Agile Governance: Tsinghua's Xue Lan on How China Regulates What It Can't Fully Predict

Summary

---

Sinica Podcast Summary

Agile Governance: Tsinghua's Xue Lan on How China Regulates What It Can't Fully Predict

Date: June 30, 2026
Host: Kaiser Kuo
Guest: Xue Lan (Dean of Schwarzman College, Tsinghua University; Director, Tsinghua Institute for AI International Governance; China’s National Expert Committee Chair on Governance of Next Generation AI)
Recorded at: World Economic Forum “Summer Davos,” Dalian

---

Episode Overview

This episode explores “agile governance,” a concept central to China's approach in regulating rapidly evolving technologies, particularly artificial intelligence (AI). Host Kaiser Kuo and guest Xue Lan dissect how agile governance works in practice, weigh its global relevance, and clarify Western misconceptions about Chinese regulatory culture. The conversation provides unique insights into regulation models, balancing innovation and risk, the role of provisionality, and the degree to which China's practices might be emulated elsewhere.

---

Key Discussion Points & Insights

1. The Pacing Problem & Why AI Is Different

• Definition: The inherent lag between technological innovation and the rules designed to govern it.
• AI’s Uniqueness:
  • Speed: AI evolves far faster than previous technologies. Regulation can become obsolete within months or even weeks.
  • Breadth: “You can't think of an area that AI would not touch on, right?” — Xue Lan (05:26)
• Context: Similar lags have existed for previous technologies, but AI’s rapid and pervasive change is unprecedented.

---

2. Defining Agile Governance

Xue Lan’s Three Core Elements (06:57–11:45):

• Mindset Shift:
  • Move away from slow, exhaustive processes (thorough reviews, broad stakeholder input).
  • “By the time you've gone through that process, the thing you regulate has already changed.”— Xue Lan (09:32)
  • Emphasis on “move fast, learn fast, revise fast” rather than comprehensive, once-and-for-all rules.
• Collaborative Regulator-Regulatee Relationship:
  • Traditional “cat and mouse”/adversarial dynamic replaced with partnership.
  • “It's do nobody's good if something disastrous happens.” — Xue Lan (10:44)
• Flexible, Gradualist Regulatory Tools:
  • Avoid severe measures (giant fines, shutting down companies) for honest mistakes.
  • Preference for “light touch,” nudges rather than heavy-handed punishment, except for grave, egregious risks.

---

3. China’s Regulatory Practice — How Does Agile Governance Look in Reality?

• Historical Example:
  • The 2017 Next Generation AI Development Plan: Initiated parallel tracks of technology and governance.
  • Regulatory responses often follow incidents, with adaptation over time (12:22–14:37).
• Iterative, Experimental Approach:
  • Regulation is often introduced as “interim measures.”
  • Revision and emendation are built-in; legal documents are assumed to be temporary, adaptable.
  • “Is it a feature rather than a bug? Regulate fast, learn, then revise?”
    • “Exactly. Yeah, that is indeed a feature.” — Xue Lan (17:35)
• Balance of Development and Security:
  • When risks are severe, caution prevails; otherwise, space is left for innovation with feedback loops (18:29–19:52).

Memorable Analogy:

“Crossing the river by feeling for stones.” — A reference to China’s reform philosophy since 1978, now applied to tech governance (15:48–15:49).

---

4. Regulatory Experimentation & The Chinese DNA

• Pilot Zones and Sandboxes:
  • Used for carefully designed, partial exposure to risks—pilot before scaling.
• Response to Tech Crackdowns (Ant, Didi, Tutoring Firms, etc.):
  • Not a failure of agility but the result of accumulated tensions, later used to set a regulatory foundation for the next, gentler phase of governance (24:17–26:03).
  • “[You] accumulate certain kind of tension and at some point… when that tension reached a certain breaking point … you generate response.” — Xue Lan (24:17)
• Learning by Doing:
  • The system is unfinished and iterative by design, benefiting from ambiguity and adaptability.

---

5. Judgment Under Geopolitical Competition

• National-level Competition (the “AI Race”):
  • Policymakers are aware but do not resort to total deregulation in pursuit of victory (20:35–21:18).
  • “We need to make sure AI will bring benefits…with minimum harms.”

---

6. Open Source and Open Weight AI Models

• Government Stance:
  • No top-down mandate for open vs. closed models; left to company discretion (28:34).
• Global South Strategy?:
  • While diffusion is supported, it is not an explicit “export the Chinese stack” strategy.

---

7. Agility vs. Predictability for Business Investment

• Challenge:
  • Rapid regulatory change introduces uncertainty for firms.
  • “For AI, the uncertainty is the DNA…”— Xue Lan (30:48)
  • Red lines are established on critical risks, accepting unpredictability as part of AI’s landscape (30:48–32:34).

---

8. Checks, Balances, and Rule of Law in China

• Western Notion: China’s agility is due to a lack of constraints (courts, free press, etc.).
• Xue Lan’s Counter:
  • Balance exists but manifests differently: legal recourse, public sentiment channels, multilateral input (33:11).
  • “Checks and balance at work… public sentiment can also percolate through different channels to really generate policy response.”
• Examples:
  • Facial recognition rules rolled back after public complaints (34:42–34:53).
  • COVID health apps quickly abolished (34:54).

