Today I’m coming to you with my first reactions to Xi Jinping’s much-hyped keynote address to the World AI Conference (WAIC) in Shanghai. WAIC has turned into China’s biggest AI event of the year, and 2026 is the first time that Xi Jinping has come in person. That’s a significant signal and this is a very significant speech.

Speeches by Chinese officials can come off as very boring — and many of them are! But the words are always chosen very carefully and some of them are worth parsing closely. This is one of those. The speech is Xi’s most in-depth discussion of AI to date, and its coming at an incredibly dynamic moment in AI development and policy.

Up top here I will give three of my key takeaways from the speech, and below I have included the full English translation of the speech, with my commentary on specific pieces of it included as footnotes. (29 in total!) This is a new format for me so we’ll see how it works — let me know how the footnotes / annotations worked for you.

Further down I’ve included a suggestion on how best to read the annotated speech. Tldr: Read it in a browser (just click on this article’s title at the top of this email) or in the Substack app.

Three takeaways from the speech:

1. Xi doubled-down on China’s pitch to the world on AI: We are the open, cooperative partner who wants to help you build your own capabilities.

This has been the pitch for 3+ years and it’s implicitly (though very clearly) framed as a contrast to the United States. Xi’s 2023 Global AI Governance Initiative called out those unnamed countries for “forming exclusive groups to obstruct other countries from developing AI.” Since then, China has made AI capacity building for the Global South core to its pitch, culminating in the founding of the World AI Cooperation Organization (WAICO) in Shanghai this week. How successful that pitch and how impactful the capacity building ends up being is an open question.

2. Xi restated the status-quo on open-source: we should “encourage open-source AI.”

Many people in international AI policy circles were watching this speech for any signal on whether China is likely to walk back or hedge it’s support for Chinese companies releasing their models open-weight on the internet. The speech didn’t signal that change. Restating the status-quo is definitely notable, but I don’t take this as a major update in either direction. OS was mentioned in just that one line in the speech and not as a stark pledge about future policy directions. Also…

3. Xi concluded the speech with stark language on remaining ready to change course as circumstances change, and preventing “loss of control.”

Chinese official discussions of AI often include references to “agile governance” and keeping abreast of new trends in the technology. But the language used here was significantly stronger than normal. It was then followed by a rich sentence on how the “staggering speed” of AI development requires, among other things, that we “constantly refine measures to forestall loss of control.” As I note below, in Chinese AI policy discourse “loss of control” is used to refer to different things in different contexts: state/sovereign control, human control, industrial control, etc. More on that in the annotations below.

The words Xi uses are significant, both because of what they signal about his preferences and because people across the system will leverage the fact that he said them to advance their own goals. In everything from regulatory fights to major grant applications, being able to tie your thing directly to Xi’s words helps. But many of those words remain open to interpretation. I’ll be closely following how they get interpreted—and, more importantly, acted upon—in the year ahead.

Please remember: There’s a lot to digest from this one, and these are my first impressions of the speech without reading in-depth commentary from others. There are almost certainly things that I overstated / missed / misunderstood in here, so take it as a first draft analysis. I hope it’s helpful and look forward to reading what others think.

Note: I recommend reading the annotations either in a web browser or in the app. That way you can hover over a footnote number or click on it. If you hover over it, the footnote’s text will appear as a pop-up. If you click on a footnote, it will take you down to the bottom and you can click on that footnote number at the bottom to return to original spot in the body text.

If you read in email form, I believe you have to scroll up and down yourself.

The footnotes in the text appear small, so I’ve highlighted each one to make it easier to scan for them. I think the annotations get more interesting as the speech goes along, so stick with it or scroll down.

You can find the original Chinese text of the speech here.

Share

Xi’s Full Speech at WAIC 2026

[Before Xi’s speech, Shanghai Party Secretary Chen Jining gave welcome remarks.1]

Distinguished Colleagues and Guests,
Ladies and Gentlemen,
Friends,

Seventy years ago, a group of young scholars proposed the concept of artificial intelligence (AI) for the first time at the Dartmouth Workshop in New Hampshire of the United States2. In the subsequent 70 years, AI scientists and researchers from around the world ventured into this unknown territory, forged ahead through twists and turns, and made breakthroughs with persistent hard work. Seven decades later today, amid the new wave of AI development, we are gathering by the Huangpu River to discuss how to promote AI globally for the positive, for good and for humanity. All this makes our meeting highly important. On behalf of the Chinese government and people, I would like to extend a warm welcome to you all.

In the course of history, the invention of the steam engine heralded the industrial civilization, the widespread access to electricity brightened up modern society, and the birth of the Internet brought the entire world together3. Each of these technological revolutions has profoundly reshaped our way of work and life, and enabled a giant leap in economic and social development.

Today, major changes unseen in a century are accelerating across the world4. The new round of technological revolution and industrial transformation is advancing at a faster pace, and the world has entered an unprecedented period of active innovation on AI technologies. Intelligent connectivity, human-machine collaboration, cross-sector integration, joint creation and sharing, and other intelligent technologies are unleashing enormous power. All this carries within it great opportunities as well as challenges to governance. We human beings must answer the questions posed by our times: How to get along with thinking machines? How to ensure security5 when algorithm is part of decision making? How to tackle ethical challenges by technologies through adaptive governance?6 How to realize AI for all when the divide keeps widening? These questions demand serious consideration and real answers from the whole international community.

In China’s view, all countries should take a people-centered approach7 and develop AI for the positive and for good. We should ensure that AI is an important driver for shared prosperity and common security. We should join hands to build a just and equitable system for global AI governance. To this end, I wish to share four observations.

Subscribe now

First, we should adhere to the principle of openness8 and win-win and boost innovation-driven development. As a new engine of world economic growth and an accelerator for the shift of growth drivers, AI is moving from the digital world into the physical world. We should seize this rare, historic opportunity to encourage open source9, openness, collaboration and sharing. We should facilitate technological innovation, industrial development and scenario-based application of AI. We should make coordinated advances in the transformation and upgrade of traditional industries10, the cultivation and growth of emerging industries and forward-looking planning for future industries, so that all sectors and businesses can benefit from AI.

Second, we should strengthen risk-awareness and ensure that AI is secure and controllable11. AI should be a trusted tool for humanity. We should take seriously the various types of inherent and secondary risks12 that AI may trigger. We should put in place laws and regulations, technological monitoring, early warning and emergency response systems in order to strengthen the line of security13, prevent abuses and malicious use, and ensure that AI is always under human control14. In the meantime, we should jointly oppose overstretching the national security concept in the field of AI and placing one country’s security over that of others.15

Third, we should encourage inclusiveness and promote mutual learning between civilizations. AI development and its application should not erode or undermine the diversity of world civilizations or the uniqueness of cultures of different countries.16 We must shape the values of AI with humanity’s common values, and make good use of AI technologies to increase understanding, tolerance, exchanges, and sharing among all civilizations. We should tend to the garden of civilizations with great care to ensure that the beauty of each civilization is appreciated and shared.

Fourth, we should advocate solidarity and improve global governance. AI is an invaluable asset that encapsulates humanity’s collective wisdom. We should practice true multilateralism and recognize the important role of the United Nations.17 We should enhance alignment and coordination on AI development strategies, governance rules and technical standards, so as to form a consensus-based global governance framework at an early date to make this frontier technology better benefit humanity. We must carry out extensive international cooperation and help Global South countries with capacity building to bridge the AI and digital divides, promote sustainable development, and prevent creating new historical injustice in AI.18

Ladies and Gentlemen,
Friends,

This year marks the start of China’s 15th Five-Year Plan. It maps out China’s economic and social development for the next five years and provides immense opportunities for the international community. In recent years, China has embraced AI with open arms. We have promoted interplay between an efficient market and a well-functioning government, strengthened AI innovation19, actively advanced the AI Plus Initiative,20 and built a healthy ecosystem for all entities to thrive in together. The core smart economy industries are worth at least RMB 1 trillion yuan.21 Smart devices in countless homes truly improve people’s livelihood. Intelligent Manufacturing in China has become another shining hallmark of Chinese modernization.

At the same time, China lays great emphasis on safety and security in AI development.22 With a deep understanding of the trends and logic of AI development, we are continuously improving laws, regulations, policies, mechanisms, application norms, as well as ethical principles to make sure that AI is safe, secure and controllable, and that this fine steed of AI gallops with both speed and stability.

As a responsible major country, China is always committed to providing international public goods relating to AI. Since I proposed the Global AI Governance Initiative, China has promoted the adoption of the U.N. General Assembly Resolution on Enhancing International Cooperation on Capacity Building of Artificial Intelligence by consensus, published the AI Capacity Building Action Plan for Good and for All, announced the AI Plus International Cooperation Initiative, and advocated for establishing the World Artificial Intelligence Cooperation Organization (WAICO). China has been contributing steadily to the global AI governance.23

We often say in China, “A single string cannot make music, and a single tree does not make a forest.” AI development should not be a solo performance by a single country, but a symphony of international cooperation.24 Thanks to our joint efforts, WAICO has come into being in Shanghai. Our vision from one year ago is now a reality. This is a major move by China to answer the call of the Global South and unite the international community together to promote vigorously AI development and governance. It will be an important milestone in the history of AI development.

To further support global AI development and advance global AI capacity building, I hereby announce that in the next five years, China will provide developing countries with 5,000 opportunities in AI training and seminar programs25; develop international AI application cooperation centers with the Association of Southeast Asian Nations, the League of Arab States, the African Union, the Community of Latin American and Caribbean States, the Shanghai Cooperation Organization, and BRICS; and enable 30 countries to use the AI-powered meteorological warning system, or MAZU,26 to safeguard homes around the world.

Ladies and Gentlemen,
Friends,

As ancient Chinese observed, “A man of wisdom adapts to changes; a man of knowledge acts by circumstances.”27 With AI advancing at a staggering speed, we must ensure its development is for the positive, for good and for humanity. We must make its oversight and governance precise and effective, and constantly refine measures to forestall loss of control.28 We should always guide AI development with human wisdom and international consensus, so that AI can truly become a mighty force that increases the well-being of humanity and advances human civilization.29

China is ready to be more open, take more practical actions, and assume a more visionary perspective. We are ready to work with all parties to seize the opportunities of AI development and meet the challenges, and join hands to create a brighter future for humanity.

Thank you!

Matt here again —

I’ll leave it there. Again, consider this a first draft analysis, and one that’s very open to change. Would love to hear others’ take on the speech and feedback on this.

Enjoy your weekend!

Share

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.

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).

Share

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.

Share

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.

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.

Share

I did a version of this in early 2024 and again in early 2026 and the results were striking. In 2024, the Chinese participants ranked AI’s impact on jobs second to last—sixth out of seven. In 2026, they ranked it second from the top. It was virtually tied with “misinformation” for the top spot and far ahead of all the other options.

These are very unscientific and small-sample surveys. The participants are ~12 Chinese AI policy people who inform government regulation and company policy. There is selection bias and no guarantee that the rankings reflect someone’s deeply held beliefs. But sometimes the results map closely onto trends that I’m seeing across the entire Chinese AI policy ecosystem. The spike in concern about labor is one of them.

Over the past two years, worries about AI displacing workers and leading to structural unemployment have shot up in China. Those fears extend from ordinary people to the wider AI policy community to (as best as we can tell) high-level CCP officials. The fears are reflected in policy documents, state media, and the way Chinese people relate to the technology itself. And on top of these existing layers of concern is a new one: the CCP is also increasingly worried about Chinese people worrying about losing their jobs to AI.

In this piece I’ll break down the drivers of this shift, how it’s reflected in official documents and discourse, and then preview the “Do’s and Don’ts” of China’s emerging policy response.

Why weren’t Chinese leaders worried earlier?

