Study Finds Beijing Embeds Political Standards in AI Models Exported Worldwide

How much reach do Chinese AI models have in the global market?

According to a Central News Agency (CNA) report dated July 14, 2026, a new study cites data from OpenRouter showing that, as of May 2026, seven of the world's ten most popular large language models (LLMs) are Chinese-developed. Measured by token usage — the metric AI systems use to process text — these seven models account for roughly two-thirds of global token consumption, according to the report [E9].

This distribution means that when global users interact with LLMs, the majority of the underlying processing volume runs through models built in China, based on the OpenRouter figures cited in the study.

How does China embed political control into AI through technical standards?

The study, titled "Authoritarian Innovation" and published on July 13, 2026 by the Resilience Innovation Lab (RIL) — a group founded by researchers and human rights advocates — analyzed Chinese government policy documents and corporate announcements published between 2013 and 2026, according to CNA [E7].

The research found that Beijing has built a set of technical standards spanning five stages of AI development: training data, data annotation, model behaviour, application testing, and content traceability. According to the report, these standards convert political requirements into concrete technical specifications, ensuring that content controls are completed during the model development stage itself, rather than applied afterward [E10].

How does the research expose China's use of AI in cross-border opinion manipulation?

The RIL study, as reported by CNA, examined China's AI governance from the angles of national security, information control, and political economy, showing how political demands are converted into technical standards and built into the entire generative AI development pipeline [E7].

As supporting evidence, the research cited internal documents leaked in 2025 from the Chinese firm GoLaxy (中科天璣), which reportedly showed the company had used generative AI to influence public opinion in Hong Kong, Taiwan, and the United States. Intelligence analysts quoted in the study described generative AI as "a new weapon for information warfare," according to CNA [E8].

What is Taiwan's policy response to AI development and China's controls?

Separately, Taiwan's National Science and Technology Council (NSTC/國科會) has been developing its own generative AI system. NSTC chairman Wu Cheng-wen (吳誠文) told the Legislative Yuan's Education and Culture Committee that the council has released an 8-billion-parameter (8B) version of its "Trustworthy AI Dialogue Engine" (TAIDE) for industry use across different sectors [E1].

NSTC further stated it will direct the National Center for High-performance Computing to work with industry to develop themed application services built on TAIDE by the end of this year, aiming to let more members of the public experience the benefits of the research, according to the council's own account of the legislative session [E2].

On export controls, the same legislative report noted that Taiwan currently lists six countries — Iran, Iraq, North Korea, Sudan, Syria, and China — as export-controlled regions under the Foreign Trade Act, while the United States is not included in that control list [E5].

Key figures at a glance

What this means

The evidence gathered by CNA and RIL shows two parallel tracks moving at the same time. On one side, Chinese-made LLMs already carry about two-thirds of global token traffic among the world's ten most-used models, while Beijing has formalized political requirements into technical standards covering training data, annotation, model behaviour, testing, and traceability — meaning content controls are built in before a model ever reaches an end user [E9][E10]. Leaked GoLaxy documents cited in the same study add that generative AI tools have already been applied to shape opinion in Hong Kong, Taiwan, and the US, which intelligence analysts described as a new form of information warfare [E8]. On the other side, Taiwan's own policy response documented in the same legislative session centers on building a domestic alternative (the 8B-parameter TAIDE model, with application services targeted by year-end) while maintaining export controls that list China among six restricted destinations [E1][E2][E5]. Whether these two tracks — global Chinese LLM reach and Taiwan's domestic AI and export-control measures — directly intersect is not addressed in the available reporting.