In a recent piece on Lawfare, Simon Goldstein and Peter N. Salib make the case that AI with cooperation is better than attempting some sort of AI race, unlike what virtually all of the relevant policymakers in the US advocate. Thus, in response to the Chinese DeepSeek model, US policymakers are doubling down on the idea that the US must “dominate” AI and win against its geopolitical rival. Goldstein and Salib write:
“In any high-stakes competition to obtain powerful military technology, the closer the game, the more sense it makes to declare a truce and cooperate. Cooperation can help to ensure that both superpowers obtain transformative AI around the same time. This preserves the current balance of power, rather than unsettling it and inviting extreme downside risks for both nations.”
The argument is immediately recognizable from nuclear deterrence theory: a balance of power is better than an asymmetry, because asymmetry – or the possibility of one coming soon – is enough to drive both arms races and preemptive action. Consider the example of Reagan’s “Star Wars” program, which was to “make nuclear weapons obsolete” by developing a satellite system capable of intercepting and destroying incoming nuclear warheads. Sounds great! Except there’s a decent argument that the act of developing such a system would be catastrophically bad for stability. On the Soviet side, a rational planner would have to assume that such a system would work when developed. There’s two responses to that. One is to strike preemptively, quick while one’s arsenal is useful. The other is to rapidly build as many nuclear weapons as possible, on the theory that any anti-missile system is vulnerable to being overwhelmed. Neither response makes nuclear war less likely. In the meantime, on the US side, a rational planner would have to assume the system did not work, because it would be impossible to test it adequately. So a US planner would want a robust second strike capability anyway.
At any rate, thus for a sample of how deterrence theory works. Goldstein and Salib analogize the AI situation to an imagined race. “Suppose you find yourself in a race. It is a 10-mile footrace, but with existential stakes. The winner of the footrace gets a very large prize—say, $50 million. The loser gets a bullet in the head.” If you start out the race 9 miles ahead, then it makes sense to run as fast as you can to the finish line, because you know you’ll win. But what if you start even with your opponent? Here they write:
“If winning and losing are the only options, you should of course run like mad. But suppose there is a third choice: You and your opponent can jointly agree to cancel the race. Then the racers get neither a special prize nor a special punishment. The status quo is preserved. This seems to us like the obvious best choice. This point can be put more formally in terms of expected value, uncertainty, and risk aversion. Like our hypothetical footrace, the AI race seems to offer massive payoffs to the winning nation: unilateral global economic and military dominance. If that prize is on offer, with little downside, a self-interested nation would be foolish not to chase it. But if the U.S. and China are at rough AI parity, the calculus changes. Then, both nations must begin to take very seriously the possibility of losing. This makes running the race less valuable, in expectation. It also raises the variance of potential outcomes—both riches and ruin start to look probable. Thus, if the United States and China are averse to the risk of losing the AI race, and being permanently dominated by the other, they should be less willing to keep running the race the closer the finish looks. As with the footrace, there is an alternative to the U.S.-China AI race: Both countries could work together to pursue a peaceful, collaborative approach to AI development. This collaborative alternative to racing has some costs to both sides. Neither party can hope to reap the future rewards that come from stable global dominance. But the risk of being among the dominated also falls precipitously.”
Thus a rational maximin strategy suggests cooperating, and they outline what such a program might look like in the form of a shared, leading-AI research program run by both countries cooperatively, which becomes a repository for the best AI advances, cooperatively pursued. As they suggest, and without minimizing the difficulties in operationalizing it, “the U.S. and China could create a joint AI lab, something like a cross between DeepSeek and OpenAI. The lab could be co-run by a U.S. and Chinese CEO. The U.S. and China could take an equal share of the profits in tax revenue and could have equal governance control over the company.”
I think this vision is right. I also think that the dip into games theory shows where some of the barriers to its happening lie (other than dumb nationalism and corporate greed, which should never be underestimated). The situation strikes me as similar to a prisoner’s dilemma in the following way. In a classic prisoner’s dilemma (PD), two players are arrested and accused of plotting a crime together. They are told that they can either confess guilt or insist they are innocent. If they confess, they will be given a nominal sentence. If they maintain innocence, then what happens next depends on the other person. If the other person confesses, the one who insists on his innocence is given a draconian sentence. If both parties insist they are innocent, they are freed.
Here, cooperative behavior is clearly collectively rational, since it generates the best outcome. The problem is that you can’t get there, because individually the rational option is to confess: the sentence for confessing isn’t as good as going free, but it’s the only way to avoid a catastrophic outcome. If you set it up right, it’s almost impossible to explain how cooperative behavior could emerge.
Something like this could explain the AI scenario. If you cooperate, that’s the best outcome because it gets to the good things AI can do faster. But if you try to cooperate or otherwise don’t look out for yourself first, you risk being on the wrong side of global domination, at least according to what the AI strategists think. So of course everybody wants to compete.
I’ve reframed the race scenario a little to get here but I think the PD logic is there. The reason to talk about this is a PD is that the PD is, at the end of the day, constructed around two premises. One is that the prisoners do not know anything about what the other will do. The other is that they only interact in this scenario (i.e., it’s a “one shot” prisoner’s dilemma). If the prisoners know each other well, then the odds of cooperative behavior improve. And Jean Hampton argued, in her Hobbes book, that if the PD is iterative – if the players know they will encounter each other again and again – then the odds of cooperative behavior increase, because I’m less likely to screw over somebody that I’m going to interact with later.
This seems to me to suggest a couple of modest steps forward. First, this narrative that we’re all in a mad rush to develop Artificial General Intelligence which will then enable whoever has it to Rule the World ™ is not just intellectually lazy and ignorant of how sociotechnical processes develop. It’s actively pernicious in terms of international relations, because it enables the construction of AI as a race with a definite finish line. If you redescribe AI in more honest and realistic terms, the conceptual model with an artificial finish line looks less attractive. It also suggests that AI will develop in a series of smaller, iterative steps - and so parties have an incentive to think about future interactions.
Second, we desperately need low-level cooperative measures – what arms control negotiators would call confidence-building measures. Here, again, the US is going in exactly the wrong direction, shutting itself off from the world and trying hard to stamp out international cooperation wherever it can be found. In order to get the confidence to be the first mover when it's important, you have to have built up trust in a series of smaller-scale measures where the stakes are lower. This, I take it, is also a lesson from Hobbes: part of the problem with the state of nature is that it is irrational to be the first-mover in cooperative behavior, because you always risk being suckered. Indeed, this is part of what the Leviathan does: it imposes a significant penalty for defection, making it rational to be the first mover and trust the other person to reciprocate. This is also how the presence of contract law enables contracts - it enables me to have recourse if someone doesn't hold up their end of the contract. That then makes it rational for me to trust them.
None of this means that China is not a global adversary, but it does mean that to the extent that an AI race is irrational, we’re not doing anything to prevent one. The project internationally is obviously difficult, as there is no global Leviathan-state, and Hobbes says that international relations are sort-of like a state of nature between sovereign states. That doesn't mean we shouldn't try. As Goldstein and Salib put it:
“If AI development will be as transformative as both nations seem to believe, the upsides will be immense, even if dominance is off the table. Economic growth will explode. Science will progress. Diseases will be cured. And much more. Racing is not necessary to achieve any of this. All of these destinations lie, at least potentially, along the path of cooperation. Indeed, the cooperative path is by far the most likely to get us there. The alternative route, of breakneck competition and existential conflict, is extraordinarily perilous.”
Recent Comments