PRC State-centric AI Governance Weakens Protection of Rights

PRC State-centric AI Governance Weakens Protection of Rights

Executive Summary:

• The Chinese Communist Party’s information control is embedded in Chinese AI models. As these systems are adopted globally, their built-in controls can be replicated at scale, entrenching Beijing-favored narratives and reasoning patterns in diverse information environments.
• The PRC promotes a global AI governance framework that foregrounds sovereignty, state control, and state-to-state cooperation, while giving limited attention to the role of nongovernmental actors and offering no commitments to rights such as access to information and freedom of expression. Wider adoption of this framework would weaken global protection of rights most impacted by AI.
• The PRC’s global AI governance approach should be read as part of the CCP’s broader effort to shape international norms in ways that strengthen its grip on power at home and expand authoritarianism globally.

In February, U.S. artificial intelligence (AI) firm Anthropic alleged that three Chinese AI companies, DeepSeek, Moonshot AI, and MiniMax, had used large numbers of fraudulent accounts to carry out “industrial-scale” distillation attacks on its AI models, extracting outputs at volume to accelerate their own models’ development (New York Times, February 23). Meanwhile, amid growing adoption of Chinese open-weight models in parts of the developing world, Washington has announced plans to send “Tech Corps” volunteers abroad to promote American AI models and support local adoption (Rest of World, February 20).

These developments reflect intensifying competition between the United States and the People’s Republic of China (PRC) around AI, not just over capabilities and global model adoption, but also over what information controls and modes of governance will travel with these systems (China Brief, March 16, 2025, April 25, 2025). In the international arena, the PRC has consistently advanced a state-centric concept of governance that constrains individual rights. In the AI domain, the same conceptual emphases and omissions recur across PRC policy documents and intergovernmental cooperation agreements. Wider adoption of Chinese AI systems and Chinese AI governance model could entrench PRC government narratives and strengthen authoritarian governance practices internationally, and narrow the space to challenge them.

Information Control in Chinese AI and Its Global Export

Chinese AI models are gaining uptake in many parts of the world (Xinhua, February 14). In February, Chinese models accounted for over half of global token use—the first time this has been the case for a single month (21st Century Business Herald, March 2). Their appeal is based in part on cost, as they are comparatively inexpensive to deploy. This is an advantage in environments with compute and budget constraints. The most successful ones are open-weight models, which also lower barriers for local developers to fine-tune and deploy systems domestically without having to rely on foreign APIs. Broader digital cooperation initiatives under the PRC’s Digital Silk Road initiative also facilitate uptake of these models. Chinese AI exports often come bundled as part of other deals. Telecommunications infrastructure built by Chinese companies—cloud services, data centers, 5G networks—provides a ready-made ecosystem for deploying Chinese AI systems.

Domestically, the PRC’s AI governance model treats information control—referred to in the CCP’s lexicon as “information guidance” (舆论导向)—as a core design requirement (China Media Project [CMP], February 9). The “Interim Measures for the Management of Generative AI Services (2023),” the country’s primary regulation for generative AI, requires providers and users to “adhere to core socialist values” (社会主义核心价值观) and to prevent the generation of content that, among other things, “incites subversion of state power” (煽动颠覆国家政权), endangers national security and interests, damages the national image, or undermines national unity and social stability (Cyberspace Administration of China [CAC], July 13, 2023) In practice, this means that data the CCP deems politically sensitive must be excluded from training pipelines, and that models must be engineered with robust output filtering to block prohibited content. This architecture is reinforced by broader national security and information control statutes, including the 2015 National Security Law and the 2017 Cybersecurity Law, which criminalize speech the state perceives could threaten the Party’s monopoly on power (Baidu Baike/中华人民共和国国家安全法, accessed February 28, 2025; National People’s Congress, December 29, 2025). [1]

A growing body of empirical testing suggests that leading Chinese large language models (LLMs) apply systemic restrictions and narrative steering on topics that the CCP views as politically sensitive. These models have been found to refuse answers, provide evasive responses, or reproduce state-aligned framings on issues such as Tiananmen, Xinjiang, Taiwan, and President Xi Jinping (NewsGuard, July 25, 2025).The extent of control moves beyond censoring information to shaping the contours of permitted knowledge, delimiting what is framed as controversial and what is presented as fact.

Three recent European assessments of Chinese AI models have found that leading models embed content controls that overreach the PRC’s domestic political sensitivities, for instance by providing distorted information in relation to Russia’s invasion of Ukraine (CEPA, February 27). A report by the China Media Project, meanwhile, found that a local-language chatbot in Uganda, built by fine-tuning Alibaba’s Qwen-3 model, not only echoes Beijing’s preferred narratives about the PRC, but also softens criticism of Uganda’s own government (CMP, December 17, 2025).

When Chinese models are exported, their embedded information controls can be applied at scale. Even if a local developer fine-tunes a model for local use, the baseline “political security” configuration can remain intact. A separate China Media Project study shows that when a DeepSeek model was deployed internationally, some firms attempted to remove pro-CCP bias, sometimes unsuccessfully, while others did not bother (CMP, March 4, 2025).

Global AI Governance Proposals Foreground State-centric Approach

Internationally, the RPC promotes an AI governance framework that foregrounds sovereignty, state control, and state-to-state cooperation. This approach, which echoes Beijing’s views in other domains, puts development and security at its core while giving limited attention to the role of nongovernmental actors and offering sparse commitments to protecting individual rights.