I think the best explanation for the earlier lack of concern about AI-driven job displacement lies in the the last 45 years of Chinese history. Since the 1980s, China’s economy and society have been in a near-constant state of disruption. A Chinese person who is 60 years old today (about the average for high-level CCP leaders) has witnessed some of the fastest and most dramatic socio-economic transformations in human history.

You can throw out a bunch of statistics to try to illustrate this1 but those can’t fully capture the visceral feeling of social, economic and physical churn that characterized China since the 1980s. And through it all Chinese people displayed a genuinely awe-inspiring ability to adapt and make use of all the new opportunities.2 Between 1980 and 2020, China’s GDP per capita multiplied 25x.3

The lesson for government officials was pretty clear: as long as the size of the pie keeps growing, the overwhelming majority of people will be dramatically better off. The social disruption and job displacement will mostly come out in the wash.4 My guess is that CCP officials looked at projections for AI-driven productivity gains and figured that the same would be true for this technology. And lots of normal people likely had a similar, though maybe unarticulated, gut feeling. People have seen their lives and workplaces go through rapid technological upgrading for 45 years—and rapid diffusion of information technology for 25 years—and during that time they’ve gotten much richer and started working much better jobs.

But over the past two years these perception have changed. Driving these changes are a specific incident and a change in the wider social zeitgeist.

The Spark: The Wuhan Robotaxi Saga

For policymakers, a key turning point was a series of events surrounding the rollout of robotaxis in the city of Wuhan.5 It’s a bit of an odd story for a couple reasons.

1. There is a good amount of public writing about this, but I was told a slightly different version of it that I can’t find anywhere online.

2. Whichever version is true, the purported impact on policymaker thinking seems quite disproportionate to what happened. Perhaps there’s a lesson in there.

I’ll start with what’s publicly reported. In 2024, Baidu began rolling out robotaxis in Wuhan under the Chinese brand name “Radish Run” (萝卜快跑; the English brand is Apollo Go). This angered traditional taxi drivers and companies who had already been under pressure from ride-hailing apps for years. In late June of 2024, a Wuhan taxi company released a public letter decrying the shrinking margins for companies and declining income for taxi drivers. The letter actually spends more time on the way ride-hailing apps have hurt taxi drivers, but it was the parts about robotaxis that stirred up online debate.

It accused tech companies of “monopolizing resources” and “snatching the rice bowls” of drivers. “The original intent of technology was to make life better. In reality, it makes it so people at the bottom can’t even get enough to eat.”6

The letter sparked lots of conversation on social media, and that conversation really picked up a couple weeks later when a Baidu robotaxi struck a pedestrian in Wuhan. Public opinion was split on the accident; the person was jaywalking and wasn’t seriously injured. But the online conversation about robotaxis “stealing people’s rice bowls” continued to gain steam, becoming a top trending topic on Chinese social media.

Here’s where I heard something a bit different. The first time someone relayed this incident to me, they said taxi drivers had gotten together and led a “strike” against the robotaxis by repeatedly hailing them and then cancelling at the last minute, effectively paralyzing the system. I’ve searched online for any documentation or discussion of this “strike” but haven’t been able to find anything. Maybe it was just a rumor or a miscommunication, or maybe any discussion was effectively scrubbed from the internet (though it’s a bit unusual to do such a good job with that kind of scrubbing).

Either way, the result was the same: the outcry over robotaxis in Wuhan led to a significant shift in how the government and policy community thought about AI’s impact on jobs. It led officials to take the threat of job displacement—and also the public’s fear of AI-driven job losses—seriously. I feel relatively confident about the Wuhan robotaxi incident contributing significantly to this shift, because multiple influential members of China’s AI policy community independently relayed that causal relationship to me at different times.

This might seem a bit weird: how could a short-lived social media kerfuffle about robotaxis actually alter official thinking about AI’s impact on jobs? And admittedly, it does seem quite disproportionate. This is partly what leads me to give credence to what I heard about drivers organizing a strike against robotaxis. Collective action, especially around economic harms, tends to really get the attention of the CCP. The Party also tends to pay close attention to what is sparking anger or anxiety online. The flipside of maintaining a very tight censorship regime is that you’re also very attuned to what issues people are angry or fearful about.7

But the robotaxi incident was also happening against a wider social backdrop of mounting anxiety about jobs and what the future holds.

The Socio-Econo-Emotional Backdrop: Deep Jobs Anxiety

This is a very important part of the story, but one that I’ll be relatively brief with here. It’s a largely vibes-based thing and I’m just not in those streets the way I once was. But over the past ~4 years, the vibes around employment in China have been very bad. Things are nowhere near crisis levels, with mass unemployment or anything like that. But after the Covid lockdowns ended in late 2022, the promised economic recovery just never fully materialized for normal wage earners. Young people couldn’t find jobs, and people who had jobs started getting their salaries cut and stopped getting promoted.

It’s a very different picture of the economy than what you’d infer from reading headlines about DeepSeek models and BYD exports. Dinny McMahon of Trivium has done a good job describing this two-track economy of booming high-tech exports contrasted with anemic wage and consumption growth.

Over the past year, AI has further inflamed this anxious social zeitgeist. wrote about this in a good piece on growing “AI anxiety.” And here’s how summarized it in an insightful short note about what’s really driving the popular frenzy over OpenClaw.

China’s labor market has been under sustained pressure since COVID. Youth unemployment remains stubbornly high. Career mobility feels stuck for millions. AI has become a psychological escape valve. From DeepSeek in 2025 to OpenClaw now, Chinese media and short video platforms have been hammering one narrative nonstop: learn this AI tool, get a high-paying job. … What looks like grassroots tech adoption is closer to grassroots career panic…

Mixing these things together and you have what could become a pretty potent cocktail of social discontent. The Chinese leadership clearly felt they needed to start formulating some responses, and during the second half of 2025 those responses started to show up in policy documents.

Initial Policy Response: Assessments, Standards, Action?

The first major policy document that touched on this was the “AI+” plan, China’s landmark plan for diffusing AI throughout its economy and society. But the document also included some of the most direct language yet on AI’s potential impact on jobs:

Strengthen employment risk assessments related to AI applications (人工智能应用就业风险评估), guide innovation resources toward areas with greater job-creation potential, and mitigate impacts on employment.

When I mentioned this to some people, they pointed out that this seemed a bit contradictory. If the government is so worried about AI-driven job losses, why is it releasing this document that’s all about turbocharging AI applications across the economy? To which the Chinese state might respond:

Do I contradict myself?
Very well then I contradict myself,
(I am large, I contain multitudes.)

or

"The fundamental cause of the development of a thing is not external but internal; it lies in the contradictoriness within the thing.”

When I first read the AI+ mention of “employment risk assessments,” I assumed these were a hand-wavy thing that wouldn’t see real implementation. But a couple months later the standards group that covers the social impacts of AI finalized a national standard that included guidance on how to conduct these employment impact assessments for generative AI applications. Below is a machine translation of the relevant part of the standard:

This remains very high-level guidance and these aren’t mandatory yet. But this is often how regulation and standardization of technology works in China. Regulators or standard setters will first take a high-level pass at the problem, and then over the course of the following years take several more passes at it, each time getting more precise and practical. This is exactly what happened with China’s mandate for applying labels to AI-generated content, a requirement that first appeared in a fall of 2021 regulation on recommendation algorithms (but wasn’t really enforced) and was then iterated on in a series of regulations and technical standards over the following years, before being fully nailed down in the fall of 2025.8

In the months after the AI+ plan, we saw policy documents and official discourse around job impacts tick up. September 2025’s AI Safety and Governance Framework 2.0 (original doc, analysis) described the risk bluntly: “The roles of capital, technology, and data in economic activities are increasingly prominent, while the value of labor as a production factor is diminished, resulting in a significant decline in demand for traditional labor.” That marked a major change from the vague and benign language in the previous year’s Framework 1.0, which described AI as “accelerating the reconstruction of traditional industry modes, transforming traditional views on employment”.

The next significant step came in late January 2026, when the Ministry of Human Resources and Social Security announced:

Our country will implement actions to stabilize jobs, expand capacity, and improve quality; roll out employment support measures for key industries; and issue documents on promoting employment in response to the impact of artificial intelligence.

Pledging future policy action made a bit of a splash, though we’re now two months past that pledge and we have yet to see anything concrete emerge.

Most recently in March of 2026, the 15th Five Year Plan included a brief but direct passage on this: "Comprehensively address the impact on employment of changes in the external environment and the development of new technologies such as artificial intelligence. Establish an evaluation system for high-quality and full employment."9

Sidebar: The Twisty Bureaucratic Politics of AI

It’s always tough to decipher what’s going in the fierce and never-ending competition among China’s many ministries and other Party bodies. Here I’ll give a bit of what I’ve heard on this as it relates to AI and especially employment issues, but take the following with several grains of salt. And I welcome feedback from anyone who has heard or seen something different.

The National Development and Reform Commission (NDRC) is a policy powerhouse that acts as China’s main macroeconomic planner. At some point in 2023-2024 (I believe the second half of 2023) China’s top leaders designated the NDRC to act as a coordinator across AI policy issues.

Up to that point, the NDRC hadn’t been very active in AI policy. The Ministry of Science and Technology had provided an early push with the 2017 national AI development plan, and it later pushed for the establishment of AI ethics review committees. And from 2021-2023 the most impactful AI policy documents were the regulations issued by the Cyberspace Administration of China (CAC), governing recommendation algorithms and “deep synthesis”.10

But in the wake of ChatGPT, the leadership decided that you couldn’t leave AI policy primarily in the hands of a control-focused content regulator (the CAC). They needed to better balance the dual imperatives for control and development of AI,11 and so the NDRC was tasked with managing this holistic balancing. Exactly how the NDRC has played that role is unclear. In one version, the NDRC itself directly took on this task. In another version, a small interagency coordinating committee was formed, with a member of NDRC leadership at the helm. I’m not sure which one it is, but my guess is the latter.

While the NDRC taking the lead would generally seem bullish for AI policy (and on balance it has been), the NDRC is also an organization that has to worry about broader economic and employment impacts. And according to at least one telling, He Lifeng — vice premier and Politburo member who led the NDRC until 2023 — became personally quite worried about AI’s impact on employment. He apparently commissioned a study on this, though I haven’t heard of any results. This last part about He Lifeng is firmly in the realm of rumor, so again take it with substantial salt.

If these rumors are accurate, it has made for a bit of an odd role-reversal. For most of the past decade, the CAC was perceived as the “bad cop” of tech policy, focused primarily on information risks and unafraid to harshly regulate and punish companies. But in the wake of ChatGPT, the AI priorities of top leaders began to shift from control to catch-up, and the CAC quickly got the memo. It greatly watered down its generative AI regulation, and has reportedly taken a far more accommodating approach to companies (“let us help you comply with these regulations”). At the same time, the more development-focused (“good cop”) NDRC is allegedly beginning to worry more and more about AI’s impacts on employment.

What’s Next for AI x Employment Policy?

The policy documents mentioned above give a picture of growing concern, but aside from “employment risk assessments” there hasn’t yet been much in the way of concrete responses.

We get a clearer picture of likely policy responses from how state media has been writing about AI and jobs. The past few months have seen a slew of editorials and commentaries that are both hashing out potential responses and “guiding public opinion.” We can lump these responses into some Do’s and Don’ts.

The Do’s are nice ideas, but probably not super high impact in the short term. The idea of steering AI development toward job-creating applications sounds good, but I start from a place of skepticism about this. Yes, the government can give nudges around the edges and it can really rev up activity in sectors that already have momentum. But short of very heavy-handed interventions, the profit motive will still be the giant magnet that pulls AI research and adoption forward. And that profit motive doesn’t care about workers—in most cases, the fewer of them the better.