---

9. China’s Multi-Regulator System — Obstacle or Asset?

• Historical Lag:
  • The Chinese regulatory system is still incomplete, boundaries between agencies are blurred but this plasticity helps.
  • Competition among would-be regulators brings flexibility and less “path dependence.”
  • “Being new, they don’t have path dependency, they have some flexibility.”—Kaiser Kuo (40:10–40:16).

---

10. Comparing Governance Models — EU, US, China

• EU:
  • Comprehensive, methodical, but possibly outdated or slow for fast-moving fields.
  • “The kind of AI they were concerned about is no longer the AI we're talking about today.” — Xue Lan (41:01–44:15).
• US:
  • Patchwork regulation; industry deference; robust legacy sectoral regulation at state and sector levels.
• China:
  • Modular, adaptive, no single comprehensive AI law but many foundational related laws and admin rules.
  • Preference for experimentation and iterative legal design.
• Advantage:
  • Modular rules permit quick adaptation (44:15–44:25).
  • “China’s approach… somewhere in between [EU & US].” — Xue Lan (41:01–44:15)

---

11. Is Agile Governance Exportable?

• Applicability:
  • Adaptive, “learning by doing” approach can be beneficial beyond China—even in liberal democracies.
  • “Why not?” (47:50–47:54)
  • All systems will need an adaptive, iterative component given the unpredictability of technologies like AI.

---

12. Global Public Goods, Fragmentation & Areas for Cooperation

• AI as a Global Public Risk:
  • Risks are not just “American” or “Chinese”—they are human risks (49:26).
  • Areas for global cooperation:
    • Existential risks: “If that really becomes a reality…is humanity’s risk.” — Xue Lan (49:26)
    • Non-state abuses, inclusion of global south, climate change, and public health.
• Fragmentation:
  • Governance is fragmenting into blocks, but there are significant grounds for continued cooperation (49:26–52:38).
• Examples of Collaboration:
  • Prior to 2016, Chinese and US scientists closely collaborated (52:23).

---

Notable Quotes & Timestamps

• “The thing you regulate has already changed.” — Xue Lan (09:32)
• “You really have to work together; you are partners. It's do nobody's good if something disastrous happens.” — Xue Lan (10:44)
• “China does not have the ambition to have a legal framework that governs everything... China has been trying to be adaptive and learning along the way.” — Xue Lan (15:49)
• “Revision and emendation are assumed. It will have legal force, but it is assumed this will be changed in the very, very near future.” — Kaiser Kuo (16:22)
• “[You] accumulate certain kind of tension and at some point…when that tension reached a certain breaking point…you generate response.” — Xue Lan (24:17)
• “For AI, the uncertainty is the DNA…” — Xue Lan (30:48)
• “Checks and balances are maybe the same as in the US except the way that’s been manifested is in different ways.” — Xue Lan (33:11)
• “Being new, they don't have path dependency, they have some flexibility.” — Kaiser Kuo (40:10)
• “The advantage of the Chinese approach... is being adaptive, learning by doing... I think that is something most countries cannot avoid.” — Xue Lan (46:35)
• “If AI goes out of control... that's not a U.S. risk or a Chinese risk, it’s humanity’s risk.” — Xue Lan (49:26)

---

Timestamps for Important Segments

• Pacing problem/AI’s rapidity: 05:02–06:28
• Defining Agile Governance: 06:57–11:45
• How China regulates: iterative and provisional rules: 12:22–17:35
• Development vs Security—Striking balance: 18:29–19:52
• Crackdowns and learning cycles: 24:17–26:25
• Global context, competitiveness, and investment uncertainty: 20:35–32:34
• Checks and balances, rule of law in China: 32:34–34:53
• Regulatory “incompleteness” and flexibility: 37:37–40:16
• Comparing EU/US/China models: 41:01–44:59
• Is this model exportable? 46:16–47:54
• Global fragmentation/opportunity for cooperation: 48:25–52:38

---

Conclusion

This episode offers a nuanced, insider account of how China’s regulatory culture has evolved to manage unpredictable tech like AI. It shows “agile governance” not as deregulation but as modular, iterative, administrative guidance—a living system responsive to real-world risks and feedback. Distinctions are drawn between Chinese practice and Western perceptions, and the potential for cross-system learning (and misunderstanding) is repeatedly highlighted. The approach, while emerging from China’s unique history, is presented as having potential relevance for any nation grappling with the unpredictable frontier of AI.

---

Transcript

[A]

Why does Progressive work hard for truckers? Because truckers unite the world. They unite kids with their first drum sets and parents with earplugs. But truckers can't do this if they're not on the road. That's why Progressive has over 360 heavy truck employees to help truckers stay on time and on track. Quote Truck insurance today in as little as eight [e-mail], progressive Casualty Insurance Company and affiliates foreign.

