ChinAI #360: Anthropic’s Dogma on US-China AI Competition

Greetings from a world where…

church is in session

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Anthropic’s Dogma on US-China AI Competition

I’ve been going to church since I was a kid. Early on at the University of Iowa, when I identified as an evangelical Christian, I went on mission trips overseas and led prayer meetings. Over the years, my faith has morphed and faded and deepened and changed. Today, as fellow “exvangelicals”, my wife and I still attend a Lutheran church almost every week, and we go to a young adult Bible study on Wednesday nights.

All this to say that I am relatively sensitive to dogmatism: an unfounded confidence in asserted opinions not based in fact. As I read more and more of Anthropic’s policy papers and CEO Dario Amodei’s blog posts, the elements of religious dogma abound: infallible faith in a set of unquestioned beliefs; justification of doctrine based on appeals to exclusive access to divine wisdom; an unwillingness to consider contradictory evidence. Sound familiar?

On May 14, Anthropic laid out its beliefs about U.S.-China competition in a new paper titled “2028: Two scenarios for global AI leadership.” This paper envisions two scenarios for the world in 2028: in the first, the U.S. has a commanding lead in AI; in the second, China’s AI ecosystem remains neck-and-neck, which results in “authoritarian AI leadership”. Their policy prescription is to intensify efforts to prevent Chinese labs from accessing export-controlled chips. Taking on the position of a religious skeptic, let us question seven of this post’s unfounded assumptions.

Assumption #1: Because transformative AI arrives in two years, the actions that decide the US-China competition will only occur in this limited window.

In January 2025, Dario posted a version of this “two scenarios” essay, in which he warned that transformative AI — “AI that is smarter than almost all humans at almost all things” — would arrive in 2026-2027. What explains the updated timeline? Why did we go from 2026 to 2028?

Looking back at my rebuttal to Dario’s post, it’s funny (and sad) how neatly this point still applies to the recent Anthropic paper: “It’s really important to note Dario’s assumed one-to-two year timeline. I’ve been in governance of AI circles since 2017. In that time, I’ve consistently heard some iteration of: AGI is two years away! It’s the Bruno Caboclo of technologies: always two years away from being two years away.”

I can’t wait until next year’s paper when Anthropic tells us that we only have a limited window of time before 2030, when the company will open up a box, and transformative AI will transform everything.

Assumption #2: Transformative AI capabilities will diffuse almost immediately.

In line with this two-year window, the Two Scenarios paper argues that frontier AI could quickly reshape the military balance in China’s favor: “These capabilities will not diffuse slowly. When a new model reaches a new capability in autonomous targeting, vulnerability discovery, or swarm coordination, for example, the regime that controls it can put it onto the field in weeks, not years.”¹

History teaches us that this is not the timeframe by which general-purpose technologies like AI diffuse across economic and military systems. My book Technology and the Rise of Great Powers verified these protracted timelines across more than ten different general-purpose technologies. Based on research into how electricity affected the military balance of power, I also found that electrification transformed militaries in a gradual, decades-long process, in which the main impacts materialized not through the eye-catching weapons systems (e.g., the drone swarms that attract Anthropic’s attention) but instead in communication, transportation, and logistics applications.

To be sure, AI could be different from past GPTs, especially if recursive self-improvement accelerates capabilities development and diffusion. Well, consider OpenAI’s fascinating reflections on “harness engineering,” based on their effort to built a software product with “0 lines of manually-written code.” What resonated with me about this early taste of the challenges associated with “AI building AI” is the organizational adjustments that OpenAI had to make. For a time, they spent every Friday cleaning up “AI slop.” They also needed to readjust their software engineering team’s primary job toward designing environments, specifying incentives for AI agents, and building feedback loops to monitor those agents. This was an experiment with a small team of three engineers. Imagine the challenge of diffusing these organizational adjustments across the entire economy or military. This will take time.

