Chinese artificial intelligence models, including DeepSeek, face a fundamental credibility deficit rooted in state censorship, leaving them distrusted by governments and mainstream users worldwide despite their competitive pricing and technical capability, according to an analysis published by the East Asia Forum.

The Censorship Tax on Chinese AI

When DeepSeek's R1 model launched in January 2025, it was widely hailed as China's most significant leap in generative AI. The model outperformed several Western rivals on mathematics and coding benchmarks — yet it refused or deflected approximately 85 per cent of prompts touching on politically sensitive topics within China. That gap between technical prowess and informational openness, the analysis argues, sits at the core of China's bid to shape the global technological order.

The regulatory architecture underpinning this behaviour is explicit. China's 2023 Interim Measures for the Management of Generative AI Services require every public-facing model to uphold 'core socialist values' and clear a state security review before deployment. The Chinese government, the analysis notes, treats the free flow of information as a political liability — a stance that directly constrains what its AI systems can say.

A 'Stability Tax' on Frontier AI

The analysis introduces the concept of a 'stability tax' — an unavoidable cost imposed on China's frontier AI ambitions by the government's overriding commitment to political control. Unlike electric vehicles, batteries, or solar panels — sectors where China leads globally — generative AI is an inherently social technology whose value depends on open-ended, unrestricted language. That is precisely the capability Beijing seeks to curtail.

Notably, the censorship problem persists even when DeepSeek's open-weight models are run on private infrastructure outside China. Research cited in the analysis found that foundation models originating from China exhibit higher refusal rates, shorter responses, and less accurate answers across a battery of 145 questions about Chinese politics compared with models developed elsewhere. Chinese state control of media, the analysis explains, seeps into training data, steering large language models toward more favourable portrayals of tightly controlled states.

Enterprise and Consumer Adoption Trends

The commercial picture reflects these trust concerns. While developers sometimes adopt Chinese open-source models for their cost efficiency and comparable performance, most large enterprises contract directly with US providers or route their inference through major US cloud platforms — where the penetration of Chinese models remains far lower.

Mainstream consumer audiences outside China continue to favour US models overwhelmingly. DeepSeek's consumer application has seen downloads fall sharply, from an estimated 171 million in 2025 to an estimated 35 million in the first half of 2026 — a decline of nearly 80 per cent.

Government Bans and Geopolitical Distrust

Australia, Taiwan, and Italy have already barred DeepSeek from use in government systems, reflecting a broader pattern of institutional caution toward Chinese AI infrastructure. This comes amid growing scrutiny of Chinese technology platforms on data security and ideological alignment grounds across Western and Indo-Pacific democracies.

As the global AI race intensifies, the East Asia Forum analysis suggests that China's self-imposed informational constraints may prove a more durable competitive handicap than any technical gap — one that regulation alone cannot bridge.