Chinese military researchers adapted Meta's Llama for military AI system
Chinese military researchers published an academic paper describing their adaptation of Meta's publicly available Llama 13B AI model to create a system called Chat B.I.T. for military intelligence and decision-support purposes, reportedly violating Meta's terms of service that prohibit military applications.
What happened
Six Chinese researchers from three institutions, two affiliated with the Academy of Military Science of the People's Liberation Army, published an academic paper in June 2024 describing their use of Meta's Llama 13B model as the foundation for a system called "Chat B.I.T." adapted for military purposes.
The researchers built an AI tool focused on military applications for collecting and processing intelligence information and providing reliable information for operational decision-making. The paper stated that Chat B.I.T. was refined and improved to perform dialogue tasks and answer questions in the military domain, reportedly outperforming some other AI systems including OpenAI's GPT-4 at approximately 90% capability.
Reported impact
Meta stated that any use of its models by the People's Liberation Army is unauthorized and violates its acceptable use policy, which prohibits military, warfare, nuclear, espionage applications and development of weapons.
Meta's terms restrict services with over 700 million users from commercial deployment without a license, and prohibit use in military or warfare applications and other activities subject to US export controls.
Index analysis
Sani Chiong, associate fellow at the Jamestown Foundation specializing in dual-use emerging technology in China, stated: "This is the first time there is strong evidence that Chinese military experts were systematically researching and attempting to leverage the power of open-source large language model programs, specifically those produced by Meta, for military purposes."
Meta acknowledged limited means to enforce its acceptable use policy on open-source models distributed publicly, relying instead on procedural measures to prevent misuse.
Governance controls
Meta's terms of service prohibit use of its models in "military or warfare industries or applications, nuclear, espionage" and other activities subject to US export controls, as well as weapon development and content intended to incite or promote violence.
However, because Meta's models are publicly available, the company has limited technical means to enforce these restrictions on end users.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.