Before moving on to the next unfounded premises, I want to underscore why it’s important to question these assumptions. This is not a navel-gazing exercise. If the timeline of when transformative AI systems actually make an impact on national power is not 2026 (or 2028, or 2029) but, say, ten years from now or later (which is more in line with the diffusion marathon of past GPTs), then the basis of policy debates changes. In a past critique of the U.S.’s export control policy, I’ve made this short-term versus long-term distinction:

We haven’t even mentioned other downsides such as the risk that the controls accelerate a) China’s development of an independent high-end chip capacity and b) the reorientation of chip supply chains in a way that insulates them from future controls. If you are operating in a world where AI does not make its mark until after at least a decade, these downsides become more salient and the benefits of the controls become more muted.

Zooming out, the fast timeline assumption fits with Anthropic’s rhetorical strategy — “we have only a limited period of time to set the conditions of the competition,” asserts the paper — to not quite manufacture a state of emergency but strategically deploy one for its own purposes. It is an approach reminiscent of some evangelicals: claiming that the end times are near, some Christians justify war (including the current military operations in Iran) because it fulfills an end-times prophecy to trigger Christ’s return. Note: Coincidentally, at church a few weeks ago, our congregation studied Acts 1:7. When apostles ask Jesus when he will restore the kingdom to Israel, he responds, “It is not for you to know the times or periods that the Father has set by his own authority.”

Assumption #3: If China’s AI models are close to US models in capabilities, China will win on adopting these capabilities across economic, military, and technological domains.

In Dario’s 2025 blog post, he makes a similar claim: “Even if the US and China were at parity in AI systems, it seems likely that China could direct more talent, capital, and focus to military applications of the technology. Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything.”

Again, my book and past research on China’s “diffusion deficit” thoroughly debunks this assertion. Take, cloud computing as an example. China has strong cloud computing giants like Alibaba Cloud, but its overall cloud adoption rate, across all businesses, is half the U.S.’s rate. As for military deployment, the U.S. military likely leads the world in leveraging private technology firms to support cloud computing adoption. In a forthcoming paper in Security Studies, I take aim at this assumption that China’s civil-military integration outpaces the United States: “When applied to China’s ‘military civil-fusion efforts,’ the four-channel measure reveals that China lags significantly behind the U.S. in terms of the efficient use of common technologies, facilities, and personnel for military and industrial purposes.”

Assumption #4: There are only two scenarios. Support Anthropic’s leadership to secure democratic AI, or authoritarian AI will rule the day. I mean, come on, it’s almost too easy to make the connection to the false dichotomies of religious fanatics. We are good, everyone else is evil.

Look, I think we should be concerned about AI’s impact on surveillance. I’ve consistently covered this issue in these pages (from ChinAI #29 to #357). But it’s a big leap from China keeping pace in AI to the entrenchment of authoritarian repression around the world. There’s some dangerous technological determinism in this thinking, which likely repeats the same mistakes of those who thought that the Internet would bring about widespread democratization — just in the other direction. As Internet governance scholars Christopher S. Yoo and Alex Mueller point out:

The premise that China threatens to increase international acceptance of authoritarianism by exporting related values through Internet infrastructure warrants greater skepticism. It appears to reflect the same type of soft-technological determinism as the United States’ early “Internet Freedom” agenda that saw the Internet an unstoppable vehicle for democracy. Just as the existing Internet failed to liberate China, Russia, and Iran, a cyber-sovereign Internet should be no more likely to increase the level of digital repression among governments that have not already embraced authoritarianism.

Core to the “two scenarios” vision is the assumption that there is a full-stack American AI and a full-stack Chinese AI. In fact, in their policy proposal to “champion the export of American AI,” Anthropic is trumpeting a plank of the Trump administration’s “America’s AI Action Plan.”

If, however, you make some effort to comprehend how AI is developing at the ground level, the notion of an all-American or all-Chinese full-stack AI system becomes fanciful at best and purposefully ignorant at worst. Top U.S. AI startup Perplexity, for instance, permits subscribers to use Chinese startup Moonshot AI’s Kimi K2.5 model; previously, Perplexity also supported a version of the DeepSeek-R1 that had been post-trained to remove censorship. I could keep going with examples. Read more about SGLang and inference engines, if you’re interested in the complex web of interdependencies that characterize the AI stack.

Here’s the thing. When talking to policymakers and pundits who aren’t steeped in this field, Anthropic can get away with these flimsy assumptions. When it tells them that the end is nigh, and we need democratic AGI to save the world, it is all too easy to just blindly trust the company’s divine technical wisdom. However, if you’ve done your homework (in other words, the equivalent of reading Acts 1:7 or any number of Bible passages that underscore no one knows the time and hour), you can see right through it.

