A new policy argument from three prominent AI figures challenges one of the loudest ideas in U.S. artificial intelligence strategy: that America should organize a Manhattan Project-style push to build AGI before its rivals do.
Former Google CEO Eric Schmidt, Scale AI CEO Alexandr Wang, and Center for AI Safety Director Dan Hendrycks make the case in a paper published Wednesday titled Superintelligence Strategy. Their central claim is direct: the U.S. should not try to secure exclusive control over AI systems with “superhuman” intelligence, also known as AGI.
The warning behind the paper
The paper argues that an aggressive U.S. campaign to dominate superintelligent AI could create a dangerous response from China. The authors say that retaliation could potentially take the form of a cyberattack, and that such a confrontation could destabilize international relations.
The argument is not that AGI is unimportant. It is that a strategy built around exclusive control may create the very insecurity it is supposed to prevent. If one country appears to be racing toward a system that could give it overwhelming power, rivals may not simply accept that imbalance.
“[A] Manhattan Project [for AGI] assumes that rivals will acquiesce to an enduring imbalance or omnicide rather than move to prevent it,”
“What begins as a push for a superweapon and global control risks prompting hostile countermeasures and escalating tensions, thereby undermining the very stability the strategy purports to secure.”
That framing turns the AGI debate away from speed alone and toward strategic reaction. The paper treats superintelligent AI not only as a technical target, but as a geopolitical signal that other governments may read as a threat.
Why the Manhattan Project analogy matters
The paper arrives after a U.S. congressional commission proposed a “Manhattan Project-style” effort to fund AGI development, modeled after America’s atomic bomb program in the 1940s. U.S. Secretary of Energy Chris Wright also recently said the U.S. is at “the start of a new Manhattan Project” on AI while standing in front of a supercomputer site alongside OpenAI co-founder Greg Brockman.
Schmidt, Wang, and Hendrycks challenge the assumption behind that approach. In their view, the best response to China is not necessarily a government-backed race to create and control AGI first. They argue that pushing for dominance over extremely powerful AI systems could resemble a standoff, not a straightforward technology competition.
The source article compares their view to mutually assured destruction. Just as global powers do not seek monopolies over nuclear weapons because doing so could invite a preemptive strike, the authors warn that an attempt to monopolize superintelligent AI could trigger hostile countermeasures.
The comparison is severe, but the source notes that world leaders already consider AI a top military advantage. It also notes that the Pentagon says AI is helping speed up the military’s kill chain. In that context, the authors’ concern is that superintelligence could be interpreted through a military lens long before it is fully understood as a civilian technology.
From winning the race to deterring the risk
The paper introduces a concept called Mutual Assured AI Malfunction, or MAIM. Under this idea, governments could proactively disable threatening AI projects rather than waiting for adversaries to weaponize AGI.
That concept sits at the center of the authors’ proposed shift. They argue that the U.S. should move away from “winning the race to superintelligence” and toward methods that deter other countries from creating superintelligent AI.
The measures described in the source include several defensive or restrictive tools:
- Expanding an arsenal of cyberattacks to disable threatening AI projects controlled by other nations.
- Limiting adversaries’ access to advanced AI chips.
- Restricting adversaries’ access to open source models.
This is still an assertive strategy. It does not call for passivity, and it does not suggest that AI development is irrelevant to national power. But it changes the objective from building the strongest possible system first to preventing the most threatening systems from emerging unchecked.
A third path in the AI policy debate
The paper also describes a divide in the AI policy world. On one side are the “doomers,” who see catastrophic outcomes from AI development as unavoidable and support slowing AI progress. On the other side are the “ostriches,” who favor accelerating AI development and hoping the risks can be managed later.
Schmidt, Wang, and Hendrycks propose a third way. Their approach supports measured AGI development while putting defensive strategy at the center. In practical terms, that means treating AI capability, access to chips, open source models, cyber operations, and deterrence as parts of the same security problem.
The position is especially notable because Schmidt has previously argued for aggressive U.S. competition with China in advanced AI. The source notes that just a few months ago, Schmidt released an op-ed saying DeepSeek marked a turning point in America’s AI race with China.
That makes the paper less a rejection of competition than a warning about how competition is framed. The authors are not saying that America’s AI choices do not matter. They are saying those choices do not happen in isolation.
What this means for U.S. AI strategy
The Trump administration seems determined to keep pushing ahead in America’s AI development. The question raised by Superintelligence Strategy is whether a maximalist race for AGI dominance would make the U.S. safer or more exposed to escalation.
The paper’s answer is that the U.S. should be cautious about a strategy built around a superintelligence monopoly. If rivals see that path as a move toward global control, they may respond before such a system is completed or deployed.
That is the practical force of the argument. AGI policy is not only about what can be built. It is also about how the attempt to build it changes the behavior of other countries watching the race unfold.
As the world watches America push the limits of AI, Schmidt and his co-authors suggest that a defensive strategy may be wiser than a new Manhattan Project for AGI.