On education, a March People’s Daily commentary on why AI is not stealing people’s rice bowls tried to calm nerves and put the focus on education. In one paragraph, it lists a number of education-related proposals from the 15th Five Year Plan (especially, “lifelong learning” 终身学习) and describes these initiatives as “providing a safety net for workers.” While education might be helpful in navigating a changing labor market, it’s telling that none of these components of the “safety net” actually involve redistributing financial resources.

These references to lifelong learning, job retraining and workers adapting show up in many other places of official discourse. As mentioned above, during my time living in China I was regularly in awe of how adaptable Chinese people were over the course of 45 years of economic transformation. But personally I think this is not going to cut it with AI.

The Don’ts are, in my opinion, more interesting:

Panic! The Science and Technology Daily made this clear in a recent piece: “AI Must be Controllable, and Technological Anxiety Must Also be Controllable.”12 It covered the recent OpenClaw mania in China and then the social currents it reflects.

“Looking back at this wave of enthusiasm, OpenClaw's impact has long since transcended the tool itself. It functions more like an emotional amplifier, reflecting a pervasive and hard-to-place anxiety of the moment — the fear of being "left behind" by the tide of artificial intelligence.”

The article calls on technology companies not to overhype capabilities or ignore security risks. And it tells online platforms to stop playing up content telling people that if they don’t use AI they will get left behind. “The technology is accelerating, but societal sentiment need to be buffered.”

This next policy response is maybe the most interesting one: Don’t let companies fire people just because AI can replace them. This specifically came out of a high-profile labor arbitration case highlighted by Beijing’s local Bureau of Human Resources and Social Security. In the case, an arbitrator ruled that firing someone because AI can now do their job constitutes a violation of the Labor Contract Law. It turned on the question of whether AI’s rise constituted “a material change in the objective circumstances on which the labor contract was based,” thus allowing for the termination of a contract. Chinese labor law is definitely not my area, so I’ll just relay how Xinhua framed it in a recent commentary:

The law is clear: "objective circumstances" of this kind must be both unforeseeable and beyond the parties' control. A company's decision to adopt AI is a voluntary business judgment, with all associated risks squarely within its own control. When an employer terminates a labor contract on the grounds that a position has been "replaced by AI," it is in substance offloading the normal risks of technological iteration onto the worker. … Companies that enjoy the dividends of AI must also fulfill their responsibility to protect workers' employment.

This came out of the Beijing bureau of the Ministry of Human Resources and Social Security, the group that has promised forthcoming action on AI x Jobs. So maybe this is a preview of that.

But it raises as many questions as it answers.

1. Is the Ministry’s position here shared across the government? The Xinhua commentary suggests a certain level of approval across the system, but isn’t the final word.

2. Is labor law an effective tool for protecting workers rights at scale? Again, not my area and my initial hunch was “no.” In the Chinese companies I worked at employees often had (or felt they had) essentially zero leverage, and I never heard of anyone taking their bosses to court. But when I asked a friend about this they said these types of lawsuit and arbitration are common and utilized successfully by workers.

3. Can this approach really hold up if competitive pressures mount on these companies, or the economy as a whole? (More on this below.)

My guess is that the authorities see this not as a permanent solution but as a useful source of friction to slow down the transition and buy some time. And maybe that’s enough.

Alright, final one: Don’t become a welfare state. Xi Jinping himself has made it very clear that he sees an overgenerous welfare state (“welfarism”) as a huge problem. He said as much in a 2021 speech and a 2022 piece in the ideological journal Seeking Truth, saying China must “steer clear of the idleness-breeding trap of welfarism.” In the speech he specifically laid into Latin American countries whose overgenerous welfare states created “a group of lazy people with unearned incomes.” That might seem surprising coming from an avowed Marxist-Leninist like Xi, but he tends to adhere to those ideologies in their more philosophical (dialectical materialism) and organizational (vanguard party) aspects.

Of course it’s possible that Xi changes his tune on this. Some of his criticism of welfare states focuses on governments going beyond their financial means and raising peoples expectations in a way they can’t satisfy. But there does seem to be a harder and more ideologically-engrained edge to this that goes beyond the pragmatic considerations. It’s a somewhat hand-wavy take, but Xi came up through a ruthless and brutally competitive system, and he came to dominate it through shows of individual strength and cunning — not really the kind of life experience that makes you big on handouts.

Putting it all together

If we try to put these policy responses together into a coherent plan, we get something like the following:

1. Don’t scare people and don’t become a welfare state.

2. Make AI companies put some thought into labor impacts and provide the government with information on this front.

3. On the margins, try to nudge the technology toward labor-intensive applications.

4. Build in some friction to layoffs and force companies that are profiting from applying AI to continue supporting their workers.

5. Encourage and provide resources for people to continue educating and retraining themselves throughout their careers.

Laid out like that, it doesn’t look too bad. To be honest, when I started the research for this piece I came in with a much lower estimation of the Chinese state’s thinking on these issues. After combing through the material, I found some more sophisticated thinking than I’d expected (especially given how recently this came on the radar) and also some alternative tools that I hadn’t considered.

But, for me at least, #4 in that list is doing a lot of the work. Maybe there’s a meaningful play to make in this space, especially given the dominant and often unique role China plays in global manufacturing supply chains. Companies with hundreds or thousands of global competitors nipping at their heels might be legitimately unable to keep already-automated workers on their payrolls. And that might have been the case for China 25+ years ago, when it was just another replaceable module in the global market for cheap labor. But given the way China has come to dominate these supply chains, and its uniquely tight-knit and sophisticated manufacturing ecosystem, maybe those pressures won’t apply to China in the same way today.

Ok, that’s all for now. This piece began as a “I’ll just spend the next couple hours throwing together a quick take on…” and then snowballed into something much bigger. I hope it’s been helpful.

Thank you to Carnegie Junior Fellow Sophie Zhuang for her help with research on this piece.

So from time to time I’ll share some of that work on here. This piece was originally published in 2017 on The China Project, the former home of the podcast and tons of great writing about China. The China Project closed up shop a few years ago but you can still access their archive on the website. Thanks to them for giving this odd story a home back then.

Last thing: I recently created a little personal website to house some of my favorite work from years past. It’s a work in progress but already has some fun stuff on there.

Alright — on to the story. I’ve left the text alone and just added a few footnotes with updates. Enjoy!

I was never planning on attending 2015’s 两会 (liǎnghuì), the annual “Two Sessions” series of political gatherings in Beijing. The meetings in question are the largely ceremonial National People’s Congress (NPC), and the even-more-ceremonial Chinese People’s Political Consultative Conference (CPPCC). If U.S. party conventions are the Broadway of American political theater, liang hui is an uninspired table-read of Soviet dramas.1

The NPC culminates in a press conference held by the premier, a rare chance for foreign media to directly address questions to him2. Accomplished Politburo watchers can parse the subtle tweaks in the premier’s phrasing, but I never took much interest in the event. As the China correspondent for Huffington Post, my reporting tended to focus on culture and the everyday lives of Chinese people. Liang hui felt like a world away.

That March, I first learned that I’d have the honor of attending Premier Li Keqiang’s press conference while being chased around the Tianjin train station by four plainclothes goons.

How to shake a tail while talking to the Foreign Ministry

It was one week before the event, and that morning, I’d hopped on the Beijing-Tianjin high-speed train to scope out a protest in Tianmu Village 天穆村, on the northern outskirts of Tianjin. Locals there had taken to the streets to demand the ouster of Mu Xiangyou 穆祥友, the village Party Secretary and a delegate at the National People’s Congress. Protesters accused him of selling communally owned land for private profit, and of running village politics like an organized crime syndicate.

I’d gotten in touch with one of the protesters, and he’d invited me out to Tianmu the next morning. He said he would meet me at the train station, provide me with documentation for their claims, and even let me spend the night at his home in the village.

So the following morning, I hauled myself across town to Beijing South Station and caught an early train to Tianjin. As I stepped off the train and onto the platform, my phone buzzed with a message from my contact: “I’m at the police station now. I will call you later.”

That’s a gut-wrenching moment for any reporter. Walking down the tunnel toward the station’s exit, I both regretted coming and felt more determined to finish the job.

Halfway down the causeway, something over my shoulder caught my attention. Turning to the left, I saw a beefy Chinese man hustling to conceal a cell phone camera in his armpit. I’m guilty of plenty of blind spots when it comes to China, but as a tall white man who spends lots of time in third-tier cities, I have a spider sense for when someone is secretly taking a picture of me.

Attempting to distinguish between an innocent fanboy and state-sponsored surveillance, I picked up my pace. The man, followed by three accomplices, half-jogged to keep up. When I arrived at the exit carousel, I took a half-step through the gate, then doubled back into the causeway. The move worked, sending one of my stalkers to the opposite side of the one-way carousel.

Three remained. Exiting the station, I managed to lose another when he unwisely boarded an escalator and I reversed down some stairs. I twisted through the winding streets and alleys around the station. Left, left, right, backtrack, hide behind a car…left, left.

As I scanned the horizon for the remaining two guys, my phone started vibrating. It was a Beijing number, but not one I recognized.

“Wei?”

“Hello, Mr. Matthew? This is Beth. Are you in Beijing?”

Beth was my main handler at the Ministry of Foreign Affairs. She oversaw a suspenseful eight-month wait for my first journalist visa, and recently chaperoned my renewal.

“Hi Beth, no… actually, I’m not in Beijing right now.”

Beth rarely called me, and it was an unusual question to lead with. I hedged on giving my exact location, trying to tease out whether Beth was supervising this dragnet from afar. Her next question surprised me:

“If you could ask a question to Premier Li Keqiang at next week’s press conference, what would you ask?”

That caught me off-balance. I’d already missed the deadline to register for liang hui and had no intention of going. Racking my brain for intelligent-sounding questions, I came up empty. I asked her if I could send her something by email. She instructed me to send along a list of three questions I’d like to ask, ASAP.

I hung up, baffled. How could they be so thirsty for questions that they were ringing me up?

A village protest

I was yanked back into the moment at the sight of one of the men following me. I ducked into a subway station and did my best Jason Bourne: Hop on a train, ride it one stop, wait until the doors start to close, hop off and board a train going in the opposite direction. I was pretty sure I’d lost them. After a few stops, I walked out of the station and jumped into a cab. I told the driver to take me to the Tianmu Village Party Committee headquarters.

On the ride over, I tried messaging my contact in the village again. No response.

When we arrived, I was surprised to see a few hundred people gathered in front of the Party headquarters. Tianmu is a village predominantly inhabited by Hui, a Muslim ethnic minority that traces its roots to Silk Road traders; some protesters were sporting the white skullcaps worn by many Hui Chinese. Some used wooden poles to hold aloft banners calling on central government leaders to investigate the village Party Secretary. But there were no chants and no speeches, just a large crowd facing an ugly building.

Unsure of what to expect, I opened the cab door and stepped out into the crowd. I tried to stay low key, but unfortunately, there’s just nothing discrete about a tall white guy infiltrating a Chinese protest. Murmurs rippled through the crowd. I twisted my face into a goofy grin.

That smile vanished quickly. I hadn’t been in the crowd for more than 30 seconds when I was surrounded by a group of six men. They grabbed my jacket sleeves and hustled me backward toward a car waiting with its door open. The ringleader was barking questions at me, demanding to know who I was and why I was there. I threw the question back at him, and he replied that he was a member of the Village Party Committee.

The tussle continued — I didn’t know what the endgame was, but I knew I didn’t want to get in that car. At this point, a group of sturdy middle-aged women came to my rescue, wedging between the men and me. Wagging their fingers and shouting at the men, they cleared a path for me to move back into the crowd.

“Take pictures!” one of them shouted.

I took a few steps and a few deep breaths, but my fingers were shaking as I took my camera from its case. A man with slick black hair slid next to me. He wasn’t wearing a uniform, but he didn’t need to; he carried an air of authority and control.