To date, the PRC has promulgated several high-level proposals on global AI governance. The most notable is the “Global AI Governance Initiative” (全球人工智能治理倡议). Introduced by President Xi Jinping in his speech at the 2023 Opening Ceremony of the Third Belt and Road Forum for International Cooperation, it calls for “respecting other countries’ sovereignty, strictly abiding by their laws, and accepting their legal jurisdiction” (应尊重他国主权，严格遵守他国法律，接受他国法律管辖) (MFA, October 20, 2023). Beijing’s Global AI Governance Action Plan, a 13-point framework released in 2025, calls for a UN-centered digital governance system based on “respecting national sovereignty” (尊重国家主权), and emphasizes support for developing countries to pursue AI “in line with their national conditions” (结合自身国情) (Xinhua, July 29, 2025). Although the action plan urges “actively safeguarding personal privacy and data security” (积极维护个人隐私和数据安全), it does not address rights central to information governance, such as access to information and freedom of expression.

This state-centric approach is also evident in bilateral and multilateral documents on AI cooperation. Some of these do not just promote Chinese models but criticize Western models as proprietary and exclusionary, arguing that technology monopolies and restrictive supply chains entrench unequal access. This allows Beijing to promote “open” cooperation—especially among Global South partners—as both a development strategy and a geopolitical counterweight to perceived Western dominance. Table 1 below provides several illustrative examples of this language taken from various initiatives, plans, communiqués, and other official statements.

Figure 1: Beijing’s State-Centric Discourse in AI Policy Plans

Rhetoric and Practice Diverge

A glaring problem in the PRC’s vision for global AI governance is the gap between its rhetoric and its practice. Beijing regularly presents itself as a champion of “inclusive” and “fair” AI and warns against monopolies and exclusionary standards. Yet when international efforts are convened outside of its preferred frameworks, it has often been reluctant to endorse the resulting commitments. For instance, in September 2024, the PRC did not sign on to the Council of Europe’s Framework Convention on AI, which aims to ensure that AI activities are consistent with human rights, democracy, and the rule of law. That same month, it also refused to endorse a blueprint for ethical, human-centric military AI use that was backed by dozens of countries at the Responsible AI in the Military Domain (REAIM) summit in Seoul in September 2024. At a subsequent REAIM summit held in Spain in early 2026, both the PRC and the United States opted out of a joint declaration (Reuters, February 5).

The contrast is even sharper when set against the PRC government’s own conduct at home. The BRICS Leaders’ Joint Statement on Global AI Governance, released in July 2025, called for respect for countries’ linguistic, cultural, and ethnic diversity, and stressed the need to mitigate discrimination and bias in AI systems (BRICS Information Centre, July 6, 2025). A joint statement with France on AI and global governance similarly says that both countries “hold that AI must provide inclusive access for all, make content available in ways that … respect linguistic plurality and cultural diversity” (MFA, May 7, 2024). But the PRC’s extensive use of data-driven and algorithmic systems to police and control minority communities—and to shape public narratives about the abuses—runs contrary to these stated principles. An Australia Strategic Policy Institute report finds that the country is using minority‑language LLMs to deepen surveillance and control of ethnic minorities, both within the PRC and abroad (Australian Strategic Policy Institute, December 1, 2025). This comes as the National People’s Congress in early March deliberates a law on “Promoting Ethnic Unity and Progress” (民族团结进步促进法), which seeks to erase minority language rights (Human Rights Watch, September 27, 2025).

Alignment with broader International Governance Strategy

Beijing’s rhetoric may diverge from its practice, but its platitude-filled policy documents are not harmless. They are designed as part of a sustained, multipronged effort to normalize a state-centric model of governance. The emphasis on sovereignty, security, and development across its AI documents aligns with Beijing’s broader vision for a new international order, which are embodied in Xi’s four signature global initiatives, covering development, security, civilization, and governance (Qiushi, October 15, 2025; China’s Diplomacy in the New Era, March 4).

The PRC approach to global AI governance also closely tracks the aim it has advanced for global Internet governance. Its digital diplomacy stresses that states should respect other states’ sovereignty and jurisdiction over data and that companies should abide by the laws of the country where they operate. These clauses normalize the view that cross-border digital services are governed first by state legal authority, not by international human rights standards or multi-stakeholder accountability (U.S.–China Economic and Security Review Commission, March 13, 2025).

It is also consistent with patterns documented in its engagement with the international human rights system: privileging state authority, limiting languages that strengthen civil and political rights protections, and narrowing the space for non-state actors, especially civil society (Brookings Institution, September 14, 2020; Journal of Democracy, July 2021; China Brief, February 6). [2] It has consistently advanced these objectives at the UN and other international fora by coordinating with “Global South” states and other authoritarian governments.

Conclusion

Beijing is systematically driving global adoption of Chinese AI models. These models are designed with information controls that ensure responses are titled in its favor. In parallel, the PRC is advancing an AI governance agenda that centers the state, marginalizes other stakeholders, and de-emphasizes rights such as access to information and freedom of expression. This posture is consistent with the country’s broader efforts to reshape international norms in ways that strengthen its rule domestically and expand its influence globally. The long-term implication is a global information environment heavily shaped by Beijing, alongside AI governance norms that legitimize expansive governmental discretion over the protection of individual rights.

Notes

[1] An amendment to the Cybersecurity Law that came into force on January 1 contains, among other things, higher penalties for violating cybersecurity.

[2] Rana Siu Inboden, “China at the UN: Choking Civil Society,” Journal of Democracy 32, no. 3 (July 2021): 124–35.