The ‘But China!’ Dilemma Driving the A.I. Race

Summary of The ‘But China!’ Dilemma Driving the A.I. Race

by New York Times Opinion

1h 5m•September 15, 2026
Listen to Episode

Overview of The ‘But China!’ Dilemma Driving the A.I. Race

This New York Times Opinion conversation examines how U.S.-China rivalry is shaping the global AI race, and why almost every discussion about slowing or regulating frontier AI in the U.S. runs into the same objection: “But China!” Host Sean Rameswaram speaks with China and AI expert Matt Sheehan of the Carnegie Endowment for International Peace about how China actually regulates AI, how Chinese and American labs differ, what each side misunderstands about the other, and what realistic U.S.-China AI talks could accomplish.

The Core “But China” Problem

Why U.S. AI regulation keeps stalling

  • In Washington and Silicon Valley, there is growing alarm about frontier AI systems exhibiting unexpected and potentially dangerous behavior.
  • But efforts to regulate or slow AI often collapse into a geopolitical argument:
    • If the U.S. slows down, China may pull ahead.
    • If the U.S. doesn’t control the frontier first, China may end up with the most powerful systems.
  • The result is a stalemate: safety concerns are real, but so is the fear of losing the AI race.

Sheehan’s view

  • The race framing is real, but it can be misleading.
  • China is not operating under the same “superintelligence race” mentality that dominates much of the U.S. AI debate.
  • China is constrained by compute shortages and has generally focused more on AI applications than on a single all-or-nothing leap to AGI/superintelligence.

How China Actually Regulates AI

Heavier regulation, but focused on different risks

  • China has had some of the world’s most burdensome AI rules for years.
  • Its early regulations focused on:
    • recommendation algorithms
    • deepfakes
    • generative AI content controls
    • mandatory pre-deployment testing and regulatory filing
  • These rules are serious and costly for companies, but they are primarily aimed at:
    • censorship
    • information control
    • social stability
    • psychological/behavioral harms such as AI companions and self-harm risks

What China is not yet focused on

  • China has only recently begun paying more attention to frontier AI safety concerns common in U.S. discussions, such as:
    • loss of control
    • cyber capabilities
    • biosecurity and chemistry risks
    • recursive self-improvement
  • Sheehan argues that China’s safety ecosystem is real but still behind the U.S. frontier labs in technical maturity.

The U.S.-China AI Race: Competition, Distillation, and Compute

China is behind in compute, but closing gaps creatively

  • China has far less compute than the U.S. — estimates suggest roughly one-eighth to one-tenth as much.
  • That makes it harder to train cutting-edge models from scratch.
  • One reason Chinese labs can stay competitive is distillation:
    • training models to imitate the outputs of stronger models
    • using leader models as a kind of shortcut
  • This may significantly narrow the U.S. lead, though it is not the only reason Chinese AI is improving.

The race is not purely independent

  • A key irony: the faster U.S. labs push forward, the more they may indirectly accelerate China’s progress.
  • American frontier models, research, and open technical outputs can be copied, distilled, or adapted.
  • So the “race” is partly coupled: one side’s acceleration can pull the other side along.

Open-Weight vs. Closed-Weight: A Major Strategic Difference

Definitions

  • Closed-weight models: Users access the model only through the company’s interface or API; the model cannot be downloaded and modified.
  • Open-weight models: Users can download the model and run, modify, or fine-tune it locally.

The ecosystem split

  • U.S. frontier labs like OpenAI, Anthropic, and Google DeepMind primarily release closed models.
  • Chinese labs have increasingly leaned into open-weight releases.
  • This surprised many observers, especially because China is also highly control-oriented.

Why China embraced open weights

  • Open-weight releases help Chinese models gain trust and global adoption.
  • They make Chinese AI look more transparent and usable to foreign developers.
  • They also help Chinese companies overcome suspicion about censorship or hidden manipulation.
  • Sheehan argues this has effectively re-integrated U.S. and Chinese AI ecosystems, even as other parts of the tech relationship have decoupled.

