Bridging the Artificial Intelligence Governance Gap: The United States' and China's Divergent Approaches to Governing General-Purpose Artificial Intelligence

Kevin Wei; Oliver Guest

REVIEW 3 major objections 3 minor 19 references

Bridging the Artificial Intelligence Governance Gap: The United States' and China's Divergent Approaches to Governing General-Purpose Artificial Intelligence

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The United States and China govern general-purpose AI along three divergent tracks—regulatory focus, core principles, and international forums—and understanding these gaps is key to safety cooperation.

desk verdict A competent, well-sourced synthesis of US-China GPAI governance; the three-axis framing works, but the proxy assumption and judgment-based axes keep it from being more than a useful orientation document. read the letter →

arxiv 2506.03497 v1 pith:NREROTG7 submitted 2025-06-04 cs.CY

Oliver Guest , Kevin Wei This is my paper

classification cs.CY

keywords AIgovernancegeneral-purposeU.S.-Chinarelationssafetyinternationalcooperationcontentregulationmodelsmultilateralforums

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the United States and China, despite shared warnings about risks from general-purpose AI, govern such systems in three importantly different ways: what domestic regulation targets, which principles guide it, and which international forums each country favors. In the United States, regulation centers on the AI models and systems themselves, including compute-threshold triggers; in China, regulation has historically targeted the content and outputs that models generate. Internationally, China pushes for UN-centered governance and capacity-building, while the United States prefers more-exclusive settings such as the G7 Hiroshima Process. The paper contends that recognizing these divergences—rather than assuming a single shared agenda—is a precondition for designing U.S.-China cooperation on AI safety and security. If correct, the analysis gives policymakers a map of where consensus-building and interoperability efforts are needed most.

What carries the argument

The analytical engine of the paper is a three-part comparative framework applied to policy documents from each country: the regulatory target (models and systems versus content and outputs), the governing principles (development rights, control, and state power), and the implementation forum for international governance (UN-centered versus allied and exclusive). The framework's load-bearing distinction is between an AI model—the algorithm that processes information—and an AI system, which embeds the model in filters, prompts, and interfaces; U.S. regulation is said to focus on the former and Chinese regulation on the latter, even as that gap may be closing. The paper uses official documents, regulations, summit declarations, UN resolutions, and readouts of the U.S.-China intergovernmental dialogue as evidence for where each country sits on each axis.

What would settle it

A concrete test would be to track China's next major AI legislation or regulation: if it imposes compute thresholds on model developers or mandates model-level safety evaluations comparable to Executive Order 14110, the paper's claimed divergence in regulatory focus would be significantly weakened. The reverse—a continued emphasis on content and deployment-side obligations—would support it. A second observable test is whether China creates an AI safety institute and seeks membership in the U.S.-led international network, which the paper itself flags as a key indicator.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that U.S. and Chinese approaches to governing general-purpose AI diverge along three axes that policymakers must understand before cooperation on AI safety can advance. The first axis is the focus of domestic regulation: U.S. proposals, from Executive Order 14110 to state bills, target AI models and systems and often tie scrutiny to training-compute thresholds, whereas Chinese regulations have concentrated on model-generated content and on the actors who deploy services. The second axis is governing principles, where the two countries differ over who should be allowed to develop and deploy GPAI, over what 'control' of AI systems means, and over the legitimacy of using AI to bolster state power. The third axis is the international forum: China consistently favors the United Nations as the main channel, while the United States prefers allied and exclusive settings such as the G7 and a network of AI safety institutes. The paper frames these divergences not as fixed facts but as a present-day picture, and it notes signs that China is beginning to move toward model-level regulation.

Load-bearing premise

The paper's comparison depends on treating each country's domestic policy documents and official readouts as reliable indicators of its true governance intentions and international negotiating positions, even though the paper itself notes that domestic policy is not a perfect proxy for international governance.

