China can build AI, but it can’t install trust

When DeepSeek’s R1 model arrived in January 2025, it was greeted as China’s leap to the artificial intelligence (AI) frontier. While R1 outperformed its Western competitors in maths and coding, it also refused or deflected roughly 85 per cent of prompts concerning politically sensitive topics in China. This distinction lies at the heart of China’s bid to shape the global technological order.

Through the 2023 Global AI Governance Initiative and 2025 Action Plan for Global AI Governance, Beijing has cast itself as a rule-maker of the digital age, championing cyber sovereignty and pitching cheap, open-source Chinese models to the Global South as ‘AI democratisation’.

Underlying, and often frustrating, these efforts is the fact that the Chinese government’s overriding priority is domestic social stability. In maintaining social stability, the government’s instinct is to treat the free flow of information as a political liability. The 2023 Interim Measures for the Management of Generative AI Services illustrated this priority by requiring any public-facing model to uphold ‘core socialist values’ and pass a state security review.

This prioritisation of social stability is not a problem for all cutting-edge technologies, since not all technologies have a social component. A battery or an electric vehicle is a comparatively ‘asocial’ technology and so carries fewer social risks. China happens to excel in such technology — it builds most of the world’s electric vehicles, batteries and solar panels and leads the world in investment in robotics.

In contrast, generative AI is an inherently social technology. Its value lies in open-ended language. Yet this very capability is exactly what the Chinese government wants to control. This constitutes a stability tax — an unavoidable cost imposed on China’s frontier AI ambitions by the government’s overriding commitment to political control.

But censorship persists even when DeepSeek’s open-weight models are run privately and research shows that foundation models originating from China have higher refusal rates, shorter responses and less accurate answers to a battery of 145 questions about Chinese politics than models built elsewhere.

Chinese state control of the media seeps into the training data on which all models draw and, in doing so, it nudges all large language models towards more favourable portrayals of states with tightly controlled media. The most visible layer of this tax can be removed in the case of DeepSeek, as Perplexity demonstrated when it post-trained a ‘decensored’ version of the R1 model.

None of this has prevented Chinese models from having global users. Alibaba’s Qwen large language models now rank among the most downloaded open-source models on Hugging Face. Developers often adopt Chinese open-source models because they offer comparable performance at a lower cost. In the first half of 2026, Chinese-origin models at times accounted for almost half of all enterprise token volume on OpenRouter. But these gains are concentrated in programming and automated ‘agentic’ workloads. It is to these cost-driven workflows that OpenRouter’s leadership has attributed the rise of the Chinese share.

Still, most large enterprises contract directly with US providers or run their inference through the major US cloud platforms, where the penetration of Chinese models is far shallower. The Chinese models that do reach international users are rarely taken as they come. Prominent developers often modify these models before offering them. Chinese models are employed for their cheap computational labour, but they are rarely trusted to communicate their outputs directly to consumers.

Mainstream consumer audiences outside China continue to reach overwhelmingly for US models. DeepSeek’s consumer app has seen downloads tumble from an estimated 171 million in 2025 to an estimated 35 million in the first half of 2026. Australia, Taiwan and Italy have barred it from use in government systems.

While consumers around the world may happily fill their homes with Chinese robot vacuums and their roads with Chinese electric vehicles, many still treat Chinese AI models with scepticism.

This stability tax is part of the reason why Beijing’s global technology ambitions run into a ceiling of its own construction. China attempts to lay the foundations of the digital order, but it struggles to build the cognitive layer of this technology in a manner that is sufficiently credible and independent to be trusted by international consumers.

China can build capable domestic generative AI models. But a government that regards unconstrained information as a threat will always seek to constrain a technology whose defining power is just that — unconstrained information. The question for China’s AI ambitions is whether it can tolerate the degree of openness necessary to make those models globally trusted.

Hannah Bailey is Assistant Professor at the Carnegie Mellon Institute for Strategy & Technology.

This is an abridged version of a presentation given by Dr Hannah Bailey at the Reimagining Global Order: China’s Order-Building Ideas, Narratives and Practices roundtable held at The Australian National University on 14 May 2026.