[B]

Welcome to this special edition of the Seneca Podcast, a weekly discussion of current affairs in China, coming to you this week from Dalian from the Davos On Air booth at the World Economic Forum's annual meeting of the New Champions, also known as Summer Davos. In this program we look at books, ideas, new research, intellectual currents, and cultural trends that can help us better understand what's happening in China's politics, foreign relations, economics and society. Join me each week for in depth conversations that shed more light and bring less heat to how we think and talk about China. I'm Kaiser Guo. For over 20 years now, I've had the distinct privilege of of working with the World Economic Forum as an official writer. And this year they've asked some podcasters to team up with them to bring you shows both under the World Economic Forum banner and under the banner of their own podcast. So I'm delighted to be able to do this this year with Seneca listeners. Please support my work by becoming a paying subscriber@seneca podcast.com I do need your help to keep doing this work and to keep bringing you these conversations. Technology has always run ahead of the rules meant to govern it. Scholars call this the pacing problem, the chronic lag between what innovators can do and what regulators have figured out how to actually handle. We've seen this with nearly every major new wave of technological innovation. Recent examples would include Uber and Airbnb, for example, or how regulation really had to catch up much earlier than that. You could go back to the advent of the automobile. Cars hit the roads before anyone had invented things like speed limits or driver's licenses or traffic lights, or even the concept of jaywalking. I was lucky enough to be in China and to have a front row seat to watch this whole thing unfold when the Internet really took off and China in the late 1990s. In recent years, artificial intelligence has turned that lag into a real chasm, and nowhere is the dilemma more vivid than in China, a country trying to do two things at once that often can pull in opposite directions. To on the One hand, unleash technological development at extraordinary speed, and on the other, to keep that development firmly within bounds that the state can manage. Out of that tension has come an idea, Agile governance, that has worked its way into the global policy vocabulary. The question I want to explore today is whether it's a genuine model that others can borrow, or whether it's something so deeply rooted in China's particular system that it can't really be transplanted. And there's no better place to ask it than here in Dalian, and where this year's theme is Innovating at Scale. It makes the governance question really more urgent than ever, because scaling at innovation, scaling innovation itself, also means scaling the risks of technology. So my guest is one of the people who has thought longest and hardest about all of this. Shulan is dean of Schwarzman College at Tsinghua University. And because I've been involved with Schwarzman for many years, really since its inception, that's how I know Dean Xie best. But he's also director of Tsinghua's Institute for AI International Governance. He's a Cheung Kong Distinguished Chair professor with a doctorate in engineering and in Public Policy from Carnegie Mellon University in Pittsburgh. He chairs China's National Expert Committee on the Governance of Next Generation AI. And some of you may recall that he was one of those from China who spoke to Senator Bernie Sanders of Vermont on cooperation on AI governance back on April 30th of this year, crucially for our purposes today. He is also one of the intellectual architects of this very concept of agile governance in a. As a concept in Chinese policymaking. So there are a few people about, few people who are better placed to tell us what it really means and whether it travels. Dean Xue, it is a real pleasure to have you at last here in Dalian. And a warm welcome to Seneca.

[C]

Thank you. My pleasure.

[B]

So, Dean, let's start with the basic problem. Every regulator faces, what's been called this pacing problem, technology evolving faster than the rules that can keep up with it. This isn't genuinely new, but it does seem newly acute. What makes it so. It's not just AI either. It's biotech, it's robotics, it's new materials, it's quantum and more.

[C]

Yes, I would still argue that AI is really unique, I think partially because AI has been changing so fast. If you look at the other technologies, I think that very few that can develop so fast. I mean, think about the. The frontier models, right? In the last few years, you know, I think it's maybe every half A year, not every two months and every month, even just few weeks. So that the rapid evolution I think is just really unseen. Yeah, I think that's, that's probably one aspect. The other thing is also the impact on know so many domains of our daily life, of our economic activities. So bio may be in certain sectors, but not in all AI. You can't think of an area that AI would not touch on, right?

[B]

Yeah, yeah, absolutely. So you are very closely associated with this concept of agile governance. So maybe in just a couple of sentences for our listeners, what does that actually mean? And just as importantly, what is it a reaction against? Agile can sound like a Silicon Valley slogan. Move fast and break things. So how is the governance version different from simply light touch regulation or from deregulation?

[C]

Sure, indeed. I think those two words actually kind of very strange combination. When you think about the governance of course it's very official, very formal, it's going to take a long time and so on. So you have that image there. Agile is something that's light, quick and the change fast. So those two combination, I mean it's kind of weird combination, but that's exactly what I think we've been trying to achieve. That is you want to achieve the function of governance, but at the same time the way you achieve that is not through the traditional official approach, but rather you try to learn from other is fierce, for example in industry and so on you try to accelerate. So that's the combination that we're trying to achieve. I think in general, I think there are three, I mean at least I think three core elements. What I see are really embodied in the so called the actual governance. And the first thing is that I'm a scholar of public policy, so. So when we think about governance, we think about regulations. You always try to study this thing very thoughtfully, I mean very thoroughly. And so you want to make sure that you're not leaving any gaps. So you want to review all the evidence, all the problems and so on and then you come up with a comprehensive, accurate kind of regulations and you very thoughtfully deliberate it. And then you go through a very elaborate process, being reviewed by experts, by stakeholders and going through the kind of a due process then to get your final thing coming out. And that's the regular governance that we have to go through. Very unfortunately for a technology like AI, by the time you've gone through that process,

