Chinese Researchers Adapt Meta’s AI Model for Military Advancements

Top Chinese research institutions, including those linked to the People Liberation Army (PLA), have reportedly utilized Meta’s open-source Llama model to develop an artificial intelligence (AI) tool with military potential.

This development, first detailed in an academic paper reviewed by Reuters, underscores efforts to harness publicly available Western technologies for strategic purposes.

The June publication highlights work by six Chinese researchers from three institutions, including two associated with the PLA’s Academy of Military Science (AMS).

These researchers adapted an early version of Meta’s Llama model, specifically the Llama 2 13B large language model (LLM), integrating customized parameters to create an AI tool called “ChatBIT.”

According to the paper, ChatBIT was optimized for intelligence gathering and decision-support in military contexts, demonstrating superior performance compared to some models approaching 90% of OpenAI’s GPT-4 capabilities.

However, specifics about performance metrics or operational deployment were not disclosed.

Meta’s Stance on AI Restrictions

Meta’s AI models, including Llama, are released with usage restrictions, prohibiting applications related to military operations, nuclear activities, espionage, and weapon development.

Despite these safeguards, the open nature of these models limits Meta’s enforcement capabilities.

Molly Montgomery Meta’s director of public policy, stated that any PLA utilization of their models contravenes the company’s acceptable use policy.

The company emphasizes the need for the U.S. to champion open-source innovation while acknowledging its potential vulnerabilities.

Meta’s response to the Reuters investigation reinforced this perspective, noting that focusing solely on an outdated American model was less relevant given China’s extensive investments in AI, estimated at over a trillion dollars aimed at surpassing U.S. technological leadership.

Scope of Chinese AI Endeavors

The research was conducted by Geng Guotong and Li Weiwei of the AMS’s Military Science Information Research Center and the National Innovation Institute of Defense Technology, alongside experts from the Beijing Institute of Technology and Minzu University.

The team acknowledged the limited training data—only 100,000 military dialogues—compared to larger LLMs trained with vast datasets.

This gap prompted skepticism about the tool’s comprehensive capabilities.

Meta’s Joelle Pineau, a VP of AI Research and professor at McGill University, expressed doubts regarding ChatBIT’s effectiveness given the comparatively small scale of its training data.

Strategic Implications and U.S. Concerns

The findings emerge amidst ongoing U.S. deliberations over balancing open innovation with national security.

President Joe Biden’s 2023 executive order highlighted AI’s dual benefits and potential risks, including removing built-in safety measures.

Meanwhile, Washington has been crafting regulations to restrict U.S. investments in Chinese technology sectors that may pose security threats.

A Pentagon spokesperson noted that while open-source models offer strategic advantages, they also present risks that demand vigilant monitoring of competitors’ capabilities.

China’s Broader AI Utilization

China’s advances in AI have extended beyond military to domestic security and law enforcement.

Another paper reviewed by Reuters detailed Llama 2’s role in training electronic warfare strategies at the Aviation Industry Corporation of China (AVIC), identified by the U.S. as a PLA-affiliated firm.

Additionally, Llama has been employed in “intelligence policing” to process data for improved decision-making by authorities.

The PLA Daily’s April commentary underscored AI’s potential to accelerate the development of military technologies, enhance combat simulations, and improve training protocols.

This broader trend has raised questions about the feasibility of restricting Chinese access to Western AI advancements.

William Hannas, a lead analyst at Georgetown University’s Center for Security and Emerging Technology, highlighted the extensive collaboration between Chinese and American researchers as a barrier to isolation.

There’s too much interconnection between top scientists from both nations to fully restrict participation, Hannas remarked.

His analysis aligns with CSET’s 2023 report that identified 370 Chinese institutions with publications on General Artificial Intelligence, reinforcing Beijing’s ambition to lead globally in AI by 2030.