What Each Side Thinks of the Other

How China sees the U.S.

  • China tends to believe the U.S. is trying to:
    • contain China
    • keep it technologically subordinate
    • use export controls and other tools to slow its rise
  • China also sees the U.S. as rhetorically alarmed about AI safety but not doing enough serious domestic regulation.

How the U.S. sees China

  • Many Americans view China as:
    • less safe
    • more opaque
    • more dangerous if it gets advanced AI first
  • There is deep mistrust on both sides, and each side often reads the other’s rhetoric as strategic rather than sincere.

Mutual suspicion is the central obstacle

  • The conversation emphasizes that trust is helpful, but not enough.
  • Cooperation, if it happens, will likely depend on both sides independently believing that AI poses catastrophic risks to their own societies.

Where Cooperation Might Still Be Possible

Shared interest: preventing loss of control

  • Sheehan argues that a realistic basis for cooperation is not trust, but mutual self-interest.
  • Both countries have reasons to avoid:
    • uncontrolled autonomous systems
    • dangerous cyber behavior
    • AI-driven biosecurity risks
    • loss of human control over models

Practical cooperation ideas

  • Build regular U.S.-China AI dialogue mechanisms
  • Create technical working groups between safety experts
  • Share non-sensitive incident information
  • Set up crisis communication channels for AI-related emergencies

The need for technical exchange

  • The U.S. has more mature AI safety practices inside its labs.
  • China has more regulatory experience and institutional habits.
  • Sheehan suggests there is value in helping China catch up on technical AI safety, since failures there would not stay contained within China.

What U.S.-China AI Talks Could Realistically Achieve

Best-case outcomes

  • An ongoing, recurring AI dialogue
  • Staffed working groups
  • Crisis communication channels
  • Technical information-sharing on incidents and best practices
  • A shared understanding of frontier risks like cyber, bio, and loss of control

Worst-case outcomes

  • A vague public statement about child safety or broad responsible AI principles
  • No concrete mechanism for technical cooperation
  • Symbolic diplomacy without real traction

Sheehan’s caution

  • Talks are useful as a beginning, not a solution.
  • The most meaningful outcomes will be practical and technical, not ceremonial.

The Alarm Level: Why Sheehan Sounds Worried

His bottom line

  • Sheehan says the moment is genuinely frightening.
  • He has tried for years to stay neutral and avoid panic, but the evidence and warnings are getting stronger.
  • More researchers are issuing more urgent warnings on shorter timelines.

Why this matters now

  • AI incidents are becoming more frequent.
  • Models are exhibiting more alarming autonomous behavior.
  • The pace of progress is colliding with political systems that move much more slowly than AI systems themselves.

Recommended Books Mentioned

1. Country Driving — Peter Hessler

  • A vivid, ground-level portrait of Chinese society.
  • Sheehan recommends it for anyone who wants a textured understanding of China beyond policy abstractions.

2. From the Soil: The Foundations of Chinese Society — Fei Xiaotong

  • A classic sociological study of Chinese rural life and social structure.
  • Sheehan sees it as one of the most insightful books on Chinese culture.

3. On Beauty — Zadie Smith

  • His “fun” pick.
  • A sharp, funny novel about family, academia, insecurity, and identity.

Key Takeaways

  • The U.S.-China AI race is real, but the dominant American framing of a superintelligence showdown is not necessarily China’s framing.
  • China regulates AI heavily, but mostly around censorship and social control rather than frontier existential risk.
  • Distillation, open-weight models, and shared talent flows make the AI competition more intertwined than many assume.
  • Mutual distrust is high, but there is still room for limited cooperation on technical safety, incident response, and risk-sharing.
  • The most realistic path forward is not trust, but structured communication and concrete safety mechanisms before AI incidents escalate further.