[B]

the thing you regularly changed has

[C]

already changed, it's totally been away. So what do you do? And that's so in a way, if you really want to be useful, you really have to change your mindset. You can't be comprehensive, you can't be accurate, everything and you really have to move fast. So that's, I think the first thing that you have to change your mindset. The second thing is that you also have to change the idea that the regulator and the regulate, the company is being regulated. You are the enemy. You are playing the cat and mouse game. I'm the regulator. I'm going to try to catch you guys. You guys are trying to do something bad so I'm going to catch you. The company is trying, okay, I'm going to try to find out the loopholes so I can slip into that. That's not the case. You really have to work together UI partners because I think it's do nobody's good if something disastrous happens.

[B]

Right.

[C]

And the third thing is the kind of regulation tools. I think of course there are heavy handed regulations like fine of billions of dollars and that can really get the small company just go bankrupt.

[B]

Right.

[C]

And so those kind of tools, you should try to avoid using that. I mean, you know, if of course we really also have to be careful for those kind of life and death issue. You do want to be sure that indeed bad guys do get punished. But in general, if there are honest mistakes and so on, you want it to be, you know, so light touch. You want to be somewhat, you know, kind of nudging rather than heavy handed punishment. Those kind of three things, I mean mindset, the relationship between regulators and regulatees and then the tools you use, those are the three core elements of this agile governance.

[B]

I think a lot of people listening right now probably don't associate the government of the People's Republic of China with the kind of mindset you talked about or with that sort of non adversarial position vis a vis those being regulated. I mean it's often caricatured abroad as a very top down regulatory monolith. You've described something that's much more iterative, much more flexible. Walk us through how a new technology actually gets governed here from emergence to effective rules. And perhaps you can cite an example we might be familiar with.

[C]

I think that if you look at the AI technology development in China, China and how the regulation has been, has been changing. That's exactly. In a way I think that it's a very good example of kind of agile governance. I think of course China has a long history of AI research and development. In the 1980s, Chinese kind of academic sort of professional association of AI research was established. So China has a long history of doing that. But the most recent development of very coordinated policy development was in 2017, when China's next generation of AI development plan was released. The most recent policy document outlining China's plan for AI development. I think until that time, I think China's AI development has been very, I think cautiously. It's cautiously and supportive in a way, allowing the AI to be used and developed and used in a very broad way unless something happens and then of course the government will respond particular on the issue, whatever being raised. So that was probably the first stage and then I think of course the second stage when this plan was released. In this plan there was indeed, I think, issues being identified. For example, along with the technology development, there's also a governance plan also being released because China recognized that there will be some risks associated with the technology development. So there's need to. For governance. So I think that's sort of the. So in parallel trying to be, you know, two tracks technology development on the one hand, the other one is more on the governance side.

[B]

Yeah. And the results have been really impressive that from this time, from 2017 on through different phases of AI's development. In the very early years when, when the technology was at the level of deep learning, the worry was about things like deep fakes, things like, about. About couch fitting and things like that. And then. So there were already regulations requiring watermarks and things like that. And it's pretty impressive how quickly they were able to then roll out rules for registering algorithms for actually generative AI, for labeling generative AI created content, and then even things like. For AI companions now.

[C]

Exactly.

[B]

I mean, I think it's probably surprising to a lot of people that it would be the Chinese government of all governments that would be able to stay on top of these things as quickly.

[C]

Yeah, precisely. I think that's actually if you think about the. In a broader way, you think about China's reform since 1978, that has been the Chinese approach.

[B]

Crossing the river.

[C]

Crossing the river. That's exactly the same way. Right. So that's what I think in a way. So there's nothing surprising for China to follow this approach. So China does not have the ambition to have an over kind of a legal document or a framework that governs everything, to be precise and to be accurate, comprehensive. I think China has been trying to be adaptive and learning along the way. So that's what China has been going through.

[B]

So with that in mind, I mean, we look at the 2023 Interim Measures on Generative AI they were explicitly called interim. Yeah, that strikes me as significant because you know one, I talked to a lot of people who worked on tech regulation on data, for example, in creating for example the personal Information protection law and things like that. One thing that strikes me is that China's system is very different than what you have in the US or in the eu. In China revision and emendation are assumed. It's assumed that this is not a final product. It will have legal force but. But it is assumed that this will be changed in the very, very near future. There aren't these long cumbersome parliamentary or congressional processes necessarily to get through in order to change that. So the regulations are written provisionally. Is that provisionality then in your thinking? Is it a feature rather than a bug regulate, fast learn than revise?

[C]

I'm not sure what you are trying to.

[B]

It's a feature, right? The provisionality of.

[C]

Exactly. Yeah, that is indeed I think a feature. Of course, I think in China if you're passing a law in a kind of very formal going through the legislative process and passing a law is very challenging, it takes a long time and so on. So many of these regulations are not. In that way but rather it's more of a administrative measure to govern the behavior of the companies.