“Don’t do anything,” he said, not looking at me. “Don’t cause a disturbance.”

I took his advice and froze. Before I was able to take a single picture, a uniformed police officer approached and demanded, in English, that I follow him to the police station in order to “verify my credentials.” I explained that this wasn’t necessary, but he insisted. I was unsure of what to do, but decided resisting arrest was likely unwise. He led me to his police car. I was plopped in the middle of the backseat, and two plainclothes men piled in on either side of me, including the man who warned me not to cause a disturbance. We jumped on the highway and headed north, away from the city center.

Taken away by the police

During my escape at the train station, adrenaline was coursing through my body, but in the police car, I just felt scared and alone.

Most foreign journalists working in China are ensconced in a bureau staffed with editors, fellow reporters, translators, and local fixers. If things get tense in the field, there’s a team that knows where they are and how to bail them out. But I was almost always flying solo. My closest editors were across the Pacific, and I only worked with translators in regions with incomprehensible dialects. My “Beijing bureau” was a small, wooden table at a mother-and-toddler café near my house called Let’s Lego! Sometimes I bounced ideas off the baristas. My most searing prose came while listening to yuppie moms singing “The Eensy Weensy Spider” during ukulele class.

No one in the world knew I was smushed in the backseat of a police car. So as our ride sped through the industrial outskirts of Tianjin, I frantically sent WeChat messages with my location to an American friend. I spammed him with photos and any geographic detail I could gather, trying to leave behind a trail of digital breadcrumbs.

I never thought of the journalism I did at Let’s Lego! as a threat to the halls of power in Beijing — for that, you needed resources and real investigative chops that I didn’t have. I mostly stuck to what I had access to: slow trains, coal towns, young people.

But I always felt the most like a target when reporting at the village level. Nothing I could unearth in Beijing was putting Xi Jinping in jeopardy. But a damning article about a village Party chief carried different consequences. That village official may be expendable to the higher-ups: If he became enough of an embarrassment, they could dispose of him with corruption charges. To reporters, it would be just another “fly” swatted in the anti-corruption campaign. But to the implicated Party chief, that one article could mean the difference between swimming in wealth or rotting behind bars. Put any person in that position and there’s no telling what they might do.

After 15 minutes, our car stopped at a police station in an industrial park. They marched me upstairs to a dim office. Inside, there was a large leather couch and watercolor paintings of fish hanging from the green walls. They poured weak tea into paper cups.

The plainclothes cop who rode with me was bristly and demanded to see my credentials. I handed them over, along with Beth’s business card at the ministry. He eyed them with suspicion and went next door to make a call.

While we waited for his return, the young English-speaking cop apologized for the hassle. I knew it wasn’t his fault, but I also wasn’t in the mood to make nice. Several minutes later, the other cop returned, looking peeved. Beth had backed me up and told him that I was free to report in Tianjin.3

He changed tactics: Yes, I theoretically could report in the village, but it was their obligation to protect my safety. Before letting me go, I had to speak with the local propaganda officials. They were on their way over now. Thirty minutes of excruciating small talk later, in rushed a middle-aged woman from the propaganda bureau, Ms. Liu. Clearly frazzled, she nevertheless put on a smile and welcomed me to their little corner of Tianjin.

For over an hour, we argued in circles. I told the group that there was no reason to detain me, and I planned to promptly return to the protests. They threw up a succession of reasons why that was impossible: People at the protest didn’t want you there (“I don’t think that’s true,” I replied); it would be dangerous for you to return (“I’ll take my chances”); you have no way of getting back there (“I can figure it out”).

They kept offering to take me back to the train station, but I declined. Every time I stood up to leave, they fussed until I sat back down. Finally, Ms. Liu offered her version of a compromise: If she could get representatives from the Village Party Committee to speak with me, would I then go back to Beijing?

Seeing no other way out of this stalemate, I agreed. I figured the village officials wouldn’t reveal anything new, but they would at least be good for a quote telling me to get out of town.

While Ms. Liu busied herself making arrangements in the next room, I chatted with the young cop. His English was pretty damn good, but he was shy about using it. He grew up in Tianjin, earned a master’s degree in England, and returned home to work in the Public Security Bureau. He seemed embarrassed about the situation.

Over the course of these conversations, my anxiety dissipated. Yes, this song-and-dance was frustrating, but none of these PR flacks or industrial-park cops appeared ready to assault or kidnap me.

Things get nasty

When Ms. Liu returned, behind her were two men, one short and built like a barrel, the other tall and thin. I stood up to greet them. They were dressed like the many village men who had become urbanites by osmosis: dirty dress loafers, dark slacks, and blazers made of glistening cheap fabric. Both men wore small white skullcaps. Scanning the room with suspicion, the men shifted from one worn loafer to the other.

Ms. Liu gestured toward the tall man with a smile.

“This man is from the Village Party Committee,” she said. “He will talk to you.”

The man hesitated for a minute, then spoke up in a quiet voice.

“This is a problem within the village and we’ll resolve it within the village. It has nothing to do with you and you’re not welcome here.”

I opened up my notebook and began to scribble, but the moment my pen touched the paper, the barrel-chested man exploded.

“Who are you?! Where are you from? What the hell is he writing in that notebook?!?”

I shut the notebook and threw up my hands.

“Look, sorry, it’s a misunderstanding. I don’t have to write anything down.”

But he wasn’t assuaged. He took a step closer, shouting at me from across the tea table.

“You’re not welcome here! Do you hear me?!?”

Ms. Liu tried to calm the man, grabbing his upraised arm and talking him down. Instead he redirected his rage toward her, snarling in her face and pushing her until she was backed up against a desk in the corner. One of the cops was yelling at the man, saying over and over that she was with the local government.

“I don’t give a fuck!!”

I locked eyes with the young cop and he gestured toward the door opposite a cowering Ms. Liu and her abuser. We slipped out of the room at the same time three new policemen rushed in. They shut the door behind them; all I could hear was a muffled shouting match. The door opened just long enough for Ms. Liu to run out in tears, and then slammed shut again. I stood beside the young cop in the stairwell, my mind spinning.

The past minute had shattered my mental model for Tianmu officialdom. I’d envisioned the cops, propaganda officials, and village party committee as a united front, members of one solid Soviet Bloc. But then the squat fellow went all Mao on Ms. Liu’s Khrushchev. Alliances were fracturing and I wanted out.

I turned to the young cop and told him I was ready to go to the train station. Two policemen squeezed out of the office. They huddled with my new bodyguard. Though they were speaking in hushed tones, it was a small staircase and I could overhear them.

The man locked in the room was making phone calls, and the policemen believed he was sending men to block key roads to the train station. There were only a few roads out of the industrial park we were in, and if his men managed to head us off, “there’s no telling what would happen.”

Seeing the fear flash across their faces, my imagination was off to the races. I visualized a cross between the Boxer Rebellion and Black Hawk Down. I was shit-brick terrified.

The group turned toward me and told me to follow another officer downstairs to a waiting van. I told them that I wanted the young cop to come with me. His superiors looked back and forth between us and nodded their approval. We hustled down the staircase and into a beige van.

The driver threw the van into high gear and we were off. As we weaved our way out of the industrial park, my cop received a steady stream of phone calls. He told the driver which roads to take and which ones to avoid. We screamed through the northern outskirts of Tianjin, slowing a bit as we followed the Haihe River through the city proper. Pulling up to a public square on the south side of the train station, the driver parked with two wheels up on the curb. We half-walked, half-jogged across the square, eyes constantly scanning the crowds milling around the entrance.

My entourage flashed badges to hustle me past security and ticketing lines. I directed an apologetic grimace at the migrant workers waiting in line. I’m sure they assumed it was just another instance of white-man-in-China privilege. They were not entirely wrong.

We cruised into the station and onto the train platform. My handlers scanned the adjacent train cars and escorted me to my seat. I asked them to promise that they wouldn’t share any of my information with the man in the office. They agreed and stepped off the train. The train slid into motion. They smiled at me and waved good-bye.

Questions for the premier

Back at my apartment in Beijing, I flopped onto my bed. What the hell just happened in that office? Was I actually in danger? Could I still be in danger? Replaying the day in my head, I thought of my contact in the village. I’d been calling and texting him throughout the day, but heard nothing.

I remembered that I had promised Beth three questions for Premier Li. Fundamentally, the annual press conference is a scripted affair. Moderators appear to be picking from a crowd of eager journalists, all waving to get their attention, but the questions are screened in advance, and the vast majority are softballs.

Foreign journalists are ambivalent about the process, with some choosing to boycott. Others, given the chance to directly address a Chinese leader, skewer them on sensitive topics, asking questions for which there’s no hope of getting an answer.

I don’t like softballs or grandstanding, so I opt for a middle ground: tough questions, but ones that don’t unnecessarily poke a finger in Li’s eye. My three questions covered social stability in the context of industrial layoffs, the economic implications of internet controls, and the topic that had fascinated me of late: the hyper-viral pollution documentary Under The Dome, produced by Chai Jing 柴静, a former investigative journalist for state TV [here is a Sinica Podcast about the film].

Filmed in the style of An Inconvenient Truth, Under the Dome is a 104-minute examination of China’s pollution woes. With narration delivered by Chai in a style similar to a TED Talk, the film weaves together a moving story about Chai Jing’s young daughter with a highly detailed look at the history, science, and politics of China’s air pollution.

Chai casts massive state-owned oil and gas companies as the key villains, arguing that they had abused their power to neuter environmental regulations. Naming and shaming the state-owned oil companies is a bold move. PetroChina and Sinopec are two of China’s biggest companies. Notoriously corrupt and politically powerful, their tentacles stretch into all corners of Chinese society.

The movie had dropped like a bombshell one week before I went to Tianmu, racking up over 100 million views over the weekend. It was the buzz of conversations on the street and the internet, and it initially received a glowing reception in official state media. In the days after its release, I contributed to a project that was crowdsourcing English subtitles for the film. China’s new minister of the environment said it reminded him of Silent Spring, the hit 1962 environmental science book by Rachel Carson that documented the environmental destruction caused by indiscriminate use of pesticides.

And then, after four days and 300 million views, Chinese censors ordered the film taken down and discussion of it censored. Chinese citizens were whiplashed. The film had taken the country by storm, then vanished just as quickly.

Many commentators viewed this as proof that Chinese leaders couldn’t handle the film, but I read it differently: If Xi’s agenda was a Venn diagram, Under the Dome stood right at its center. Xi had called out corruption, pollution, and economic stagnation as the three scourges on Chinese society. For three years, he had directed corruption investigations like a cudgel to wipe out obstinate officials whom he deemed obstacles to reform.

Of course, his chief political rivals just so happened to lead some of China’s most corrupt, polluting, and stagnant industries: coal and oil. Zhou Yongkang 周永康, an ex-Politburo member and architect of the security state, spent decades building a power base in PetroChina’s parent company, CNPC. Ling Jihua 令计划, a top aide to Hu Jintao, had deep-rooted family and patronage networks in China’s coal heartland of Shanxi. In 2014, both men became some of the highest-ranking leaders ever taken down on corruption charges.

What Chai Jing did in Under the Dome was spin a damning narrative about these industries, and tie it up with a bow. The film explained why Chinese people’s lungs and pocketbooks were suffering, and it directed all of the outrage at Xi’s rivals. Xi couldn’t have done it better if he had directed the film himself.4

The sudden censoring of Under the Dome wasn’t about what the film said, but who got to say it. Chai was the perfect messenger to get the story out, but Xi’s cohort wanted to make sure she didn’t turn her fame into an independent source of power.

I’d been tracing these strands at the grassroots for the past six months, visiting run-down coal villages and steel towns, victims of both economic restructuring and environmental crackdowns. Those towns stuck with me, and they struck me as core to figuring out where China was going.