Editorial extensions

If this is right

Policymakers should design cooperation around the areas of clearest shared interest, such as preventing bioweapon proliferation aided by AI tools, where narrow agreements can build trust for broader efforts.
International governance mechanisms that address both AI models and specific harmful outputs—for example, content that could contribute to bioweapons—would be more interoperable across U.S. and Chinese regulatory cultures.
Key terms such as 'control' and the Chinese term 'anquan' (safety/security) need explicit definition in shared statements, since the same word can carry different regulatory meaning in each country.
The creation of a Chinese AI safety institute and its inclusion in or exclusion from the U.S.-led network would be a leading indicator of whether the two countries can cooperate on AI risk mitigation.
The paper's analysis implies that the U.S.-China Track I dialogue on AI is likely to remain limited by differing delegation priorities—national security concerns on the U.S. side and bilateral relationship management on the Chinese side—until these structural divergences are addressed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the paper's observation that China is shifting toward model-level regulation is correct, the regulatory gap may narrow, but the more consequential divergence could shift to questions of model access and open-weight distribution, where U.S. export controls and Chinese open-source advocacy collide.
The paper's domestic-policy-as-proxy method could be stress-tested by comparing Chinese policy texts with enforcement actions, such as the targets of regulatory fines and service suspensions, to see whether content control or model safety drives actual state behavior.
The forum divergence suggests that global AI governance may develop into two parallel tracks—UN-based capacity-building and alliance-based safety standards—making the paper's interoperability recommendation a practical bridge rather than a harmonization ideal.
A testable extension is whether third countries align with the U.S. or Chinese governance model based on their dependence on either AI supply chains or Chinese digital infrastructure; the paper's framework would predict a split.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a competent expert synthesis, not a research result. The three-way comparison—focus of regulation, key principles, international forum—is a sensible organizing device, and the paper is honest about its own limits. It explicitly flags that domestic policy is not a perfect proxy for international governance (p. 2), and it notes China's possible move toward model-based regulation. That is real credit.

The sourcing is strong and the paper is careful not to overclaim. It makes clear that some divergences could shrink: no official compute threshold in China may reflect drafting timing, and the two academic drafts already contain thresholds. The self-citation (Elmgren and Guest) is a side observation about Chinese AISI counterparts, not load-bearing. The citation pattern overall looks healthy.

The soft spots are real but not fatal. The entire comparison leans on reading official Chinese documents as evidence of GPAI safety-governance intentions, but many of those documents serve domestic content-control goals. The paper acknowledges this but does not test it. The absence of an official compute threshold is treated as divergence even though timing could explain it; the paper itself undercuts this by citing academic drafts with thresholds. And the selection of three divergences is an analyst judgment, not derived from a systematic method—another author might pick different axes. These are moderate concerns for an interpretive piece.

Who gets value: policymakers and researchers wanting a compact, well-referenced map of where the US and China stand on GPAI governance. It will not change anyone's mind, but it is a reliable orientation document for people new to the topic.

Recommendation: if you are deciding whether to referee, yes. A serious reviewer with China policy expertise could sharpen the proxy discussion and check the document interpretations. This is a perspective piece for a policy venue, not a research claim for a technical journal. Send it to review with the expectation that revisions should make the proxy limitation more central.

Referee Report

3 major / 3 minor

Summary. This RAND-style policy perspective compares U.S. and Chinese approaches to governing general-purpose artificial intelligence (GPAI) and argues that three areas of divergence are especially important for policymakers: the focus of domestic regulation (models versus content), the key principles of governance (who may develop and deploy GPAI, control, and state power), and the preferred forums for international governance (the UN versus more exclusive, alliance-based settings). The paper reviews recent policy documents, summit declarations, UN resolutions, and intergovernmental dialogues, and it closes with three recommendations for U.S.-China cooperation: focus on narrow aligned interests, design interoperable governance mechanisms, and clarify contested terms such as 'control.' Throughout, the authors are careful to note countervailing evidence and to hedge claims about China's trajectory, including the possibility that China is moving toward model-based regulation.

Significance. If the three claimed divergences are real, this paper provides a timely and useful map for policymakers working on U.S.-China GPAI safety cooperation. Its strengths are substantial: it is heavily sourced across English- and Chinese-language materials; it explicitly flags its own limitations, including that domestic policy is an imperfect proxy for international governance; it notes the translation ambiguity of 'anquan' (safety/security); and it acknowledges signs of convergence, such as China's possible shift toward model-focused regulation. The paper does not attempt a formal causal or statistical test, and it contains no fitted parameters or circular derivations; it is best read as an interpretive policy analysis. Its practical recommendations are modest and actionable. The main risk is evidentiary: the divergences rest heavily on official Chinese documents whose purposes may include domestic censorship and ideological management, and the paper's response to this risk is acknowledgment rather than systematic testing.

major comments (3)