[B]

Guidelines, guidelines. China talks about giving development and security equal weight in practice though when the two collide, which tends to win out and has that balance shifted at all over the past few years in your mind?

[C]

Well, I think there's always of course you have to balance and you have to have the trade offs. So really depending on what kind of risks we are talking about, if this life and death really concerning the human lives and you tend to be on the more cautious side. So I think that's for sure. Sure. And then I think on some other things that may be lesser severe and then you may try to figure out how to balance the two and maybe align this space and then see how that goes and then try to see how you want to move forward. For example as you mentioned, the generative AI I think that China the first time I think there was in 2023, I think in April, first version came out was very strict and many comments said no, no, no, we can't do that. And then there's a lot of feedback and I think argument and so on and so that gone through a few processes and by July when the document came out I think was much more tolerant. So I think that of course now that's already also been updated, that's the process you go through.

[B]

So all of this is happening in a context where there is an AI race whether you accept this framework or not. I think it's a terrible frame. But the fact is that there is a sort of national level competition where the United States is quite deliberately trying to kneecap China, trying to prevent it from having access to the compute that will allow it to really, you know, compete on the frontier. All this is happening then. So how has that affected the mindset of regulators? Does that make them think, well, we need to give the, the companies more room to run, we need to regulate even more lightly?

[C]

I think it'd be naive to say that policymakers does not have that in mind at all. I think that's probably not true. But at the same time to say that given that okay, we're in this race, we're going to let the company do whatever they want, that's not true either. I think for Chinese regulators and that's what they want to make sure AI will bring benefits to people but with minimum potential harms. And I think if it comes to making a trade off and sometimes I just mentioned if the life and death of people, then there's no leniency.

[B]

China often has things and I think this is deep in its reform and opening DNA pilot zones, regulatory sandboxes, tiered or classified supervision. It seems like experimentation is, is very central to the Chinese approach. So what makes that work here in China? This, this approach of experimentation?

[C]

I think, you know, experimentation. I think, you know, as you mentioned, I mean it's already, you know, in this generation DNA. I think some people like to try that. I think that that's probably in terms of the institutional kind of behavior and that's what people. I mean the learning itself is critically important. So when you experiment of course when possible you design so that you won't have everything brought into to the risk. So you want to let's experiment on this side first and then try if it works so in a partial way that you get exposed to various risks and then you gradually learn from that and then you kind of make some change and here and there. So I think that's sort of the kind of process that people have learned over the years how to, you know, in a way the experiment process itself is not a ad hoc one, but rather often it's very carefully designed ones.

[B]

Yeah, I think of things like self driving cars and you know, the robo taxis and having you know, large municipal pilot zones like when Wuhan for Baidu's Apollo and it's Actually they're probably more, I don't know, I don't know what the numbers are compared if you look at Waymo versus the Chinese competitors now, but it's pretty impressive. Again, looking at it from the outside, there's often a surprise to people. People look at the 2020-2021 crackdown on tech companies, this is how it's described and I don't see it this way necessarily, but on companies like Ant Financial who's very lucrative IPO was pulled quite suddenly about, you know, ddtrucin that had its app taken down from app stores because of data violations abroad, new Oriental and other after school tutoring programs that were suddenly shuttered. This looked to a lot of outsiders like just the opposite of Agile. It looked just sudden, quite sweeping, quite costly. Was this a failure of the model or agility of a different kind or what did policymakers take away from this period in the early Covid days?

[C]

I think of course any of this kind of process you of course in the ideal world that will go very smoothly. Right. But at the same time in reality it's not. I think sometimes you accumulate, you accumulate certain kind of tension and at some point when that tension reached a certain breaking point and then you generate response and of course in order to be able to address the perceived risks and maybe the response can be very strong and in a way I think precisely because of that response is strong enough and then there's a whole set of really regulations, including legislations that actually came out and then set up the framework for the next generation of the agile governance to take place. So in a way I think that this is kind of a. The reality that you accumulate, maybe that you've been too soft for a long time and then you reach that breaking point and then you get strong and then you move on to the next phase of relatively gentle ones that I think that of course it's very hard to predict. I think we are now in enter I think in a more because anyway this is a learning process.

[B]

Yes.

[C]

So I think that the tension there, I think around 2020, 2021, that's indeed I think in many ways have accumulated enough complaints from the society, enough concerns on the behavior of the big companies and so on that generated that response.

[B]

Yeah, I mean just to be clear, I see a very compelling logic in why they would have cracked down on Ant Financial because of the moral hazard of this sort of this lending of these, you know, local banks money through the. Yeah, anyway, it was a real textbook moral hazard case and I think it was, it was sensible. Let's talk about the advent of Deep Seek, what many call the Deep Seek moment in January of 2025. How did that moment reshape conversations inside China about innovation versus control and the balance between them?