Given space constraints, I’ve included rejoinders to the following three assumptions in the footnotes:

• Assumption #5: If a PRC AL lab had developed a model like Mythos, the CCP would have had first access to a system that could autonomously discover and penetrate critical American infrastructure.²

• Assumption #6: Chinese AI labs depend on large-scale distillation attacks that illicitly extract the innovations of American companies.³

• Assumption #7: Neck-and-neck competition risks disincentivizing responsible AI.⁴

The broader purpose here is not to debate every single point line-by-line but rather to show the lack of reflexive thinking in this recent Anthropic paper.⁵ Good research and good argumentation — and good faith, for that matter — must exist with a healthy amount of doubt and uncertainty. I’ve just finished grading 47 research papers, and the number one piece of feedback I give to my students is consider how an opposing perspective might answer their research questions and then either cede some ground back or engage with hypothetical counter-arguments. When I read good writing about AI policy, the authors are open to updated assumptions and disconfirmatory evidence. I don’t see these qualities in Anthropic’s paper. It’s dogma all the way down.

This is a peculiar moment. When one of the world’s most valuable private companies publishes a geopolitical thesis with so many holes, I would have expected it to generate more productive debate in these past two weeks. There are many smart people who have a better mix of expertise and insights than me to give critical takes on this paper.

It’s just that many of them are now working for Anthropic (including several former Biden administration officials who shaped AI policy) or hope to be hired by Anthropic in the future.

The Church of Anthropic is growing, and they pay handsomely. Call me a nonbeliever.

Thank you for reading and engaging.

These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

1. When Anthropic claims that a military can put a new AI model on the field “in weeks”, they cite a discussion about the war in Ukraine, in which the research, development, testing, and evaluation cycle of military development has shrunk to weeks. But there’s a lot of evidence that AI’s integration in drones has been overhyped. Also, as this post articulates, there are a lot of issues to fix before AI drones can scale in terms of military applications.

2. There is some validity to this claim. China has implemented Regulations on the Management of Network Product Security Vulnerabilities that might pertain to whether Chinese AI labs must first share Mythos-like capabilities with the government. Yet, one study pointed out the ambiguity of this regulation: “It is not made explicit … what kind of entity is considered a network product provider, or what degree of affiliation an individual can maintain with a network product provider without being subject to the law. In addition, it is unclear what level of severity a vulnerability must have to require reporting (and the law provides for no rating or review process)…Even the MIIT’s ability to cope with the significant volume of vulnerability research the RMSV seems to demand is doubtful.” The study also finds that Chinese researchers continue to contribute to vulnerability disclosures. I’m also not convinced that AI will provide more of an uplift to cyber attackers compared to defenders. If you read the Mozilla-Mythos collaboration post closely, Mythos is a wake-up call, but it also revealed that many of Mozilla’s systems were relatively hardened. On this question of how AI will scale the offense-defense balance in cyber, see Garfinkel and Dafoe 2019; Lohn 2025.

3. On distillation attacks, Nathan Lambert wrote up a great piece that demonstrates how Anthropic’s broad language could hurt the U.S. domestic AI ecosystem. Setting aside what policymaking action would actually look like to disrupt distillation attack, I’m not sure that the company that recently settled for $1.5 billion regarding accusations of pirating half a million books is the right voice to be making these claims.

4. I think this is a significant assumption for researchers to explore more closely. I’m not convinced that a large U.S. lead in AI necessarily produces safer outcomes. In an influential paper on this issue, Armstrong et al. conclude, “The private information case is more complicated. At low capability, adding more teams will certainly worsen the situation, but at higher capability, the effect is reversed.” I am glad to see that this Anthropic paper expresses a desire for successful engagement with China on AI safety issues. If that desire is genuine, it’s hard to ignore that intensified chip controls will undermine cooperation, especially if on-chip verification becomes a useful method for AI governance. The juice might be worth the squeeze on that front, but the downside should be acknowledged.

5. If I had unlimited time, it would be a useful exercise to sketch out multiple alternative scenarios that are more likely than the two choices that Anthropic gives.