AI Frontiers' Three Models of Sino-American Competition Miss the Real Stakes

The debate forum AI Frontiers published a piece by Bill Drexel on June 30 that asks a question American policymakers cannot answer: how to win the AI race with China without losing American values. Drexel lays out three models of competition. Each one assumes the United States has a coherent set of values to defend and a meaningful choice about how to defend them.

That assumption is the article’s weakness. The publication’s own archive suggests a more uncomfortable reality. The United States is already deep into a competition it does not fully control, and the values question may have already been answered by market forces and export controls that neither model fully accounts for.

Drexel’s three models are worth summarizing because they capture the current policy debate. The first model, which he calls “containment,” treats AI as a strategic asset like nuclear weapons. The goal is to restrict China’s access to advanced chips, models, and talent through export controls and investment screening. The second model, “co-evolution,” accepts that China will develop frontier AI capabilities and focuses on building superior American institutions, norms, and safety practices. The third model, “cooperation,” argues that the risks of uncontrolled AI development are global and require joint governance, including shared safety standards and arms-control-like agreements.

Each model has prominent advocates. The Biden administration’s October 2023 and November 2024 export control rules on advanced semiconductors embodied the containment approach. The co-evolution model echoes the “build better” ethos of Anthropic and DeepMind. The cooperation model has support from figures like former Google CEO Eric Schmidt and some AI safety researchers who argue that catastrophic risks do not respect national borders.

Drexel’s framing is useful as a taxonomy. It fails as a diagnosis. The three models share a silent premise: that American values are stable, legible, and widely shared. They are not. The values debate inside the United States is itself a contest between open science and national security, between free markets and industrial policy, between privacy and surveillance. The CHIPS Act and the export controls already represent a massive departure from American free-market orthodoxy. The Biden administration’s AI executive order of October 2023, which mandated safety testing for frontier models, was contested from both sides. Technology companies argued it was too restrictive. Safety advocates argued it was too weak.

The AI Frontiers archive makes this tension visible. Afek Shamir’s June 19 article on export controls on Anthropic’s most advanced models notes that the restrictions came suddenly and could be “Europe’s wake-up call on AI sovereignty.” The United States is not just competing with China. It is also competing with allies, and the export control regime treats European AI labs as potential adversaries. That is not a values-driven policy. It is a power-maximization policy dressed in values language.

The deeper problem is that none of Drexel’s three models adequately addresses the pace of capability advancement. Govind Pimpale’s June 25 article on nuclear deterrence makes the point starkly. Pimpale argues that a large AI capabilities gap could destabilize nuclear deterrence because a power with superior AI could disable an adversary’s retaliatory capability. That scenario does not fit neatly into containment, co-evolution, or cooperation. It is a race dynamic where the first mover gains a decisive and potentially irreversible advantage. In that environment, the values question becomes secondary to the survival question.

The AI Frontiers piece on “The Quadrillion-Dollar Disagreement on AI and the Economy” by Anton Shenk, published May 11, adds another layer. Shenk traces the divergence in economic forecasts to three specific assumptions about AI’s impact on productivity, labor substitution, and capital accumulation. The range of outcomes is enormous, from stagnation to explosive growth. If the economic impact is as large as the optimistic forecasts suggest, the winner of the AI race may not need to defend its values. It may simply outgrow its competitors.

What Drexel’s article misses is that the United States is already pursuing a fourth model that combines elements of all three without resolving their contradictions. The export controls on advanced chips are containment. The billions of dollars flowing to domestic AI labs through private capital and government contracts are co-evolution. The tentative safety dialogues with Beijing are cooperation. The result is not a strategy. It is a series of tactical responses to an adversary that is also improvising.

The AI Frontiers editorial board includes Helen Toner, Miles Brundage, and Bruce Schneier. These are people who understand the stakes. The publication’s quality is high. But the Drexel piece, like much of the policy discourse, treats the values question as a matter of choice when it is increasingly a matter of path dependence. The United States chose containment when it imposed export controls on NVIDIA’s A100 and H100 chips. It chose containment again when it restricted ASML’s lithography machine sales to China. Those choices have consequences that constrain future options. The values debate is real, but it is happening inside a box built by earlier decisions.

The most honest line in Drexel’s article may be the one that gets the least attention. “American leaders agree that the AI race will shape the balance of power with China. But they can’t agree on how to ensure the technology advances American values.” The disagreement is genuine. What the article does not say is that the disagreement may not matter. The race is already shaping the balance of power. The values question is being answered by default, through the accumulation of tactical decisions that no single model governs.

The AI Frontiers piece on Chinese audiences reading Western AI safety discourse, by Calvin Duff, published May 18, offers a small but telling data point. Chinese tech media engage seriously with Western AI safety treatises. The discourse is not one-way. The values are not entirely separate. The three models of competition assume a clean divide between American and Chinese approaches. The reality is messier. Ideas cross borders. Safety frameworks get adapted. The competition is not just between two value systems. It is between two systems that are already entangled.

What the Drexel article does well is force the reader to confront the question. What are American values in AI? Open science? National security? Market competition? Democratic oversight? The article does not answer because the answer is not settled. The three models are placeholders for a debate that has barely begun.

The real takeaway for AI builders and policymakers is that the window for deliberate value-setting is closing. Every export control, every funding round, every model release is a de facto values choice. The question is whether the choices are being made with awareness or by default. Drexel’s article is a useful map of the terrain. The next step is to recognize that the map is not the territory, and the territory is moving faster than the cartographers.