[C]

I think actually deep sea. I think certainly in China, deep sea doesn't really generate that much of a concern, but rather it's more of a. Really gave the Chinese community confidence. Exactly. And so people, because I think prior to that you see more of the frontier models mostly I think was really generated by the US and there seems to be a kind of Chinese models are lagging behind. And so deepsea really just came out of way nowhere prior to that. Nobody knows about the company and about the model. And also I think it's not just a model itself. It has a unique feature of its own, innovation, improvement and so on. That really gave many of the Chinese companies and Chinese society a sense, a confidence that actually, indeed, there are different ways to innovate, to move forward in frontier models.

[B]

Open weight models really complicate everybody's playbook. China has very much leaned into open weight models and more broadly into open source. How does China think about governing models that it can't fully control once they're released into the world? I mean, this seems like it would give them some headaches.

[C]

Yeah, I think first of all, I don't think this is a really a particular deliberate government policy. Again, I think whether you're open source, open weight or you're model with closed source and close weight, it's very much of a company's choice. So I think in China we have closed source models, we have open source models, and the same is true for the US So I think that itself, I think, indeed, I think the Chinese government certainly, I think in general is very supportive of open source in the sense that China wants to have a broader diffusion both in terms of domestic and also internationally. You know, China has been very active in supporting capacity building and, you know, inclusive development at the un. So I think that itself, I think, you know, China has that tendency, but certainly, you know, it's fully of company's choice.

[B]

Yeah, I think the reason I ask is because I think there's now a kind of conventional wisdom among a lot of observers of AI who that, that China has a preference for open source, that it's, you know, in part because of this, in this deliberate diffusion. They want to make sure that it's these Chinese models that are being used in the nations of the global south, that that it sort of hooks them into a Chinese software ecosystem, they become part of the Chinese stack. But I mean, I, I could be completely wrong about that. Speed has a cost, right? Firms face rules that frankly they can change just from month to month. How do you weigh regulatory agility against the predictability that long term investment often requires demands and needs?

[C]

I think that indeed that's a fascinating issue. I think for AI, I think the uncertainty is the nature is the DNA of AI. And so I think that, I think particularly for frontier models, I think it's very hard to, I think as we've talked to many of the AI people and they said indeed we, you know, I mean, you know, traditional people think, okay, maybe there's an information asymmetry the government don't know. I mean the company, they actually know everything. So there's a kind of information asymmetry in terms of regulation. But actually when we talk to the companies, they often say that's really, you know, we are wrongly blamed. They say in many cases, you know, we don't know what the model will do right when we move to the next state. So that's where I think that this is indeed, I think the issue with AI. So that's why I think we need to extra cautious in terms of AI risks and the governance. So I think that I joined a few groups in discussion on how actually we can make sure that indeed we have guardrails properly set up to guard against potential risks. And so the idea of so called red lines is precisely because of this, you don't know what might happen next. So you want to be cautious enough so that actually there are certain potentially risky boundaries that you don't want to cross.

[B]

Right. Western critics will often argue that China can be agile precisely because it isn't constrained by courts, by an independent judiciary, or by a free press or a constitution that really has a lot of teeth. Is agility partly just a function of fewer checks and balances? Where does rule of law fit in when the model is sort of built for speed?

[C]

Well, I think that's totally untrue if they know the Chinese society, I think in a way that, I mean the checks and balances are maybe the same as in the us, except that the way that's been demonstrated manifested in different ways. For example, if AI generate harms, for example, like the, even for the so called surveillance and you know, facial recognition technology and there are indeed, I think, you know, many, in many places people do have complaints and actually there is, there are, you know, legal cases being contested, you know, about facial recognition. So if you now you go to different places, I think they are actually, for example, in going to the railway station in the past, you have to go through this facial recognition and now that no longer. And so I think you see that checks and balance at work. Except that of course in China's case, China does have a lot of laws and so on in governing many of these activities. But in addition, the public sentiment can also percolate through different channels to really generate policy response.

[B]

Yeah, that's something not well understood outside of.

[C]

Not at all. Not at all.

[B]

Yeah, it's, it always bothers me. I think there's, there are a few governments that, that monitor public sentiment quite as closely as China.

[C]

China, yeah. Yeah.

[B]

Often you see quite a reversal of things. I mean, I remember when, during COVID when we all had the apps, you know, the, the health check apps, everyone assumed that would become a regular feature of life in China, that it would last long beyond. And you know, of course they, they disappeared in very, very quickly. I, I always go back to my own experience in the Internet world in the late 1990s and the early 2000s. And you know, it was really remarkable how you have this, this, this pacing effect. You had these companies that were moving fast and breaking things and doing, asking for, for forgiveness rather than asking for permission and doing some pretty radical things. I like to remind people that you have here, despite rules that explicitly forbid foreign investment, most of these Internet companies were founded with foreign capital run by people who, if they weren't not themselves actually foreign educated, had at least a very foreign facing aspect to them. People like Jack Ma, for example. But a lot of, you know, returnees

[C]

like, you know, Jack Ma didn't have a foreign education.

