Who should the U.S. Talk to in China on AI?

A who's-who of Chinese AI policy, with colorful org charts and a shareable deck.

Following the summit of President Trump and Chinese leader Xi Jinping in mid-May, the two sides announced that they would start a government-to-government dialogue on artificial intelligence (AI).

Agreeing to start a dialogue is a significant step, but it raises as many questions as it answers. Is productive US-China dialogue on AI even possible? (I think it is, if you come at it right.) If so, what should the countries discuss? (Open question, I lean toward best practices in testing models for mutual threats that cross borders.) Who should the U.S. try to engage with in China on AI?

In this piece, I’ll try to answer that third question. There’s lots of uncertainty here and I’ll do my best to account for parts of it by outlining the strengths and weaknesses of different Chinese interlocutors across different potential dialogue topics.

This piece will have three sections:

I. Executive Summary / TLDR: Who should the U.S. talk to in China on different aspects of AI?

II. Mapping China’s AI Bureaucracy: Colorful charts of the whole apparatus.

III. Strengths and Weaknesses of Different Actors in China: More in the weeds breakdown on who does what, and why certain orgs are best to talk to on certain topics.

I’ve also taken all the images and org charts from throughout the piece and turned them into this shareable Google Slides Deck. You can get the same effect by simply scrolling through this piece and checking out each of the charts.

Before beginning, three important caveats. First, the internal politics and bureaucratic / personal turf wars are really hard to understand from the outside, and so parts of this should be seen as (informed) speculation. Second, I’ve never been a U.S. diplomat or arranged governmental (Track I) dialogues, so I’m sure that I’m missing many key dynamics. I view figuring all this stuff out as a team sport, and I really welcome reader feedback on these previous two things.

Finally, the U.S. doesn’t simply get to pick it’s counterparts or independently decide the agenda for the dialogue. These will all be part of a negotiation with the Chinese side, including who they want to talk to from the U.S. government.

One final note that shapes my thinking on all this: I believe that (at least for now) the most important actions on AI governance and safety will be domestic actions that the U.S. and China take — not international or bilateral agreements. There may be some level of light touch coordination, but it’s unlikely to be “We do X because you’ve promised to do it too.” That said, I think engagement between the two countries can really matter, both for communicating about emerging risks/responses and for sharing certain safety-enhancing practices (such as approaches to testing and evaluation).

Part I: Executive Summary / TLDR. Who should the U.S. talk to on different aspects of AI?

Let’s begin. Any dialogue will include a delegation of representatives from different organizations on both sides, with one or two of those designated as the leads in the delegation. With that in mind…

• Given that it appears Secretary Bessent is leading on the U.S. side, his most logical counterpart is He Lifeng, Politburo member, lead on U.S.-China economic issues, frequent Bessent counterpart, and former head of the NDRC (see below).

• For a conversation on mandatory testing, evaluation, and standards for frontier AI models, the best counterpart would be the Cyberspace Administration of China (CAC) and its subordinate organizations.

• For a broad conversation on AI policy—including both development and big picture governance—the best counterpart would be the National Development and Reform Commission (NDRC).

  • The NDRC would also make for a logical lead of the delegation given its role in coordinating AI policy across the government (more detail below) and its strong ties to He Lifeng.

• For a conversation on the scientific trajectory of AI and long-term impacts, the best counterpart would be the Ministry of Science and Technology.

• For a conversation on real-world applications and related industry standards, the best counterpart would be the Ministry of Industry and Information Technology (MIIT).

• For a conversation on international or multilateral AI governance the best counterpart would be the Ministry of Foreign Affairs (MFA) and the related CCP commission that, the Central Foreign Affairs Committee (CFAC).

  • If the NDRC isn’t the lead of the Chinese delegation, it would likely be the MFA’s Department of Arms Control, which has been given the role of coordinating AI affairs across the ministry.

Putting this together, into a potential configuration for the Chinese side.