So, facing my computer screen on that Friday afternoon, I wrote out my third question:

In Chai Jing’s recent documentary Under the Dome, she argues that PetroChina and Sinopec are obstructing the setting and enforcement of environmental standards, especially in terms of gasoline standards and natural gas. Do you believe these companies are obstructing environmental enforcement, and if so, what will the central government do to break through that obstruction?

Hitting send, I knew the question had no chance of being picked. It shouted out a banned documentary and laid down the gauntlet on factional politics. Even if it did jive with the Xi agenda, they wouldn’t let those words be spoken live on national television. I likely wouldn’t get to see the Great Hall of the People, but I was fine with that.

After I snapped shut my laptop, my phone buzzed. It was a message from my contact in the village. He had been held at the police station all day, but he was out now. He said he was doing all right. We promised to stay in touch. I went to the kitchen, cracked a Yanjing, and took a deep gulp of warm beer.

A surprising phone call

By Wednesday, I had forgotten all about the questions I’d sent to Beth. But that afternoon, my WeChat buzzed with a message.

“Thank you for your questions, Mr. Matthew. You are invited to ask the third question during Premier Li’s press conference. Come to the NPC media center by Saturday to get your credentials.”

I checked my email outbox — the third question was the one about Under the Dome. I read and reread the message and the email in disbelief. Could this possibly be right? By this point, the film and the discussion of it had been thoroughly scrubbed from all sites, except for Caixin, a business news site with ties to China’s political elite.

Unsure of what to think, I translated my question into Chinese and began rehearsing. I asked two former Chinese teachers to go over the question with me before the press conference on Sunday. I would be directly addressing one of the most powerful men in the world on live national TV, and I didn’t want to fuck it up.

On Saturday, I picked up my credentials, then met my teachers at a café. They tweaked my grammar on fuel-economy standards and told me where to pause. They were equally baffled that my question was approved, but wished me luck.

I went straight from the café to meet some friends in the hutongs. Spring was creeping into Beijing after a frigid and polluted winter. We celebrated with a healthy amount of drinking.

The pants

When my alarm rang at 6:30 a.m. the next morning, my body begged to stay put. I mind-over-mattered my way out of bed, and opened my closet. I didn’t own a suit, so I grabbed a wrinkled blue button-down shirt and dark khaki pants out of the closet. I gathered the essentials — notebook, keys, wallet, phone — and threw on my dad’s worn leather jacket.

When I directed the cabby to the Great Hall of the People, he asked if I would be at the press conference. I told him I would, and that I might even ask a question. He was thrilled. I was not.

All week, I had nursed a lingering fear that there had been a miscommunication. They must have meant a different question. Or meant to send that to a different journalist. If I had misunderstood Beth, the ministry would think that I switched the question on purpose. Whether or not I was right about the factional politics, the leadership took pains not to flaunt divisions in the party-state.

Puzzling out reasons for allowing the question, all I could come up with was that it gave Li a platform to endorse the film and throw PetroChina under the bus. Zhou Yongkang was already behind bars and corruption investigations had wreaked havoc on the company. Maybe the question was a chance for Li to essentially announce the execution of PetroChina under the guise of environmental stewardship.

Whatever was going on, these were all heavy hitters, and I feared my visa would become collateral damage.

The towering white pillars of the Great Hall of the People flanked Tiananmen Square to the west. Once past security, I roamed the dome-ceilinged rooms and found my way to the hall reserved for the press conference. At the front, an elevated platform contained a desk for Premier Li and his colleagues. Dozens of rows of red velvet-covered chairs were arranged facing the stage, and behind them stood hundreds of video teams and photographers with absurdly large lenses.

I found my seat on the right side, put down my jacket and notebook, and started pacing around the room. I repeated the question under my breath and people-watched for my journalism heroes. Going over the Chinese names of PetroChina and Sinopec for the tenth time — I always mixed the two up — I caught sight of someone waving at me. Squinting through the crowd, I saw it was Beth’s boss at the Ministry of Foreign Affairs. She was instrumental in getting my visa through, and had always been friendly.

I smiled and started walking in her direction, my right hand absent-mindedly fiddling with the button on my back pocket. And then I felt it. A loose thread, a frayed edge. Walking toward the woman, my fingers frantically traced the fault line.

My smile turned to plastic and my eyes grew wide as it dawned on me. My pants were sporting a Grade-A tear, maybe six inches north to south. My back pocket was hanging out, and my gray Costco boxers were on display for my colleagues and the cameras.

I greeted Beth’s boss, but my mind was retracing my steps across the hall to try to keep track of everyone I had just flashed. I nodded my way through our short conversation. The only thing I processed was her final reminder: “Don’t forget to raise your hand high during the questions so that you’ll get picked!”

I clasped my two hands together behind my back to cover the hole and made a beeline for the bathroom. Partway there, I was intercepted by Beth herself. She was excited for the presser, and stuck her right hand out to shake mine. Pulling my right hand out from behind my back meant exposing the heart of the tear. To make up ground with my left hand, I was forced to dip my left shoulder, arch my spine, and lean back slightly, all while extending my right hand forward. It looked more like a Twister move than a handshake.

My mind blotted out the conversation, then I booked it for the back wall, sliding along sideways until I got into the bathroom. Once in a stall, I assessed the damage. The tear stretched from the top of my right pocket down to my upper thigh. It was big enough to put my fist through. I remembered where it came from. Earlier in the week, I had found myself in a deep glute stretch outside of a subway station, really leaning into it with my foot on an elevated ramp. I’d heard a pop and felt a little bust in the seam. But when I got home, I just tossed the pants right back into the closet.

Out in the hall, I walked quickly and avoided eye contact. Back at my seat, I grabbed my dad’s leather jacket and strapped it around my waist, covering my backside down to my knees. The worn-down leather kept slipping, so I double-knotted it above my belly button.

The ambient chatter swelled, and then I saw Premier Li Keqiang stride onto the stage, waving to the cameras. After some introductory remarks, the questioning began.

I was secretly hoping they had forgotten about me. I didn’t know what I was more nervous about: drawing a total blank when addressing the premier, being kicked out of the country, or losing my jacket and showing millions of people my underwear.

When the moderator solicited new questions, the Chinese journalists all stood up and waved, and I followed suit. But each time I sat back down, I could hear the tear chewing up new fabric. It had expanded past the seam and entered the open waters of the pant leg. If it kept heading south, it would soon be too big for my leather fig leaf.

Finally, while waggling my notebook above the crowd, I heard my name announced over the loudspeaker. The crowd around me sank into their seats and I was handed a microphone. I took a deep breath and launched into the question, tip-toeing through a minefield of potential mispronunciations. I stumbled a bit, but managed to spit it all out, including the four key proper nouns: Chai Jing, Under the Dome, PetroChina and Sinopec.

Handing back the microphone I sat down and stared at my notebook. The translator restated the question in English, using all the key phrases. Li, almost smiling, began his answer. He talked about how he had ‘declared war’ on pollution at a previous NPC. He said that despite government efforts, the results hadn’t lived up to the people’s expectations. He said the government will strictly enforce the new environmental policy, and that they will hunt down all those who interfere with or violate these policies.

It was tough talk, but to my ears it was no different than the tough talk of years past. Most importantly, in his lengthy answer he never mentioned Chai Jing, the documentary, or the oil companies. As soon as Li finished, an Australian reporter in the row in front of me turned around:

“Did you change your question? Was what you asked approved?”

I nodded. The presser continued, but my mind was tying itself in knots.

What just happened? Why did they want me to ask that question if Premier Li wasn’t really going to answer it? Had I missed something in what he said, or was it possible this was all a mistake?

When the press conference adjourned, a small gaggle of Chinese and foreign reporters surrounded me, asking me about the question: Was it approved? What did I think of Li’s response? My answers were truthful but intentionally vague. A reporter from the People’s Daily asked to take a picture of me holding the logo of the paper’s Weibo account, but I declined.

I excused myself and headed for the exits. On the way out I ran into Beth, and she told me I did well. Descending the white stone steps outside the Great Hall of the People, I looked at my phone. It had been blowing up for the past hour, with friends from around the country sending me pictures and videos of my face on their TV screens. I waved down a cab and cruised home.

Up in my apartment, I finally untied the leather jacket and inspected the damage. The tear had cut its way north to my belt line and south to the back of my knee. I bent over, maneuvered my phone behind my back and snapped a photo. I tweeted the picture out and described my wardrobe malfunction.

Within a minute, veteran New York Times reporter Chris Buckley replied on Twitter: “You’ll regret sharing this with the world.”

Looking out my window onto Beijing’s Chaoyang District, I was relieved. Between the language, the tear, and the subject matter, I felt like I had just skipped through a minefield unharmed. But there was a nagging sensation, like when you hear all of a joke except the punchline. I believed the significance of the question lay in the proper nouns — the naming of Chai Jing and her nemeses — but Li had ignored those and just regurgitated officialese.

That evening I struggled to write up an article on the event. The New York Times already had a piece out that mentioned my question, but I couldn’t figure out whether anything meaningful actually happened. I eventually gave up, stranding a few paragraphs in a Word document.

The next day my editor in New York was baffled as to why I hadn’t published anything. She (correctly) told me to put something together quickly. I was procrastinating, flicking through Twitter, when a tweet from the People’s Daily caught my eye:

I sprang to attention, searching Baidu for mention of the case. On the website of the Party’s corruption watchdog, I found a short notice — just 34 characters — announcing the investigation of Liao Yongyuan, the general manager of CNPC. Liao had been nicknamed CNPC’s “Northwest Tiger” for his exploits in the oil fields of China’s far western deserts. He was now likely in shackles at an undisclosed location.

There it was:. Li’s actual answer to my question.

He had said nothing direct because he didn’t need to. A foreign reporter asking the question provided the perfect veneer of deniability for the administration. No one could accuse Li of endorsing the film or throwing PetroChina under the bus, but those words were still broadcast on national TV.

Xi and Co. placed the horse’s head in the bed, but let someone else be the messenger.

I sat down to write.

Two years later

Exactly two years have passed since that press conference.

Protests have continued on and off in Tianmu Village. My contact continued to post about the protests on WeChat, and he once sent me a picture of a man’s arm with a deep gash. He said thugs had come to his house and slashed one family member with a knife.

Mu Xiangyou is still the party chief of Tianmu Village, but he hasn’t shown up in accounts of NPC delegates since 2015.

After releasing Under the Dome, Chai Jing disappeared from the public eye. Later that year Time magazine named her one of the world’s 100 Most Influential People, but she hasn’t been seen or heard from since before the press conference.5

I had no problems renewing my visa, but I moved back to California last year, in the middle of liang hui 2016.

This January, Liao Yongyuan, the “northwest tiger” of CNPC, was sentenced to fifteen years in prison. Court records accused him of accepting 13,394,101.73 RMB in bribes.

On Wednesday, Li Keqiang will once again hold his annual press conference at The Great Hall of the People.6

Note from today: I hope you enjoyed this little relic of an earlier time. In future posts, we’ll be back to regularly scheduled programming.

Subscribe now

Today I’m writing to share a new piece on China’s draft regulation on AI companions, and also to sketch out two hypotheses and a research agenda. The piece, co-authored with the excellent Scott Singer, just came out on the Carnegie Endowment’s website: China Is Worried About AI Companions. Here’s What It’s Doing About Them.

It covers China’s draft regulation on “anthropomorphic interactive AI.” You can find a full translation from China Law Translate here. The regulation itself is ambiguous on what exactly counts as anthropomorphic AI: just “AI companions” or a much broader range of general-purpose chatbots? (More on this below.)

But it’s pretty clear on the harms it aims to reduce: addiction and self-harm due to extensive emotional interactions with AI. This is a significant broadening of the scope in Chinese AI regulation, moving away from the focus on content/information control that dominated early regulations from the Cyberspace Administration of China (CAC). It also moves directly in parallel to state laws from California and New York that aim at reducing these same harms.