[B]

He didn't. But he had a foreign facing personality for sure. He loved to speak English, you know, you know, like Zhang Chaoyang, you know, came back. Liam Hong, who came from a lot of these early Internet entrepreneurs were attorneys, some of them were American green card holders. And then they went to American capital markets. And this is in industry that was clearly a commanding height, that was clearly strategically very, very vital and very important. And so I have often wondered what kept the regulation from keeping up. And one thing that often people would tell me is that there was confusion as to who had regulatory authority. I worked for a while for Yoku, which was video site before I went to Baidu. I went, I was at Yoku. And what was interesting was there was always a regulatory fight over who regulated them. Was it the state administration of radio, film and television or was it the, you know, the precursor to the, you know, was it mit, was it before CAC came along? Of course we didn't. And I wonder whether having multiple regulators actually helps agility or hinders it, because right now even there's still some confusion.

[C]

Yeah, yeah. I think to respond to that, I think you really have to understand the evolution of China's regulatory system. I think that for those of you who knows about China's reform process, prior to 1978, China was a so called planned economy. In a planned economy there's no need for regulatory agency. Right. And everything's planned by the state agency and so on. Why do we need a separate regulation? And so when China making the transition from a so called planned economy to a market based economy. And so along this process, the regulation, the idea of regulation began to gradually taking shape. But I think we have to admit China's regulation system is still in the development. Sure. And because I think that the leaders, I think particularly the pioneers in trying to push for the economic reform, trying to bring the market economy to China, they don't like regulators. So I often talk to many of these government officials and they are really very strong support for the reform. So when they hear about regulation, they shake their head. Yeah. I think Chinese regulation system is still, I would argue that still incomplete. Yeah.

[B]

A work in process, a work in progress.

[C]

Progress. So I think. So the boundaries are not always clear, particularly now Internet AI. This is something new that we've not had this before. The impact is very much in different areas. And do you regulate the source or do you regulate at the application side? So there are many of those issues. So I think in a way, I think it's, I would argue probably it's a good idea. We don't have a very experienced fixed regulator. So in a way, I think this different regulatory agency, they also try to work among themselves trying to figure out where's the boundary, whether this belongs to you or belongs to me.

[B]

Being new, they don't have path dependency, they have some flexibility.

[C]

Exactly.

[B]

Adaptability. I think that's a very, very good point. I'm going to be talking tomorrow to a scholar who's recently written a paper talking about how law is actually sort of a sixth layer of AI.

[C]

Okay.

[B]

That, that she, she's got a very interesting idea, but that'll be, it'll be a good conversation. Set the three big approaches alongside each other for us. The eu, which has a real kind of comprehensive AI Act. Right.

[C]

Yeah.

[B]

The US has a very market led patchwork of regulation and then you have China. What does each of these get right when it comes to having agile governance of this new technology?

[C]

I think that I'll just comment, I think along these three, you know, the three ways you just talked about EU. I think let's go talk about EU. I think first of all, I think I admire EU's ambition, right? And they really want, I think this is in a way, I think very from academics, this is the ideal. You want to have a comprehensive framework in governing the risks and then going very methodic about everything. In the particular the potential risk, analyzing that and then finding ways to govern that. I think of course that takes a long time. And that's one thing, the kind of AI they were concerned about is no longer the AI we're talking about today. The other thing is Also I think EU's approach, given the comprehensiveness and given that it's somewhat late, so now in implementation they do have a new challenge. So now they are actually trying to figure out how they can have another set of rules that allows those big ambitious framework to really be implemented. So I think that's sort of the issue they are facing. I think the us, I think is, I mean, people certainly look at the US feel somehow US doesn't have any regulations, which is not true. I think first of all, I think US has more of the traditional kind of a regulation framework that actually has been very comprehensive in different industries and so on that has been in place. So one can argue that in many situations those regulations can still be useful in governing their behavior of the AI companies. And I think the other things also I think that in many states in the U.S. they also have the regulations. Right? So I think currently I think there are many states actually are having the regulations on various aspects of AI development. So I think it's more of the, at the federal level, I think that's probably not, as, you know, as well developed. Yeah, well developed in the other. What's been seen. So I think the China's approach, I would say it's somewhere in between. China does not. So far China does not have a comprehensive AI law, but China has a lot of foundational laws related to AI and Internet development, cybersecurity law and so on. Those foundations are still very useful. At the same time, also there are some administrative regulations as we've been talking about, governing as various aspects of AI development. So maybe at some point of time those will be integrated into a broader framework, a legal framework. So that could be in the process.

[B]

For now, the very modular nature of these seems to be an advantage. It means that you can change one and. Right, right.

[C]

Yeah, indeed.

[B]

Brussels spent years negotiating this AI act in a field that's moving as fast as this. I feel like comprehensiveness is, is a liability. On the other side, the American.

[C]

But, but I think you have to say that there's a Brussel effect. Right.

[B]

This is their biggest export. Right, Exactly.

[C]

So that really provides a, a good framework, a good thing as a reference. So they do have their own externality with a positive or negative, you can argue. Yeah.

[B]

My sense is the American approach is just too deferential to industry. They are sort of jerked around. David Sacks can pick up a phone and tell them, but China. And then they will drop whatever we saw happen very recently.