Potential US-China AI Dialogue Configuration

Note on the Ministry of Finance as a potential co-lead of the delegation: MOF Vice Minister Liao Min was reportedly involved in talks with Secretary Bessent to set up the dialogue before the Beijing summit. MOF has little direct involvement in AI policy, but maybe this early role (and the Finance<>Treasury parallels) means they will retain a coordinating role.

Aside from Secretary Bessent’s likely role as the delegation lead, I won’t speculate on the configuration or roles and responsibilities of the U.S. side. Those respective roles seem to be very much up in the air and I’ll leave that reporting and analysis to others. I imagine someone in China is writing the inverse version of this piece analyzing potential U.S. counterparts…

For reference, these were the delegations at the 2024 US-China AI dialogue, with the Chinese participants in the order listed in their readout:

• China: MFA (delegation lead, North America Dept), MOST, NDRC, CAC, MIIT, Central Foreign Affairs Commission。

• US: NSC, State, Commerce (I believe that’s it).

Ok, before diving deeper into the bureaucracy and strengths/weaknesses of each organization, it’s worth getting a little more detail on who does what in Chinese AI policy — I’ll expand on these below.

• The CAC — The Regulator. The CAC is China’s primary internet and data regulator, and has been the lead author and enforcer of China’s binding regulations on AI. It has regulated recommendation algorithms, deepfakes, generative AI, labeling requirements for AI-generated content, and more. Orgs under CAC — particularly the standards group TC260 and CNCERT — are critical players in the nuts and bolts of regulatory compliance, risk management, and related standards.

• The NDRC - The Coordinator. The NDRC’s core role is as a macro-economic planner. It previously had a limited role in AI policy, but after the debut of ChatGPT the top leadership gave it the role of coordinating AI policy across the different ministries. That coordination role wasn’t made public, and the mechanisms and effectiveness of this coordination aren’t clear. The NDRC has the lead role in implementing China’s AI+ (re: AI diffusion) plan, and also plays a critical role in compute and data issues.

• MOST — The Big (Scientific) Picture. MOST played a major role in China’s early AI development plans, and led on some “soft law” approaches to AI governance. After a 2023 re-org, it became the office of new CCP Central Commission on Science and Technology, sharpening its focus on strategic planning and reform of the S&T system. People there now describe it as “the OSTP of China.”

• MIIT — Tech Support. MIIT leads on industrial (broadly defined) applications of AI. It’s a co-signatory on the CAC’s AI regulations and does lots of the technical work to inform and support the regulations. Much of that is done via MIIT’s subordinate orgs, such as CAICT and industry standards bodies, which set standards and carry out various testing, evaluation and certification functions.

• Other relevant orgs include: the Ministry of Public Security (MPS), which focuses on the law enforcement side of regulation; the Ministry of Foreign Affairs (MFA), leads global AI diplomacy; the Ministry of Commerce, which deals with export/controls; and the Ministry of State Security (MSS), the foreign intelligence service that is involved in testing and evaluation related to cyber vulnerabilities.

I asked Claude to mock up a graphic based on the above text:

Alright, onto the charts.

Part II: Mapping China’s AI Bureaucracy

Note: Below are *simplified and illustrative* charts intended to highlight the key actors and their rough place in the bureaucracy. They are incomplete, inconsistent in some ways, and fuzz certain relationships that are either difficult to illustrate or unnecessary to communicate for this specific topic. I welcome any suggestions or corrections.

Simplified Chart of China’s AI-relevant Bureaucracy

Some notes on the chart:

• The entities in blue are formally part of the Chinese government, and the entities in red are CCP bodies; the purple ones have mixed party-state identities. Despite these divisions, the real political power lies with the Party, either through direct reporting chains or internal CCP Party Committees within state organizations.

• Almost all of China’s regulations on AI have been put out by the bottom row of actors: ministries, the CAC, and the NDRC. The entities above them are very important for political direction and approval, but are not usually in the weeds on regulatory requirements. Major development plans (2017 New Gen AI Plan and 2025 AI+ initiative) have come from the State Council.

