies considered strategic.
Source: Reuters
Photo: Shutterstock