The piece from Scott and me does a few things:

1. Breaks down what’s in the draft regulation and what’s likely to change.

2. Explains the growing role of technical standards and gives a preview of what might become the key standard for this regulation.

3. Compares the Chinese approach to regulation with state laws in California and New York, pointing out similarities and also a pretty core difference in approach.

Below I’ll give a TLDR for #1, followed by a hypothesis each for #2 and #3.

Subscribe now

Key Requirements of the Anthropomorphic AI Regulation

Operational requirements:

• Companies can’t train on data gathered during conversations unless users opt-in to allowing it.

  • Big change from the existing “opt-out” status-quo. I think this provision is likely to be scoped down or eliminated in the final version.

Addiction mitigations:

• Monitor users’ emotional state for “extreme emotions or addiction”

• Periodic reminders to users that they’re interacting with an AI (not a human), periodic reminders to take a break

High-risk conversations / self-harm mitigations:

• Template responses for various high-risk situations, comforting users and providing resources for professional help

• If a user clearly brings up self-harm, there must be a manual takeover of the conversation by a human

  • Another provision that I think is likely to change

Minors and elderly users:

• Must add a guardian or an emergency contact upon registration, and that person will be notified if the user brings up self-harm

• For minors using AI companions, guardians have a number of parental controls, including limiting the characters available and reviewing summaries of chats

• For the elderly, a prohibition on the AI imitating family members

There’s lots more than just this TLDR, so check out the piece for that.

Two Hypotheses and a Research Agenda

Here I’ll lay out a couple hypotheses and a preview of my upcoming research agenda.

Hypothesis #1: Technical standards are becoming increasingly central to how China regulates AI. These standards—particularly those put out by TC260—will end up being the key determinants of how Chinese frontier AI models are tested and evaluated.

Hypothesis #2: Chinese and U.S. AI regulations are likely to take fundamentally different approaches to mitigating AI risks. U.S. laws will rely largely on public transparency, liability and the threat of lawsuits — essentially, banking on corporate self-interest and giving companies wide latitude in how to avoid or reduce those harms. The Chinese regulations (and technical standards) will be far more prescriptive in mandating specific technical interventions, they will focus on government-facing transparency, and will be enforced through central oversight.

Here I’ll draw from the piece to give a bit more context on each hypothesis, and lay out where this is leading me.

The Role of Technical Standards in the Anthropomorphic AI Regulation

Perhaps the key open question with the draft regulation is what products or model behaviors fall within the scope. The definition for “anthropomorphic AI” could easily be interpreted to cover virtually all user-facing chatbots, though that seems pretty unlikely. But if the regulations only cover products marketed as “AI companions,” that would seem far too narrow. Some of the worst incidents so far have emerged out of interactions with general-purpose chatbots like ChatGPT.

From the piece:

So how will regulators find a middle ground between solely focusing on AI companions and regulating all chatbots? One potential answer is that the regulation won’t be scoped to target a product category, but rather a type of behavior displayed by different AI products. Doing that would require creating precise definitions of when an interaction transitions from a standard, unemotional dialogue into an “emotional interaction” covered by the rules. That’s a technically ambitious approach, and one that will rely on China’s new favorite tool for AI regulation: technical standards.

How Standards are Shaping Chinese AI Regulations

Chinese authorities have long used technical standards to complement and further specify the requirements in regulations. But over the past three years, standards have taken on an increasingly central role in how China regulates AI. …

The generative AI regulation included many high-level demands on the outputs of models (“Uphold the Core Socialist Values”) but without any concrete benchmarks or testing procedures for companies to follow. Those came several months later in a document issued by the CAC’s technical standards group, TC260, which laid out exactly how to test training data and model outputs for illegal content. When the CAC released a subsequent regulation on the labeling of AI-generated content, the national standard [really, a whole basket of standards] spelling out the technical details of these labels was already prepared. Both the regulation and the national standard officially went into force on September 1, 2025.

China’s AI policy community celebrated this synchronization of regulation and standardization as a “pioneering breakthrough,” making regulations more practical and enforceable. “China’s governance of generative AI has officially transitioned from principle-based norms to a new stage of refined regulation,” wrote legal scholar Zhang Linghan in an officially sanctioned commentary on the regulation.

My bet is that this is the direction of travel for AI regulation in China. Going forward, technical standards that accompany regulations (or some standards not tied to any regulation) will end up containing the most concrete and actionable requirements that shape model development. Some of these will be focused on testing and evaluation (as with the generative AI standard), while others will be technical specifications for the models and accompanying systems. Some of these will be “recommended” standards, while others will be mandatory. But from what I’ve heard, most companies treat TC260’s “recommendations” on how to test a model as essentially mandatory — if the regulators give you their playbook for compliance, why try to create your own?

For the anthropomorphic AI regulation, TC260 has already directed one of its working groups to begin drafting an accompanying standard with the “basic safety/security requirements” for the technology. We don’t have any more detail on that standard yet, but a few weeks ago I found an article proposing a three-part test for when a conversation with a chatbot becomes an “emotional interaction” covered by the regulation. From our piece:

We have already seen one detailed proposal for this laid out by Hong Yanqing, an influential Chinese legal scholar who is a member of TC260 and regularly participates in standards drafting. In the proposal, Hong laid out three criteria. First, the AI has an emotion recognition capability that is currently active. Second, the results of that emotion recognition affect the model’s outputs. And third, the model exhibits a “structured empathetic style.” When an AI system meets two of the three criteria, the dialogue would be categorized as an “emotional interaction” and subject to the regulation. The piece goes on to describe in detail the technical underpinnings, legal logic, and potential pitfalls of this system.

I’ve spent the past several years mapping out in painstaking detail how Chinese AI regulations get created (e.g. the chart below and this report). A big motivation for that work was to identify key actors, institutions and processes that might be early signs of where future regulation is heading. So when I get to watch that process in real time — and maybe even get an early glimpse of what’s coming — I get a real kick out of it.

One key part of my research agenda going forward will be doing a version of what I did for the AI regulatory process for the AI standardization process. For now that will focus on Chinese domestic standards for AI. Years ago I dug into on how China interacts with international standards bodies, including a deep dive on the formulation of a specific AV standard. But when it comes to really shaping how Chinese AI models are trained and tested today, I think the action is in domestic standards. I’m currently working on a piece breaking down China’s standards for labeling AI-generated content, so “watch this space..”

Side note: the topic of Chinese AI standards is way bigger and more sprawling than I could ever cover in my own research, so I’d love to see more people jumping in and tackling pieces they’re interested in.

Differing Approaches: Liability vs. Technical Mandates

For hypothesis #2, I’ll be brief:

But there are also fundamental differences that reflect the divergent political systems. California wrote some explicit technical requirements into the law, but on self-harm issues, it relies largely on corporate transparency and an individual’s private right of action that empowers users to sue the companies for damages. This gives companies latitude in how they choose to mitigate these risks, knowing they will be liable for damages if something goes wrong. In contrast, China’s regulation is more prescriptive in the mitigations required, and it will likely become more so when accompanying technical standards are released. China’s enforcement will be top-down, primarily driven by administrative oversight by the CAC rather than lawsuits and legal precedent. …

Neither China’s top-down intervention-based model nor California’s statutory liability-based model has been tested at scale. Across the wide array of potential AI risks, it may be that a liability-based approach works best for ongoing harms to individuals, while a more technically prescriptive approach works for larger-scale threats.

There are lots of caveats to be made here (particularly if the CAISI and NIST take on a greater role in shaping / enforcing binding AI regulation in the U.S.), but it’s late and I’ve got to catch a very early flight. One of my goals this year is to not let my procrastinating pedantry get in the way of publishing, so I’ll leave it here for now.

If you’ve got thoughts, pushback, additions or questions on any of this stuff, I’d love to hear from you.

Best,

Matt

Welcome to my first Substack post. A quick introduction to me: I’m a senior fellow at the Carnegie Endowment for International Peace, where my research focuses primarily on Chinese AI governance, specifically how the AI policymaking process works over there. (An overview and a deeper dive on that). Before that I was a journalist in China and wrote a book about California-China ties.

I’m not sure exactly how I’ll use this Substack — probably a bit of experimentation in terms of form and content — but I’ll primarily be covering China, AI, and related odds and ends.

[Note after I’ve finished writing this piece: I initially intended this to be a pretty short sketch of an argument, but as I got going it grew into something much more in depth. It gets into big questions on the trajectory of Chinese AI development, so I think/hope it’s worth your time. But if you’re not looking for that kind of a longread every time I’d still encourage you to subscribe — subsequent pieces will take different formats and likely be more concise. Famous last words for a writer, I know…]

OK - let’s dig in.

On August 26th, China’s State Council released a much-anticipated policy document on how to implement the national “AI+” initiative. In terms of Chinese AI policy documents, this is quite a big deal. I’d rank it below the 2017 national AI plan and the 2023 generative AI regulation, but above a raft of other regulations and standards that have come out over the past ~8 years (my incomplete catalogue of those here). As with the 2017 AI plan, this is also likely to set off a chain reaction of sector-specific subsidiary “AI+” policy documents, such as last week’s “AI+Energy” policy.

Below I’ll explain:

1. Goals and structure of the AI+ policy

   1. And what the hell does “AI+” mean?

2. The bearish case against AI+

   1. Cash-strapped local governments, slumping private investment, difficulty of integration.

3. The bullish case for AI+

   1. “Policy as permission”, (training) data unleashed, AI+Science

4. Conclusion

   1. Where I land, tech booms as a driver vs. product of economic booms

Links to the text: I’ve had Claude do a machine translation and side-by-side comparison with the original Chinese. I haven’t checked the translation line by line, but what I did check looked good. You can find the original Chinese document here.

1. What is the AI+ plan and how does it work?

Overall framework:

The AI+ plan aims to turbocharge the diffusion and integration of AI applications throughout China’s economy, society and governance structures. In literal terms, the “AI+” refers to “AI+[a specific industry or sector]” — e.g. “AI+manufacturing” for integrating AI into Chinese factories, or “AI+healthcare” for integrating AI into the healthcare system. This “[Technology]+[Sector]” formulation originates from the “Internet+” plan of 2015, which I’ll describe more below.

The “big idea” articulated in the policy is that the deep integration of AI applications will re-structure China’s economy and society in fundamental ways, and this is the Chinese government’s roadmap for guiding and accelerating that integration. The really-quite-modest goals of the plan are using AI integration to:

“reshape the paradigm of human production and life, promote a revolutionary leap in productive forces and a profound transformation of relations of production, and accelerate the formation of a new type of intelligent economy and intelligent society.”

This all dovetails nicely with Xi Jinping’s big push for “new quality productive forces” (新质生产力) to become the driver of China’s next decade of growth. In a nutshell, “new quality productive forces” is shorthand for using technological innovation to upgrade Total Factor Productivity in China’s vast industrial economy.

Notably, the document doesn’t cite AGI as the driver of these changes, and it pays relatively little attention to frontier AI development in general. Instead, it’s focused on how the AI of today-ish can be leveraged to achieve the CCP’s economic, social and political goals. Analysts such as Julian Gewirtz have reasonably questioned whether this public focus on applications might be a smoke screen for the CCP’s actual focus on AGI. That can’t be ruled out and it’s possible a considerable amount of resources are being devoted to such a program (I’m sure our intelligence agencies are monitoring for it). But at this time, a plan like this will be the Party’s largest push to mobilize China’s scarce AI-related resources (money, chips, talent, bureaucratic energy) in the coming years, and directing that push towards applications is likely a sign that these remain the Party’s top priority.

As one Chinese security scholar working on AI issues phrased it on WeChat, “The next stage of generative AI isn’t AGI. It’s application scenarios!”