• Within this bottom row, the CAC and NDRC are elevated slightly above their peers because of their increased power and importance on these issues.

• All of the ministries in the bottom row report up to a party commission, but I haven’t included most of them here.

Detailed (but incomplete / inconsistent) Chart of China’s AI-Related Bureaucracy

Some notes on this more detailed chart:

• Don’t be scared! This won’t be on the test. But it forms the basis for some of the arguments below on the strengths/weaknesses of different interlocutors.

• The names I’ve added here represent the Politburo or Politburo Standing Committee (PBSC) member with the most direct or indirect influence / oversight of an organization.

• The new bottom rows include several different types of organization that aren’t directly comparable to each other, including functional departments within ministries, subordinate agencies, standards groups, etc. I included them all because they play a meaningful role in Chinese AI policy and relate to the question of who to engage with over there.

• The dashed lines here indicate something other than direct reporting or control, but the nature of the relationships varies. Sometimes it’s on-paper (but not real) control, other places it’s real (but not on-paper) control, etc.

Ok, with those charts for reference, let’s dig into the strengths and weaknesses of different counterparts. One consistent tension you’ll see throughout: the people who are best positioned to engage with Americans often don’t hold real power within the system, and those with real power within the system often don’t have the experience or inclination to engage externally.

Part III: Strengths and Weaknesses of the Key Actors in China

Note: I highly recommend the Asia Society Policy Institute’s Decoding Chinese Politics project for deeper and more accurate dives into specific organizations and leaders. I’ve learned a lot from it and any errors below are, of course, mine.

Mandatory testing of AI models: the Cyberspace Administration of China + subordinate organizations

Over the past five years, the CAC has emerged as the key actor when it comes to binding regulation and testing of algorithms and AI models. The catch is that for most of that time, the CAC has focused its attention on the political risks of those models, what they call “content security.” The core motivation for its regulations on generative AI and recommendation algorithms were to ensure that AI’s role in creating and disseminating content didn’t upset China’s existing controls on online content. To implement those regulations, the CAC set up systems for the mandatory registration and pre-deployment testing of AI models, including accompanying technical standards on specific testing regimens.

But over the past ~2 years the CAC and its subordinate orgs have widened the aperture and begun to target other types of AI risks, including psychological harms, AI-enabled cyber attacks, and various practical risks related to AI agents. Last year, two key subordinate orgs — standards body TC260 and cybersecurity coordinator CNCERT — published the AI Safety & Governance Framework 2.0, an expansive taxonomy of risks with a significant focus on extreme risks like loss of human control. (My + Scott Singer’s analysis of the Framework.) That taxonomy is now forming the roadmap for batches of new technical standards. Many of those are being led by TC260’s newly established “AI Safety/Security Working Group” (WG9, see the China AI Bulletin for more on their work). Though new, WG9 appears likely to become a key node in how the Chinese party-state understands and deals with frontier AI risks.

The CAC’s role within the wider Chinese party-state bureaucracy offers both advantages and disadvantages. On the advantage side, the CAC serves as the office of the Central Cyberspace Affairs Commission, which was previously chaired by Xi and is now chaired by his chief of staff and right-hand man Cai Qi. That is a quite direct line to real power.

But there are also disadvantages. The CAC is fundamentally a Party-centric, content/ideology-focused, and largely domestic-facing organization. In the early 2010’s, the CAC had extensive engagement with U.S. tech companies (when they still thought they could get into China) and it still hosts the World Internet Conference in Wuzhen. But to my understanding it doesn’t have extensive experience engaging with foreign officials on technical or regulatory issues.

If a U.S.-China AI dialogue were to have a significant focus on testing, evaluation, and technical standards for frontier AI models, the CAC could be a good interlocutor on those issues, potentially as the lead of a working group. But it wouldn’t make for a likely lead of the whole delegation.