[C]

Yeah, I mean if I can make a comment there. I think of course the narrative of so called China, US and China in this AI race, I think that's very. Unfortunately it's not the reality. I think Chinese companies, when they develop the frontier models, they do their work. They're not, I don't think they are in their mind they want to compete with the US model. That's not the case. I mean they wanted to of course push the boundary and they want to see how that can be applied in whatever the industry they are focusing on. So we've had quite some interesting discussion with Alibaba people and they certainly definitely want to push for applications which is somewhat different from deepsea. DEEPSEA really wanted to push, you know, for this frontier models.

[B]

So heart of our topic and I said at the beginning, can China's model be applied in other national contexts? When is it genuinely portable and what of it is bound up with China's particular system?

[C]

Yeah, I think, of course, I think that, you know, different countries have different legal culture, different regulatory system and different technological capabilities. So each country has to find its own way. But the one thing that I would say that the advantage of Chinese approach of just kind of being adaptive, so learning by doing, I think that is something that most of the countries cannot avoid. So somehow either way, I mean whether in different format, but you do have to go through that process. And I think that the kind of Chinese approach in the learning by doing and adaptive approach I think can be conducive to that learning.

[B]

I can certainly see a mid sized country in the developing world, in the global south, actually taking a page from China's government's playbook and there will be things that you should be wary of. But I wonder, is there a version of Agile governance that is in your mind Compatible with liberal democracy, with the developed.

[C]

Why not?

[A]

Yeah, why not?

[C]

Right. Why not?

[B]

Yeah. I mean, good.

[C]

But of course, of course people have to define what is liberal democracy.

[B]

Let's pretend we know.

[C]

I think that's a problem. I think that in today's world, we have two minutes of this kind of cold world. I think that really often, I think simplifies many of the reality, which is often so complex and nuanced. And that I think is unfortunate.

[B]

Yeah, well, I mean AI governance is, let's face it, it's fragmenting. We've recently tended together a conference in Hong Kong. I mean I. One, one point that I brought up is that there is this tendency in the west to see AI governance in China simply as a form of censorship, to think of it only in terms of controls, whether at the input or the output level. But you know, we have these rival summits happening, declarations that key players won't sign. Some will somewhat. You've called managing AI risk a global public good. I completely agree. How do we govern a global technology with no global governor? And are we drifting toward rival governance blocks? Can we still see these converge? I mean, you've signed international statements on AI emergency preparedness alongside Western scientists. What are the areas where we can find common ground?

[C]

I think there are many, many areas of a common ground. I think first of all, I think it's that we think for example, the risk for AI out of control if that really becomes a reality. I think there's no U.S. risk. Chinese risk is humanity's risk, so existential risk. I think that's the kind of issue that we want to prevent from happening. So I think that's sort of what we have been trying to argue, that there's a need for a common interest for us to work together to address those issues. Another thing is for example, the so called, the third group of maybe in terms of using AI to generate harms. So there's a possibility that indeed some extreme groups might use AI. Non state actors, non state actors that generate huge harms to human society. That again, we have the common interest to work together to address those. So I think those are the certainly very obvious ones, but also I think they are more, less obvious ones. But I think people should, should understand. Another thing is that there's also common interest for the US and China to work together to see how AI can be brought to many other developing countries for countries which may not have the capability to develop frontier models. Because you don't want to have the world that only US and China have the frontier models can use that to the advantage of those two countries. But the rest of the countries are impoverished with nothing. That kind of world is also very, very dangerous. And that's what the point I was trying to make with Sanders discussion. I said the world needs to be given that we see such a huge potential, why not bring all this benefit to everybody rather than having just in the hand of a few. So there I think that the two countries can work together to see how actually this can be brought to the broader society to have an inclusive AI development. So that's probably another area that we can work together. Of course there are many other areas, for example like climate change, public health. So it'd be so wonderful to have the teams of the US teams and Chinese team to work together and to address those problems. By the way, I think prior to 2016, Chinese scientists, US scientists have worked very closely.

[B]

Yes, absolutely.

[C]

If you look at the cost of the paper, that's probably the highest among all the cost of papers. So I think, I don't think anything have fundamentally changed since then. So why can't we get to those areas?

[B]

Couldn't agree with you more. And Dean Xue, this has been exactly the conversation that I had hoped that we would have. Thank you so much for your time and for your candor and for imparting your wisdom to our listeners. I am Kaiser Guo. This has been the Seneca Podcast recorded in collaboration with Davos, on air here at the World Economic Forums and annual meeting of the new Champions in Dalian. Thanks for listening and Dean said, thank you so much once again.

[C]

Thank you. Thank you.

[A]

This episode is brought to you by Progressive Insurance. Do you ever think about switching insurance companies to see if you could save some cash? Progressive makes it easy to see if you could save when you bundle your home and auto policies. Try [e-mail] Progressive Casualty Insurance Company and affiliates. Potential savings will vary. Not available in all states. Truckers aren't just moving goods. They're making sure bakers get their chocolate chips and hotels get their tiny soaps. But truckers can't do this if they're not on the road. That's why Progressive has over 360 heavy truck employees to help truckers stay on time and on track. Quote Truck Insurance today in as little as eight [e-mail], progressive Casualty Insurance Company and affiliates.