The "But China!" Trap: Why Our AI Strategy is Failing

The prevailing "race" metaphor in AI policy is a distraction. By framing AI development as a zero-sum sprint against China, both nations ignore the shared, catastrophic risks of frontier models. This obsession with supremacy creates a perverse incentive: we accelerate development to "win," which forces our adversaries to do the same, effectively tethering our safety to their speed. The true competitive advantage does not lie in reaching superintelligence first, but in building the institutional capacity to govern it. Readers who shift their focus from the "race" to the "governance gap" will gain an analytical edge, identifying where systemic cooperation creates more security than unilateral speed ever could.

The Hidden Cost of the "Race" Metaphor

In the current geopolitical climate, the "But China!" objection acts as a stop for any serious discussion about slowing down or regulating AI development. The logic is simple: if we pause, China wins. If China wins, we face a future of "Chinese killer robots" rather than American ones.

However, Matt Sheehan’s analysis reveals that this binary view is misaligned with reality. China is not a monolithic, unconstrained juggernaut. In fact, they are compute-constrained and have adopted a diffuse, application-heavy approach to AI rather than a singular focus on "takeoff" scenarios.

"The idea that this is just a total binary of like any obligations you put on companies automatically puts you behind this totally wild, unconstrained Chinese juggernaut. That is just not true. That is not based in reality."

-- Matt Sheehan

The systems-thinking error here is the belief that the "race" is a fixed path. In reality, the leader (the US) is dragging the follower (China) along through open-weight model releases and shared research ecosystems. By racing faster, we are not just outrunning them; we are refining their capabilities, making the "race" a self-defeating feedback loop.

Why Immediate Discomfort Creates Lasting Moats

Conventional wisdom suggests that strict regulation destroys competitiveness. Sheehan points to the irony that China has maintained some of the world's strictest, most burdensome AI regulations for years while simultaneously closing the gap with the US.

The competitive advantage here is counterintuitive: practical regulatory experience. While US labs debate theoretical safety, Chinese regulators have spent four years building iterative, reusable tools for model registration and content control. They are building muscle that the US lacks.

When we avoid regulation to maintain "speed," we are sacrificing durability. A system that cannot be governed is a system that cannot be reliably deployed. The teams and nations that invest in the work of safety testing and crisis communication protocols now will be the ones capable of managing the next generation of recursive self-improvement, while others are paralyzed by the chaos of their own unconstrained models.

The System Responds: How Distillation Undermines Supremacy

The "But China!" argument ignores the technical reality of how knowledge flows. If US labs continue to lead, Chinese labs will continue to use "distillation," training their models on the outputs of ours.

"Maybe if you slow down China wouldn't be going so fast either."

-- Matt Sheehan

This creates a systemic dependency. If we treat the relationship as a race, we are subsidizing their progress. The smarter move is to shift from a "race" to an "information-sharing" framework. By establishing crisis communication lines, specifically using slow but reliable methods like formal document exchanges, both nations can mitigate the risk of non-state actors or autonomous agents triggering a conflict that neither government intends.

Key Action Items

• Prioritize Governance over Speed: Over the next quarter, shift internal focus from "time-to-market" to "time-to-governance." Investing in robust safety testing now builds the institutional muscle needed for long-term stability.
• Establish "Costly Signals": Instead of vague rhetoric about safety, push for concrete, technical information-sharing on AI incidents (e.g., the Hugging Face hack). This creates a foundation of mutual interest that transcends political distrust.
• Invest in Crisis Infrastructure: This pays off in 12 to 18 months. Advocate for, or build, formal, low-speed communication channels (like document-based protocols) between technical safety teams. Speed is a liability during a crisis; clarity is an asset.
• Reframe the "Open-Weight" Debate: Recognize that open-weight models are a strategic reality. Instead of trying to "close" the ecosystem, focus on building sovereign AI safety standards that can be adapted by other nations, effectively setting the global rules of the road.
• Embrace "Boring" Regulatory Iteration: Accept that initial regulations will be imperfect. The goal is to build a system that can be updated. Over the next year, prioritize the process of revision over the perfection of the first policy.