Overarching AI Policy Conversation (and delegation lead): the National Development and Reform Commission

The NDRC is China’s lead macro-economic agency and is known as a policy powerhouse and “super ministry,” a place that attracts top talent and wields significant power over funding. It’s early (pre-2023) role in AI policy came via projects on building out China’s compute infrastructure—originally the “Eastern Data, Western Compute” project and now the integrated national computing network.

After the debut of ChatGPT, the alarming gap with the U.S. in AI led top leadership to re-assess the balance between asserting control and encouraging development of technology companies. (This was in the wake of the 2020-2022 tech crackdown.) As part of that reassessment, they apparently tasked the NDRC with serving as a cross-departmental coordinator for AI policy.

This coordinating role was never made formal or public, and the exact mechanism by which the NDRC does this isn’t clear. One formulation I’ve heard is that one of the vice-chairs of the NDRC leads a cross-governmental working group on AI policy, but I can’t confirm that. How effective this coordination is also remains unclear, but since mid-2023 the NDRC has been listed as the second-leading signatory on key regulations. More recently, the NDRC was given the role of leading and coordinating implementation of the State Council’s AI+ plan, the State Council’s major AI promotion and diffusion initiative that I wrote about here.

What makes the NDRC particularly interesting here as a counterpart — and potential delegation lead — is the constellation of higher level leaders above it.

The NDRC is the focus of this chart, though MOF is included due to its role in establishing the dialogue (and thus its potential role in coordinating the dialogue).

Directly or indirectly, the NDRC is linked to three key high-level leaders: He Lifeng, Ding Xuexiang, and Li Qiang. The He Lifeng link is probably the most impactful. He was a vice-chair and then the chair of the NDRC from 2014-2022, and his protege Zheng Shanjie currently runs it. (The roles at the NDRC are also sometimes called “deputy director” and “director.”) He Lifeng is now a Vice Premier on the State Council and runs the office of the Central Financial and Economic Affairs Commission (CFEAC), the CCP commission that oversees work in these areas. He has also been Bessent’s counterpart on US-China economic issues. Ding Xuexiang is a member of the Politburo Standing Committee and an Executive Vice Premier of the State Council (higher ranked than He). His formal State Council portfolio there encompasses the NDRC and MOF. Both He and Ding also operate under Premier Li Qiang through their roles on the State Council and through He’s role on CFEAC (which has Xi as chair and Li Qiang as vice chair).

So the NDRC (and MOF) have diverse ties to multiple members of the senior leadership. To be honest, I don’t know what that means or how it nets out in terms of role, influence, etc. But on balance I’m inclined to think its a useful thing to have in a counterpart for a dialogue like this.

Finally, a note on two organizations under the NDRC. The National Data Administration (NDA) was created in 2023 with the mandate to catalyze data’s role as a productive force in China. This took some power over data out of the hands of the CAC, which previously led on data issues but focused on regulating data, not unleashing it. Notably, the first director of the NDA came from the CAC, perhaps as a political compromise. The CnAISDA (China AI Safety and Development Association) is a bit of a complicated story. The TLDR is that it was created as China’s first-pass attempt at an AI safety institute, but it’s not really equivalent to places like the UK AISI or the U.S. CAISI in either institutional design or mandate. (Bit more detail in this footnote.¹)

Scientific trajectory of AI and long-term impacts: Ministry of Science and Technology

MOST is a pivotal player across China’s science and technology apparatus, though its role is changing and the new equilibrium is not entirely clear. The biggest change came in 2023, when the CCP revived the Central Science and Technology Commission (CSTC) and declared that MOST would be the home of the commission’s general office. The CSTC-MOST relationship parallels the way that the CAC acts as the office of the Central Cyberspace Affairs Commission. This shifted MOST from being a primarily state organization to an increasingly party-facing entity.

The (re)creation of the CSTC marked a political consolidation and elevation of S&T policy, moving it from the purview of State Council ministry to the remit of a CCP commission. Politburo Standing Committee member Ding Xuexiang was named as the chair of the commission, and MOST Minister Yin Hejun was named the director of the office of the commission.