One final big picture note before digging into specifics. This is the highest-level document I’ve seen that directly addresses unemployment risks from AI:

“Strengthen employment risk assessments of AI applications, guide innovation resources toward areas with high employment potential, and reduce impacts on employment.”

This is just one sentence in a long document aimed at turbocharging AI deployment. But I think it’s a very significant one. Up until quite recently, unemployment concerns were not a major topic in Chinese policy discussion around AI — in informal polls of policy scholars it ranked quite low on a list of potential risks. But that seems to be changing. A spate of protests in Wuhan by taxi/Didi drivers over Baidu’s robotaxis drew national attention, and I’ve heard rumors that the NDRC (the economic planning agency in charge of AI+) is growing increasingly concerned about AI-induced unemployment, including commissioning a major study on the issue. This isn’t brand new — when I was in China in 2017, a researcher with the State Council’s main think tank told me about a study they were conducting on this. But both the technological and economic forces here have changed dramatically in the years since, and I’ll be watching this space closely going forward.

Ok, back to the policy document itself.

Targeted sectors:

The plan selects six priority areas for this integration:

1. AI+Science and Technology

2. AI+Industrial Development

3. AI+Consumption Upgrading

4. AI+People's Livelihood and Well-being

5. AI+Governance Capacity

6. AI+Global Cooperation

Each of those priority areas then has a number of more specific focus areas, some smart and some... questionable. For example, the S&T bucket includes a call to “accelerate the construction and application of scientific large models”… along with applying AI within philosophy research. The People’s Livelihood bucket includes calls to increase AI-assisted medical diagnoses… along with AI for “strengthening interpersonal connections, providing spiritual comfort and companionship” and more.

Aside from these priority sectors, the plan also includes several sections with (pretty vague) suggestions for improving China’s deployment of data resources, compute capacity, talent, etc. to support AI diffusion. These largely rehash existing government policies / goals, but I’ll flag some potentially new ideas and language below.

Milestones and metrics:

The policy sets ambitious goals for AI adoption in these areas, as well as broader societal impact:

2027: The penetration rate of AI devices, agents and applications will exceed 70%;1 “the role of AI in public governance will be significantly enhanced; and the AI open cooperation system will be continuously improved.”

2030: AI will “comprehensively empower high-quality development” and the penetration rate will exceed 90%, with the intelligent economy becoming an important “growth pole” for China’s economic development.

2035: China will “enter a new stage of developing an intelligent economy and intelligent society” that supports the CCP’s existing big-picture goal of “basically realizing socialist modernization” by that year.

Including specific metrics and timelines is nice, and these years + targets are guaranteed to be used by U.S. politicians pointing to the fact that China has a “plan” to use AI to dominate us in all of these sectors.

But when analyzing these high-level policy documents, the specific targets are often written vaguely enough to defy precise measurement. That’s because it’s not the targets themselves that matter for policy execution. It’s the signal they send, and the reaction they spark from actors across the CCP, state and wider economy.

How these plans typically work: “Call-and-Response” policies

The AI+ initiative follows in a tradition of Chinese tech stimulus policies that aim to accelerate the development and deployment of a new technology. Some examples of this include the 2015 “Internet+” campaign (the namesake for today’s AI+ policy) and the 2017 national AI plan.

These policy docs aren’t like pieces of legislation in the U.S. that prescribe exactly what will be done, by whom, under what authority, and with what level of appropriations. They are closer to Executive Orders, if EOs were vaguer and could galvanize action by all state and local officials, along with businesses, venture capitalists, university presidents… even Chinese parents deciding what their kid should study or what job they should take after college.

The documents often begin by announcing grandiose goals for the technology (“reshape the paradigm of human production and life”) and then rattle off hundreds of things that might contribute to that goal: specific applications, various types of incentives, and lots of nice things to “promote” and “encourage.” They make for strange and overwhelming reading if you take them too literally.

Essentially, these documents operate like a “call-and-response” between the top leadership and ministries + local officials / businesses around the country. The high-level policy document sends a clear signal to all Chinese officials (+private companies/investors and society writ large) that this technology is a priority for the top CCP leadership, and they will be rewarded for doing something, anything (see below) to advance it. If they do, officials can expect better performance marks and faster promotion. And historically, private companies and investors will rush into the sector, aiming to ride the wave of subsidies and preferential policies to windfall profits.

Here is a small sampling of these types of actions from a 2018 piece I wrote about how local officials were reacting to the 2017 National AI Plan. (The links are now broken, but all of these were all publicly announced projects at the time.)

Are you in charge of transportation for the new megacity of Xiong’an? Partner with Baidu’s self-driving project, Apollo, to demonstrate autonomous vehicles in the city. Head of the Changping branch of Beijing’s Public Security Bureau? Spend 2.75 million yuan ($437,000) procuring AI person-tracking software for security cameras. President of a mid-tier engineering university in Shandong? Open the province’s first AI research center focusing on medical and marine AI. Party chief of a Nanjing economic development zone? Pour 8 billion yuan ($1.3 billion) into an AI-focused venture capital fund and dole out 5 million yuan ($794,000) in R&D subsidies to each firm that sets up shop there.

This approach floods the zone of the targeted sector with a wide range of physical, financial, human and technical resources. Basically, it revs up the engine (or “primes the pump”) of the industry and associated research. Looked at individually, many of these actions are clearly wasteful — lots of technological “bridges to nowhere.” But if a technology is general-purpose enough, and if it’s economic upside is large enough, then in the aggregate these campaigns can pay off by turbocharging development and adoption.

Has it worked in the past? It’s impossible to fully disaggregate the impact of a policy from the existing industrial/technological momentum. And sometimes these policies are just a way for the gov officials to claim credit for what was happening anyway.

But for at least the 2015 Internet+ campaign and the 2017 National AI Plan, I watched them play out up close, and I talked with entrepreneurs, investors and students about them at the time. I believe they had a real material impact in turbocharging these fields. At the time, local governments around the country had the cash to throw at all kinds of projects, and the government had the ability to galvanize the private sector (and society writ large) to rush headlong into these industries in droves.

In the years following Internet+, we saw a tidal wave of new internet startups and applications that drove actual economic productivity and government efficiency. And in the wake of the national AI Plan of 2017 we saw a similar surge of startups, along with a major increase in the academic and research base of the industry.

2. The Bearish cases against AI+: Money, Trust and Integration Friction

In short, (1) China’s local governments don’t have the cash to run the typical playbook, (2) China’s VC / tech investment sector has been decimated and is less likely to rush into the new sector due to broken trust with government, and (3) deeply integrating AI into traditional sectors is simply far harder and more labor intensive than the relatively lightweight integrations of the “Internet+” era, such as mobile payments and online orders.

Historically, “Call and response” policies worked well when two conditions were met:

1. There was a ton of capital sloshing around the economy (both private and public coffers).2

2. Private investors and companies strongly followed the government’s lead because they believed there was lots of money to be made in the new industry-of-choice.

But this time around I don’t think those conditions hold.

I’ll start with government cash, and I’ll be relatively brief here. For local and provincial Chinese officials to carry out their “response” in the “call and response” policy cycle, they usually need to spend money (or forego the opportunity to collect it in taxes). There are some low- or no-cost policy actions that can be taken (and I’ll describe some for AI+ in the bullish case), but many of the actions in the typical playbook require significant financial outlays.

And Chinese local governments just don’t appear to have the surplus of cash on hand. That’s largely due to declining revenue from selling of public land (which basically printed money for local governments during the multi-decade housing boom), and largely stagnant tax revenues since 2021. See the below chart on declining fiscal revenues from the Rhodium Group, particularly the “% of GDP” line, which has gone from ~25% in 2015 to a projected ~19% this year.

This has effectively forced local governments around China to implement a variety of informal austerity measures (like not paying their employees and contractors) and illicit revenue generating schemes (like extorting businesses and rich people into paying new fines or alleged back taxes). These practices are not systematic across China, and we only get glimpses of what’s happening, but the picture that emerges is of local governments — and public institutions like hospitals and universities — who are really struggling to pay the bills.

For AI+ to really work you need these institutions to be ready to dole out cash on all kinds of related projects, but that doesn’t appear to be the case. Without the money to subsidize, procure, and pilot AI+ projects, you’re left with a lot of exhortations but no new incentives.

Ok, if local governments / public institutions can’t put down the money themselves, they can surely still galvanize (or force) the private sector into investing, right? I’ve got my doubts.

Chinese private investment in AI companies 2012-2025 (YTD)

The data comes from ITJuzi.com, one of the better providers of data on Chinese VC and tech funding (though none are perfect). The orange bars represent deal count while the grey line represents capital invested. This screenshot is missing it’s left axis (and the right is in Chinese) so I’ll pull some numbers.

Both deals and total capital invested peaked in 2021 at 974 investments worth approximately $25 billion USD. In 2024 those numbers were down dramatically to 510 investments worth $8.6 billion USD. In 2025 we’re on pace to substantially increase the deal count (already at 507 in early Sept) but to roughly equal the capital invested in 2024 (currently at $5.7 billion). Just looking at the direction of the lines, you’d think that we were in the midst of an AI winter. (And no, the DeepSeek moment from earlier this year didn’t set off an ever-mounting wave of VC investment — May through July saw a slight uptick but then back to previous levels in August.)

Contrast that downward arrow in China over the past three years with:

Global (non-China) investment in AI companies 2012-2025 (YTD)

Again, that grey line going steeply up is the total capital invested, reaching around $100 billion USD in 2025 YTD. Very different trend line. Data on investment totals can be all over the place, so for context here’s a comparable chart from the 2025 AI Index Report by Stanford HAI, pulling from different sources.

The bottom line is that in the nearly 3 years since the debut of ChatGPT, funding for AI companies in China has decreased. This is a reflection of the fact that the venture capital industry in China has been essentially decimated. For a good read on that decimation, check out this April 2024 piece from Robert Wu at Baiguan. It could be that the industry has now bottomed out and is ready to rise, but for various structural reasons (exit of USD funds, end of the first-ever 7-10 year cycle for lots of funds born in the 2010s, etc.) I don’t think it will be getting back to its 2012-2018 heyday anytime soon.

Aside from the specifics trendlines on funding, I think there’s a deeper issue at stake: lots of private investors and companies felt severely burned by erratic and largely unfriendly policymaking from 2018-2023, and they’re not in a hurry to rush back into whatever the government tells them is the next big thing.

Beginning in mid-2023 the Party has attempted to publicly re-embrace private companies (including the platform tech giants they spent the previous few years beating up), but that hasn’t turned investor or consumer sentiment around the way these governmental public displays of affection did in years past. It seems that trust has been broken in a way that’s not so easy to repair.

That hesitance undermines a key — maybe the key — mechanism driving the successful tech stimulus plans of the past. Yes, direct subsidies can be helpful, and government procurement + the opportunity to pilot certain technologies gives lots of startups some helpful runway to achieve liftoff. But those things also tend to come with a fair amount of paperwork and bureaucratic entanglement. Traditionally, the best startups in China try to avoid these government entanglements. Instead, it was the way major policies galvanized private capital and entrepreneurial energy that drove the technology forward.

Finally, beyond the financial and business sentiment issues, I think the complexity of integrating AI into most industries will pose a significant obstacle. During the Internet+ era, millions of restaurants and nail salons and drivers were pretty easily able to plug into a user-friendly internet platform (WeChat, Meituan, Didi, etc.) and both take orders and receive payments on there. That relatively lightweight, low-complexity connection was enough to generate a real economic boost for many small businesses, and to even upend whole industries. Some integrations were more extensive and complex, such as the way local governments used WeChat as a portal to accessing public services, a linkage that removed a whole lot of headaches for regular people used to long lines and lots of 没办法’s (“whattayagonnado”). But when compared with the complexity of integrating AI (whether LLMs or vision models) into industrial and corporate workflows, that all looks relatively simple.

