Schmidt Warns Against AI Arms Race — But His Fix Has New Risks
In a groundbreaking shift from the prevailing views among U.S. policymakers, former Google CEO Eric Schmidt has co-authored a thought-provoking policy paper, cautioning against the widely speculated “Manhattan Project” strategy for developing artificial general intelligence (AGI). Collaborating with Scale AI CEO Alexandr Wang and Center for AI Safety Director Dan Hendrycks, the paper, titled “Superintelligence Strategy,” issues a stark warning regarding the repercussions of an aggressive American push for superintelligent AI, potentially leading to heightened tensions with China and unsettle global diplomacy.
Schmidt and his team critically examine the presupposition that international competitors will passively accept American AGI superiority. They argue that a “Manhattan Project” for AGI could catalyze hostile counteractions, undermining global stability rather than securing it. “[A] Manhattan Project [for AGI] assumes that rivals will acquiesce to an enduring imbalance or omnicide rather than move to prevent it,” they caution. “Such a strategy risks prompting hostile countermeasures and escalating tensions.”
This discourse surfaces at a pivotal juncture, as a U.S. legislative commission has recently recommended a massive “Manhattan Project-style” initiative for AGI, echoing the scale of the 1940s atomic bomb program. U.S. Secretary of Energy Chris Wright has advocated for the inception of a new era in AI development, with the Trump administration revealing a $500 billion “Stargate Project” to boost AI infrastructure.
The authors highlight that the United States finds itself in an AGI impasse reminiscent of Cold War-era mutually assured destruction dynamics. Just as superpowers have historically avoided exclusive control over nuclear arsenals to avert preemptive strikes, they propose caution in dominating AI systems.
Schmidt and his co-authors suggest refocusing efforts from triumphing in the superintelligence race to formulating methodologies that deter other nations from advancing superintelligent AI. They propose an innovative concept called Mutual Assured AI Malfunction (MAIM), where governments would preemptively disable potentially harmful AI projects instead of waiting for rivals to weaponize AGI.
The paper identifies a divergence in AI policy thinking between “doomers,” who foresee catastrophic AI outcomes as inevitable, and “ostriches,” who support rapid AI advancement without trepidation. Instead, the authors promote a middle ground—an approach that emphasizes defensive strategies over unfettered competition.
Schmidt, Wang, and Hendrycks recommend enhancing cyberattack capacities to neutralize threatening projects, controlling adversaries’ access to advanced AI components and open-source models, and ensuring domestic production of AI chips. This strategy focuses on deterrence through sabotage, restricted access to weaponizable systems, and bolstered domestic technological foundations.
This perspective marks a departure for Schmidt, who formerly pushed for vigorous competition with China in AI development. Just recently, he remarked on how DeepSeek signified a turning point in the U.S.’s AI rivalry with China.
While Schmidt’s deterrence plan bears promise within a framework akin to nuclear arms, it might underestimate other countries’ potential to advance AGI while thwarting U.S. interventions. China’s proven prowess in AI and cybersecurity suggests a formidable counterforce.
In an increasingly complex landscape where several nations boast advanced AI and cyber capabilities, the feasibility of such deterrence strategies hits an impasse. If numerous countries advance AGI alongside robust defenses against sabotage, the MAIM concept might inadvertently fuel an AI arms race rather than prevent it. As global leaders grapple with this daunting task, finding equilibrium between competition and collaboration in AGI development emerges as an urgent and monumental challenge on the international stage.