[A Short History of U.S. and Chinese GPAI Governance, p. 2] The paper states that 'domestic policy is not a perfect proxy for international governance, but some strategic lessons might still be drawn,' but it never tests this proxy. This is load-bearing because the second and third divergences—key principles and international governance implementation—are inferred primarily from Chinese domestic documents and official statements that may serve domestic censorship or ideological-management goals rather than expressing GPAI safety governance. I recommend that the authors either add a systematic comparison of Chinese domestic versus international-facing AI documents (e.g., coding for the stated purposes of content-control measures) or explicitly reframe the claims as about domestic governance only, with international implications as hypotheses.
[Differences in the Focus of AI Regulation, p. 4-5] The claim that China's regulatory focus is on outputs and content rather than models is supported by citing two regulations and by the observation that 'no official policy document references a compute threshold.' The absence of a compute threshold is weak evidence, as the paper itself notes that two academic-led draft AI laws do contain compute thresholds, and absence may reflect drafting-stage timing rather than a substantive divergence. To make this divergence robust, I suggest either removing the compute-threshold absence as a supporting data point or presenting a more systematic comparison of the regulatory targets (e.g., which actors are subject to obligations: developers versus deployers) using a clear coding scheme applied to both countries' documents.
[Differences in Key Principles of AI Governance, p. 5-7] The selection of the three divergences is presented as self-evident ('Three areas of divergence are notable for policymakers'), but the paper does not explain how these three were chosen from the many possible differences. Since the central claim is that these particular divergences are the important ones, the authors should either make the selection criteria explicit (e.g., based on a structured review of policy documents) or frame the paper as an interpretive essay that offers one plausible framework, not the definitive set. Without this, a reader cannot assess whether omitted divergences (for example, export controls versus data governance) might be equally or more consequential.

minor comments (3)

[Differences in Key Principles of AI Governance, p. 7] The discussion of 'control' notes that China's usage must be understood against its history of content moderation, but it does not give the reader a method for distinguishing safety-oriented control from censorship-oriented control; a brief operational definition or example would sharpen the argument.
[Differences in Approaches to Implementing International AI Governance, p. 8] The statement that the U.S. delegation was led by national security authorities while China's was led by foreign ministry officials is presented as suggestive of different priorities; it would be helpful to note that delegation composition can also reflect bureaucratic division of labor rather than strategic intent, to avoid over-interpretation.
[References] Several citations are to blog posts and think-tank commentaries; this is appropriate for a policy perspective, but adding page numbers or paragraph numbers for direct quotations (e.g., from the UN resolutions and the Bletchley Declaration) would improve verifiability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an interpretive synthesis of external policy documents and official readouts, with no fitted parameters or self-derived predictions.

full rationale

The paper makes no formal derivation or predictive claim; it compares U.S. and Chinese GPAI governance by citing primary documents (e.g., Executive Order 14110, Cyberspace Administration regulations, UN General Assembly resolutions, and official readouts of the U.S.-China AI dialogue). The central three-way divergence is presented as an interpretive finding from those external sources, not as a consequence of the paper's own definitions or fitted values. The only reference co-authored by a current author, Elmgren and Guest (2024), appears in footnote 62 to support the side observation that China already has institutions somewhat analogous to an AI safety institute; that observation is also supported by the cited Central Committee resolution and is not load-bearing for the paper's central comparison. The stated caveat that domestic policy is not a perfect proxy for international governance is a limitation, not a circular step. No equation, parameter, or construction reduces the output to the input.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper relies on interpretive assumptions about what policy documents signify, but it does not introduce free parameters, fitted values, or new theoretical entities. Its main burden is the proxy assumption and the selection of documents.

assumptions (3)

domain assumption Domestic policy documents are a usable proxy for international governance priorities.
The paper states 'Domestic policy is not a perfect proxy for international governance, but some strategic lessons might still be drawn' (Section 'A Short History'), and the whole comparison rests on this inference.
domain assumption The selected policy documents are representative and sufficient.
The paper selects EO 14110, Chinese regulations, UN resolutions, and readouts without systematic criteria for inclusion or weighting.
domain assumption Chinese policy terms can be meaningfully compared to U.S. terms despite translation ambiguity.
The paper notes 'anquan' can mean both safety and security and 'control' has different connotations, yet proceeds to compare them (Section 'Differences in Key Principles').

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bridging the Artificial Intelligence Governance Gap: The United States' and China's Divergent Approaches to Governing General-Purpose Artificial Intelligence." pith.science (2026). https://pith.science/paper/NREROTG7

@misc{pith2026250603497,
  author       = {Pith},
  title        = {Pith review of: Bridging the Artificial Intelligence Governance Gap: The United States' and China's Divergent Approaches to Governing General-Purpose Artificial Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NREROTG7}},
  note         = {Machine review of arXiv:2506.03497}
}

read the original abstract

The United States and China are among the world's top players in the development of advanced artificial intelligence (AI) systems, and both are keen to lead in global AI governance and development. A look at U.S. and Chinese policy landscapes reveals differences in how the two countries approach the governance of general-purpose artificial intelligence (GPAI) systems. Three areas of divergence are notable for policymakers: the focus of domestic AI regulation, key principles of domestic AI regulation, and approaches to implementing international AI governance. As AI development continues, global conversation around AI has warned of global safety and security challenges posed by GPAI systems. Cooperation between the United States and China might be needed to address these risks, and understanding the implications of these differences might help address the broader challenges for international cooperation between the United States and China on AI safety and security.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