In China, this move from mere “connection” (internet) to “empowerment” (AI) is being discussed as a reason why AI+ will be so transformative. Patrick Zhang describes this well in his fantastic Geopolitechs newsletter on AI+:

According to some experts, “Internet Plus” was about “connection” (连接)—linking information to redesign processes and improve efficiency. “AI Plus” builds on that by adding cognitive ability, moving from “information connection and diffusion” to “knowledge application and creation.” Its essence is “empower”(赋能). It promises to reorganize production factors, upgrade value creation models, transform organizational structures, and reshape governance.

I think they’re right about the potential economic upside over the medium- to long-term. But that reorganizing of production factors and transforming of organizational structures adds many, many degrees of difficulty to the process. Deeply integrating an LLM into the guts of a corporation, or intelligent robots into a factory’s production process, is a complex and friction-filled process. It will likely require either very sophisticated in-house technical talent, or lots of highly specialized startups that build unique products and do the leg-work of deep integration.

But both of those paths are rocky right now, and likely will be for the foreseeable future. Specialized startups need the kind of funding that just doesn’t appear to be there right now. And the companies that the CCP has the easiest time mobilizing are state-owned enterprises (SOEs), a group not famous for either their in-house technical talent or their nimbleness with new technology. Much was made of how SOEs and local governments rushed to adopt DeepSeek in the months after its release, but see the ChinAI newsletter from just a few months later for a reality check on that adoption (“shallow, narrow, and slow”).

Taking all of these things together, we have an extraordinarily ambitious tech stimulus plan — one that might even be very far-sighted in how it sees AI transforming economies and societies — but one that lacks the traditional levers or incentives that have made such plans so impactful over the past decade.

3. The Bullish Case for AI+: Policy as permission, data unleashed, and AI+Science

I’ll start the bullish case with some rebuttals of key points in the bearish case. Given how successful China has been at climbing the technological value chain (with smart policies providing key nudges along the way), you could argue that the burden of proof is more on the bears than the bulls here.

• Local government funding shortages won’t undercut implementation:

  • The cash shortage isn’t equally distributed across Chinese local governments, and generally speaking the richer and more technologically advanced cities and provinces (Beijing, Shanghai, Guangdong, Shenzhen) are doing better. These are the places where AI+ will really launch from: Beijing, Shanghai and Hangzhou are the hubs for software-focused AI startups; Shenzhen is the global center of hardware innovation; and Guangdong provides a vast industrial playground for experimentation with AI integration. Local and national support for AI+ experimentation will work out the kinks and demonstrate the business case for this integration, at which point the private sector will naturally move in and take the campaign nationwide.

• Private investment in AI is finally ready to rebound:

  • Yes, it’s been a rough ~3 years for private investment in AI, especially at the VC stage. But that was due to some extraordinary circumstances: the big tech crackdown (2020-2022/3), the huge Covid lockdowns of 2022, broad uncertainty about the economy coming out of Covid, and some very mixed signals from the Party-state when it came to AI. The government’s first reaction to the post-ChatGPT generative AI boom was extreme concern over the implications for online speech and information controls.

  • The CAC basically blocked the wide release of LLMs for the first 8 months of 2023, and then was still making companies wait months for a (de facto) license for these. It wasn’t until 2024 that the CAC began shifting from de-facto-licensing toward more streamlined registration of models, and it wasn’t until the DeepSeek moment of early this year that we really saw the government full-on embracing these companies. The AI+ plan is now the clearest and the highest-level signal that the state is fully embracing AI as a major economic force. It may take a minute for that to translate into concrete investments, but there is tons of pent-up capital that is ready to surge into these fields now that it knows where to go. Add to this the “quiet bull” in China’s stock markets and we are seeing many more realistic exit options for tech investors than ever before.

• AI integration is hard, but no country is better positioned for this than China:

  • China’s vast industrial economy, paired with its unmatched hardware innovation ecosystem, will make it an absolute playground for the integration of robotics and intelligent automation into industrial processes. We’re already seeing the early stages of this with the boom in robotics startup valuations. Yes, this will be a process full of friction and failure, but that’s true everywhere and China has the benefit of close proximity between lab and factory floor. With the rise of China’s new generation of robotics startups, we will see an extremely tight cycle of iteration and experimentation. (Side note: I highly recommend subscribing to the Substack linked in the previous sentence, Hello China Tech by Poe Zhao. It’s like Stratechery for China tech and hugely informative.)

Ok, beyond those rebuttals there are also additional positive cases to be made for the AI+ plan. I’ll make three here: (1) Policy as permission, (2) data unleashed, and (3) AI+S&T.

“Policy as permission”

This policy document might not provide the concrete funding needed to deeply integrate AI into every industry listed. But it does give entrepreneurial Chinese officials a certain amount of political cover to go out and experiment with AI-related activities without fear (ok, with a little bit less fear) that they will be punished if those experiments don’t pan out. There remains the question of how you fund these experiments, but there are some low-cost policy interventions that could prove impactful.

Just a couple examples of low-cost interventions from the policy itself (I’m sure enterprising officials will think of far, far more):

• “encourage universities to include open-source contributions in student credit recognition and teacher achievement recognition”

• “Improve data property rights and copyright systems suitable for AI development, and promote lawful and compliant openness of copyrighted content formed by publicly financed projects.”

Counting open source contributions toward university credits is pretty novel and interesting, though admittedly quite marginal within the big picture. But…

Data unleashed

The data stuff might turn out to be far more impactful. Legal decisions are nominally “free,” and I think this document sends a signal to judges nationwide that their rulings on cases involving data, copyright, fair use, and AI should be “suitable for AI development.” Chinese courts have issued a mixed bag of rulings on these issues, with some major decisions on whether training data constitutes “fair use” still pending. This document could be an important political nudge toward rulings that give AI companies a freer hand with using copyrighted works as training data. When you compare that with Anthropic’s $1.5 billion settlement in a copyright case this week, that nudge could go a long way. Add in the forced opening-up of copyrighted content that was created by public funding (an overwhelming portion of academic work, and a lot of creative work as well) and that nudge goes even further.

And it’s not just loosening up on copyrighted data. China’s recently-created National Data Administration (NDA) has a mandate to turn data into a new “factor of production” (“生产要素”), alongside the traditional factors of land, labor, capital and technology. The NDA is under the National Development and Reform Commission (NDRC), the country’s economic planner. Over the past two years the NDRC has been charged with guiding Chinese AI policymaking across the bureaucracy and corralling the different ministries. The NDRC was the actual author of the AI+ plan (which was then released by the State Council), and it quickly followed that with the “AI+Energy” policy, co-released with the National Energy Administration. In describing ways that AI can turbocharge the energy sector, that document made better sharing and utilization of energy data a core component, mentioning “data” 33 times.

If it turns out that a key component of AI diffusion involves the creation and open-sourcing of high-quality public datasets, the NDA could be a major accelerant in that process.

AI+Science and Technology

I saved this for last because (a) it might be the most important part, but (b) I don’t have a lot of useful stuff to say about it. Lots of the people who I most respect in this space think the most promising AI applications are in the “AI for science” realm. By placing this first among the six “key domains” for AI+, the Chinese leadership is likely indicating that they agree with that assessment. This week the People’s Daily published a piece by Zhou Bowen, director and chief scientist of Shanghai AI Lab, on the paradigm shift toward AI-driven scientific research (Claude machine translation here).

If this predictions of AI becoming a driving force in scientific breakthroughs pans out, and if the AI+ initiative significantly accelerates that nexus of work, then this could turn out to be a monumental decision. But to be totally frank, I don’t know whether the document will accelerate that nexus of work, because I don’t know what the main obstacles or bottlenecks to this kind of work are.

If the main bottlenecks are bureaucratic hurdles (lab or university policies that limit the use of AI in research), then this could have a major impact, unleashing a wave of wide-ranging experimentation with AI-empowered research by scientists across China. If it’s a matter of a marginal access to resources, the policy could also have a significant impact, pushing universities and institutions like the Chinese Academy of Sciences to reallocate some financial and computational resources towards this type of research.

But if the limiting factors on the AI+Science nexus are different from the above (e.g. hard technical capability limits) then it’s unclear how much impact the policy will have. I’m not the right person to answer that question, but I really welcome feedback from those who know this space well.

4. Conclusion: Where I land and what I’m watching

When I started writing this piece, I intended to just argue the bear-ish case. But in the process of doing that I found myself throwing in a bunch of caveats and “but maybe if” qualifiers. In the end, I decided better to lay out the bear and bull case in full. But I don’t want to fully duck the question with a “maybe this, maybe that” piece.

On balance, I lean toward the bear-ish case. Despite the theoretical 10-year time horizon for the AI+ initiative, I think these type of policy documents have the largest impact in the ~2 years after they’re released. And right now I think the fiscal conditions and business environment severely undercut the mechanisms by which these policy documents traditionally have impact.

If you’ll allow me one final hedge: There’s a decent chance that over the next 5-10 years China does end up successfully diffusing AI applications throughout its economy, maybe even faster than the U.S. does. But if they do succeed, it won’t be because this policy turbocharged adoption. It will be due to other factors: major recovery in the Chinese economy; technical innovations that ease integration frictions and make AI adoption far less costly; and perhaps important of all, the truly awe-inspiring tenacity of Chinese entrepreneurs who always find a way to squeeze any possible business opportunity for all that it’s worth.

In this way, I actually differ from Jeffrey Ding’s thesis on China’s “diffusion deficit” in technology: I think that over the past ~20 years China has done a pretty amazing job at continually using technological diffusion to lever itself up global value chains, from traditional manufacturing upgrades all the way through “Internet+” and even early waves of AI. But at this particular economic and political moment, I think the traditional policy playbook for turbocharging diffusion just won’t work as well as it used to. This relates to another idea that keeps nagging at me and that I hope to explore in a future post: when I look back at the 2010-2020 period, I increasingly think of China’s technological boom of that era as being as much (if not more) a product of the wider economic boom as it was a driver of it.

But I’ll save that for another time. Thanks for sticking with me (or skipping to the end). I plan to keep future posts shorter and sharper. And I welcome your feedback.

Best,

Matt

What this will be used for: Occassional emails to share new work that I’ve put out into the world.

What this will not be used for (for now): Regular dispatches containing links and my thoughts on what happend this week.

There are lots of great newsletters out there sharing news and commentary on these issues, and I’m not looking to get into that space. But I’ve learned so much by staying connected with other people interested in these same issues, and I want to make sure those connections don’t live and die by the fates of a certain other platform.

So if you’d like to stay in touch and get the occassional (once a month? once a quarter?) email with new work of mine, please do subscribe below.

A few housekeeping details and links:

Why did I get this email? Either because you signed up for my new newsletter, or because you were previously signed up for my old Chinafornia newsletter. If you don’t want to receive any more emails just unsubscribe and you won’t hear from me again.

Who are you? I’m Matt Sheehan, a Fellow in the Asia Program of the Carnegie Endowment for International Peace. I study China’s tech ecosystem, with a particular focus on artificial intelligence. My current project is dissecting how China makes AI and algorithm governance policy: what are the political, intellectual, socio-economic and international forces shaping how China governs algorithms? Here is my bio and work at Carnegie.

I’m also the author of The Transpacific Experiment: How China and California Collaborate and Compete for our Future (Counterpoint, 2019). Before that I worked as a reporter in China, and took a stab at social media stardom through videos teaching ultimate frisbee in Chinese.

Other links: My Twitter is @mattsheehan88. Lots of my earlier analysis work is on MacroPolo. Lots of my earlier reporting from China can be found on this interactive map.

Anything else? Please feel free to get in touch if you share an interest in these topics and would like to chat. My email is [first name] . [last name] at ceip.org

Subscribe now