As part of the restructuring, MOST also shed some of its more peripheral responsibilities (e.g. agriculture-related S&T) and refocused on core tasks: strategic S&T guidance, planning (including how China can break through “chokepoint” technologies) and reforms to the S&T system. People in China have told me that post-2023 reforms MOST is supposed to be “the OSTP of China.” I’m not sure exactly how strong of a parallel there is—especially given MOST’s more hands-on role in allocating funding—but that’s part of their self-conception.

MOST has historically played a major role in key AI policies, but since the 2023 reforms it hasn’t been as involved in domestic regulations. Back in 2017, MOST was likely the lead author and was definitely the lead implementor of national New Generation AI Development Plan. In the following years, its expert committees released ethics and governance principles for AI, some of which have continued to influence the language of major documents such as China’s Global AI Governance Initiative. MOST also led the early push for mandatory S&T ethics reviews for sensitive fields, including AI, but when these were written into a final regulation in 2025 MIIT took the lead.

In terms of external-facing diplomacy, when China was invited to the 2023 AI Safety Summit in the UK, it sent MOST vice-minister Wu Zhaohui, a former AI scientist. Notably, for the 2025 Paris AI Summit China sent Zhang Guoqing — a Politburo member and vice premier who oversees MIIT — perhaps reflecting the shift in emphasis of the summit, as well as China’s stronger engagement with the summit. (Chinese leaders believed it had received a last minute invite to the UK summit, but got an early leader-level invite to Paris).

Today, MOST’s role should be seen as the body responsible for advising top leadership on S&T issues (via the CSTC) and a key funder for major national projects. It’s role in domestic AI regulation is largely indirect, but given its role in advising the leadership on big picture AI trajectories, if the U.S. government hopes to communicate things of that nature MOST should be a key node of engagement.

Quick notes on the MFA and party committees

There is lots more to say about other ministries—particularly MIIT and MFA—but for now I will spare you (and me) too much additional writing on that. Some quick notes here on the MFA and party-vs-state participation:

• At the MFA, the Department of Arms Control has been given the role AI policy coordination, with its department chief given the title of the “coordinator for AI affairs” (外交部人工智能事务协调员).

  • If the MFA is given the role of leading the Chinese delegation—a strong possibility—I would expect it to run through this department. That would mark an improvement over when the Chinese delegation to the 2024 US-China AI dialogue was led by the MFA department for North American and Oceanian Affairs.

• Historically, state (not party) organizations have been used for most aspects of diplomacy. This has evolved somewhat, and the Central Foreign Affairs Commission (CFAC) was listed as a participant in the 2024 dialogue. We’ll have to see if that trend continues or grows to encompass other commissions —considering that’s where real political power rests, I think it would be a good thing.

Conclusion

Whewww — we’ll call it there. If you got this far, thanks very much for reading and engaging. As mentioned above, I welcome any and all feedback on this, especially where people have insights from their own work. Finally, if you enjoyed the piece, please subscribe and please share. I’d love to widen the group of people thinking and working on this.

1. In 2024 the Biden Administration was preparing to host the first meeting of the International Network of AI Safety Institutes that November. China didn’t want to be excluded from this new group, and it was told by international AI policy interlocutors that to even be considered for an invitation it would need to form its own AI safety institute. Some leading AI policy / technical people in China who are invested in safety issues began pushing the government to form an AISI.

Through the late summer and fall the Chinese bureaucracy debated this, and I was told that there was competition over which part of the bureaucracy would own this new thing. Eventually they opted to create an “association” of existing organizations — CnAISDA — instead of a single AISI. That association is not formally part of the government but I and others were told that it is under the NDRC (another signal of the group’s growing role in AI policy). CnAISDA was formally launched at the 2025 Paris AI Summit. While several of the constituent orgs (Shanghai AI Lab, CAICT) and people (Andrew Yao, Xue Lan) play quite significant roles in AI testing and policy, its unclear if the organization itself has additional power beyond those existing institutions. For a detailed analysis of CnAISDA from 2025, see this piece.