[3]

China Lashes Out at Latest U.S. Export Controls on Chips,

As of December 4, 2024: https://www.qinshui.gov.cn/xwdt_361/gwyyw_49700/202405/t20240509_1979919.shtml “China Lashes Out at Latest U.S. Export Controls on Chips,” Associated Press, October 8,

work page 2024
[4]

Who is leading in AI? An analysis of industry AI research

Concordia AI, “What Does the Chinese Leadership Mean by ‘Instituting Oversight Systems to Ensure the Safety of AI?’” AI Safety in China blog, August 2, 2024a. Concordia AI, “China’s AI Safety Evaluations Ecosystem,” AI Safety in China blog, September 13, 2024b. Cottier, Ben, Tamay Besiroglu, and David Owen, “Who Is Leading in AI? An Analysis of Industry A...

work page Pith review arXiv
[6]

CAICT’s 7th Batch of AI Model Evaluations,

As of December 4, 2024: https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm Ding, Jeffrey, “CAICT’s 7th Batch of AI Model Evaluations,” ChinAI Newsletter blog, August 5,

work page 2024
[7]

The Bletchley Declaration by Countries Attending the AI Safety Summit, 1–2 November 2023,

Government of the United Kingdom, “The Bletchley Declaration by Countries Attending the AI Safety Summit, 1–2 November 2023,” policy paper, November 1,

work page 2023
[8]

Historic First as Companies Spanning North America, Asia, Europe and Middle East Agree Safety Commitments on Development of AI,

Government of the United Kingdom, “Historic First as Companies Spanning North America, Asia, Europe and Middle East Agree Safety Commitments on Development of AI,” press release, May 21, 2024a. Government of the United Kingdom, “AI Seoul Summit: Participants List (Governments and Organisations),” webpage, last updated June 13, 2024b. As of November 26, 20...

arXiv 2024
[9]

The 2024 China-US AI Dialogue Should Start with an Eye on Chem- Bio Weapons,

Heslop, David, and Joel Keep, “The 2024 China-US AI Dialogue Should Start with an Eye on Chem- Bio Weapons,” The Diplomat, March 9,

work page 2024
[10]

As of November 25, 2024: https://idais.ai/ International Dialogues on AI Safety, statement at IDAIS-Beijing 2024, March 10–11,

International Dialogues on AI Safety, homepage, undated. As of November 25, 2024: https://idais.ai/ International Dialogues on AI Safety, statement at IDAIS-Beijing 2024, March 10–11,

work page 2024
[11]

Is Xi Jinping an AI Doomer?

As of December 4, 2024: https://idais.ai/dialogue/idais-beijing/ “Is Xi Jinping an AI Doomer?” The Economist, August 25,

work page 2024

Show all 19 references

[13]

U.S.-China Track II Dialogue on the Digital Economy,

National Committee on U.S.-China Relations, “U.S.-China Track II Dialogue on the Digital Economy,” webpage, undated. As of November 26, 2024: https://www.ncuscr.org/program/us-china-track-ii-dialogue-digital-economy/ National Technical Committee 260 on Cybersecurity of Standar...

2024
[14]

As of December 4, 2024: https://cset.georgetown.edu/publication/china-safety-requirements-for-generative-ai-final/ National Technical Committee 260 on Cybersecurity of Standardization Administration of China, 网络安全标准实践指南—人工智能伦理安全风险防范指引, January

2024
[16]

What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring,

Shavit, Yonadav, “What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring,” arXiv, arXiv: 2303.1134, May 30,
[17]

AI Advisory Body,

UN—See United Nations. United Nations, “AI Advisory Body,” webpage, undated. As of November 27, 2024: https://www.un.org/ai-advisory-body United Nations General Assembly, resolution on Seizing the Opportunities of Safe, Secure and Trustworthy Artificial Intelligence Systems fo...

2024
[18]

As of December 4, 2024: https://cset.georgetown.edu/publication/china-ai-law-draft/ Zhou Hui, Zhu Yue, Zhu Lingfen, Su Yu, Yao Zhiwei, Wang Jun, Chen Tianhao, He Bo, Hong Yanqing, Fu Hongyu, et al., The Model Artificial Intelligence Law (MAIL) v.2.0— Multilingual Version [人工智能...

2024
[19]

His research focuses on AI in China, coordination between AI safety institutes, and AI’s geopolitical effects

As of December 4, 2024: https://zenodo.org/records/10974163 16 About the Authors Oliver Guest is a research analyst in the international governance team at the Institute for AI Policy and Strategy. His research focuses on AI in China, coordination between AI safety institutes,...

2024
[2020]

Shanghai Declaration on Global AI Governance,

Ministry of Foreign Affairs of the People’s Republic of China, “Shanghai Declaration on Global AI Governance,” July 4, 2024a. As of December 4, 2024: https://www.mfa.gov.cn/eng/xw/zyxw/202407/t20240704_11448351.html Ministry of Foreign Affairs of the People’s Republic of China...

2024
[2021]

As of December 4, 2024: https://www.tc260.org.cn/file/zn10.pdf National Telecommunications and Information Administration, Dual-Use Foundation Models with Widely Available Model Weights, July

2024
[2022]

Interim Measures for the Management of Generative Artificial Intelligence Services

As of December 4, 2024: https://www.cac.gov.cn/2022-12/11/c_1672221949354811.htm Cyberspace Administration of China, “Interim Measures for the Management of Generative Artificial Intelligence Services” [“生成式人工智能服务管理暂行办法”], July 13,

2024
[2023]

Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims,

Brundage, Miles, Shahar Avin, Jasmine Wang, Haydn Belfield, Gretchen Krueger, Gillian Hadfield, Heidy Khlaaf, Jingying Yang, Helen Toner, Ruth Fong, et al., “Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims,” arXiv, arXiv:2004.07213, April 20,

2004 arXiv
[2024]

An AI Race for Strategic Advantage: Rhetoric and Risks,

Cave, Stephen, Seán S. ÓhÉigeartaigh, “An AI Race for Strategic Advantage: Rhetoric and Risks,” AIES ’18: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, December

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

China Lashes Out at Latest U.S. Export Controls on Chips,

As of December 4, 2024: https://www.qinshui.gov.cn/xwdt_361/gwyyw_49700/202405/t20240509_1979919.shtml “China Lashes Out at Latest U.S. Export Controls on Chips,” Associated Press, October 8,

work page 2024

[2] [4]

Who is leading in AI? An analysis of industry AI research

Concordia AI, “What Does the Chinese Leadership Mean by ‘Instituting Oversight Systems to Ensure the Safety of AI?’” AI Safety in China blog, August 2, 2024a. Concordia AI, “China’s AI Safety Evaluations Ecosystem,” AI Safety in China blog, September 13, 2024b. Cottier, Ben, Tamay Besiroglu, and David Owen, “Who Is Leading in AI? An Analysis of Industry A...

work page Pith review arXiv

[3] [6]

CAICT’s 7th Batch of AI Model Evaluations,

As of December 4, 2024: https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm Ding, Jeffrey, “CAICT’s 7th Batch of AI Model Evaluations,” ChinAI Newsletter blog, August 5,

work page 2024

[4] [7]

The Bletchley Declaration by Countries Attending the AI Safety Summit, 1–2 November 2023,

Government of the United Kingdom, “The Bletchley Declaration by Countries Attending the AI Safety Summit, 1–2 November 2023,” policy paper, November 1,

work page 2023

[5] [8]

Historic First as Companies Spanning North America, Asia, Europe and Middle East Agree Safety Commitments on Development of AI,

Government of the United Kingdom, “Historic First as Companies Spanning North America, Asia, Europe and Middle East Agree Safety Commitments on Development of AI,” press release, May 21, 2024a. Government of the United Kingdom, “AI Seoul Summit: Participants List (Governments and Organisations),” webpage, last updated June 13, 2024b. As of November 26, 20...

arXiv 2024

[6] [9]

The 2024 China-US AI Dialogue Should Start with an Eye on Chem- Bio Weapons,

Heslop, David, and Joel Keep, “The 2024 China-US AI Dialogue Should Start with an Eye on Chem- Bio Weapons,” The Diplomat, March 9,

work page 2024

[7] [10]

As of November 25, 2024: https://idais.ai/ International Dialogues on AI Safety, statement at IDAIS-Beijing 2024, March 10–11,

International Dialogues on AI Safety, homepage, undated. As of November 25, 2024: https://idais.ai/ International Dialogues on AI Safety, statement at IDAIS-Beijing 2024, March 10–11,

work page 2024

[8] [11]

Is Xi Jinping an AI Doomer?

As of December 4, 2024: https://idais.ai/dialogue/idais-beijing/ “Is Xi Jinping an AI Doomer?” The Economist, August 25,

work page 2024