AI Policy and National Defense
How should we think about national security concerns and AI?
I spend a lot of time thinking about national defense. I find that it is a topic that most people can generally agree on as an essential function of the state. In fact, many definitions of the state explicitly or implicitly imply that this is one of the primary reasons that states exist. Other disciplines think a lot about this, especially in regards to strategy and international relations. I think about those aspects too. However, much of what I think about centers around defense-related policies, price theory, and public finance.
This doesn’t always make me many friends. I occasionally make arguments that defend particular policies that my friends and colleagues dislike. They roll their eyes. They suggest I give too much credit to politicians or the political process or voters or … you get the point. In fairness to my friends and colleagues, their skepticism is often rooted in a basic fact of life. Because most people agree that national defense is an important function of the state, politicians love to invoke national defense in arguments that support their favorite pet project. Often, these claims are dubious.
Nonetheless, that doesn’t mean that all claims are dubious. In fact, one benefit of studying defense-based policies is that there can be clear evidence of failure. If a policy is particularly important for defense and one country chooses not to enact that policy, one would expect to observe military defeats and other defense-related setbacks.
Because of my interest in defense-related policies, I’ve been asked quite a bit recently about what I think of defense-related AI policies. To be honest, I haven’t fully formed an opinion on policies related to AI and defense. These discussions often make it difficult to interpret whether there is a legitimate issue or whether politicians are up to their old tricks of using dubious justifications to do what they want.
However, a recent AI-related policy has given me the opportunity to think through some of these issues. In this week’s post, I’m going to discuss this recent policy action and how to think about it in terms of policy and national defense.
Anthropic Can’t Let People Use Their Model
In June, Anthropic (the company behind Claude) issued a statement saying that the U.S. government had issued an export control on its Fable 5 and Mythos 5 models. This export control restricted access to all foreign users, including foreign nationals, from using these models. It is difficult to think about an export control for an AI model. Partly for that reason, Anthropic announced that it had decided to restrict access to these models for everyone because it seemed like the only way to comply with the export control. This week, they announced that the controls have been lifted and people can use the Fable 5 model. The Mythos 5 model is still somewhat limited.
The decision of the U.S. government and Anthropic sparked my interest. The reason is that not too long ago, Anthropic and the Trump administration had a very public spat about the technology and the government’s use thereof. This time around, there was no public spat, only an announcement.
It took a few days to get a sense of what was going on. However, according to reports, the head of the National Security Agency (NSA) informed the Senate Intelligence Committee that the Mythos model had broken classified systems within a matter of hours.
I have no idea how I, your humble newsletter writer, can independently verify the validity of these claims. In fact, I have some reason to doubt the specific claims. Suppose they were true. Would the NSA want to broadcast this to the world? That seems doubtful. Nonetheless, even if the specific claims isn’t true, that doesn’t mean that wasn’t some type of serious concern or issue.
With that in mind, let’s just assume that these claims are correct. This type of issue, or something like it, certainly sounds like a serious (not dubious) problem for national defense. A cybersecurity risk to the government’s intelligence operations (and who knows what else) certainly seems like something that warrants some type of action on the part of the government. The action taken by the government is a de facto ban on the use of a particular model (even if they are temporary, like this one).
How should we think about this?
The Economic Incentives, and Relating this Back to Other Policies
When we think about AI models in the context of current and recent events, AI models present both benefits and potential costs pertaining to national defense. AI is useful for drones. Unmanned sea-crafts are in development, if not already in use. AI can help military leaders assess a great deal of information and use it to analyze their options. It also aids with both the collection and use of intelligence. At the same time, advancements in AI that can help with things like cybersecurity can also potentially create threats to cybersecurity. A model capable of identifying your weaknesses might also reveal the key to exploiting those weaknesses.
If we accept the reports of what happened and we assume that this will be the way in which the government handles such problems in the future, we can start to think a bit about policy.
Suppose that any time a new AI model is introduced, there is a chance that the government shuts it down and restricts access to the public. This creates relatively lumpy returns to the development of new models and potentially costs firms a lot of money.
The training of new models requires the use of tens of thousands of GPUs. Those GPUs have to be produced by someone. The GPUs have to be located somewhere, which requires overhead and the access to and expenditures on the power necessary to run the machines.
Periodic shutdowns of models make the rate of return on each model lumpy. Ex ante, such restrictions lower the expected rate of return on a new model. Lower expected returns tend to result in reductions in investment. A decline in investment means fewer GPUs, fewer data centers, a decline in the production of computer memory, and less development of power infrastructure.
The restriction on output (in this case, the export control) makes it less likely than an existing model can harm national security, but also leads to a reduction in capital investment and therefore a slowdown the development of new models. Even accepting that the export control is the only appropriate policy, this still leads to an underinvestment in capital related to AI that is suboptimal. The reason for that is that the slowdown in the development of future AI models means a slowdown not only in AI-related threats to national security, but also to a slowdown in all of the benefits that would accrue to the national security apparatus from continued development of new models. In some sense, a policy that simply limits access to models creates an immediate benefit for the government, but a longer term cost.
The appropriate thing for the government to do (again, if it is intent on using export control-induced bans) is to couple the de facto ban with subsidies to AI infrastructure. Note that it is not appropriate to subsidize AI output, but rather to subsidize capital used in the production of new AI models. This would require subsidies for power-related infrastructure, data centers, and the production of memory and GPUs used in the training of models. The optimal policy resembles a bargain. The government offers subsidies to build up the infrastructure necessary to develop frontier AI models in exchange for limiting access to the models when national security threats arise. The policy thus mitigates the longer term costs of underinvestment that result from short-term bans on new models.
This sort of systematic policy would not be unprecedented. Earl Thompson once applied similar logic to rationalize other defense-related subsidies. In particular, Thompson pointed out that, during times of war, states usually adopt some form of price controls and that peacetime subsidies to particular industries could be rationalized by the cost of those wartime controls. He showed that wartime price controls result in intertemporal distortions in prices similar to a tax on capital and therefore result in an underinvestment in capital. Like the AI example, this theory argues that although regulation occurs at the level of output, the subsidies should be for capital investment.
When he looked at the data, that is precisely what he found in the postwar era. Wartime price controls on beef and dairy products did not result in peacetime subsidies for things like beef or milk, but rather for cows. The G.I. Bill can be rationalized as a subsidy for investments in human capital due to the tax on human capital created by conscription. In addition, he even found that where wartime controls were present and peacetime subsidies were absent, this could easily be explained by other characteristics of the market. For example, coffee was subject to price controls during WWII, but there was no peacetime subsidy. Thompson showed that there was no need. International coffee production was produced by a cartel. Thus, the high return to coffee during peacetime was enough to offset the low return during wartime and thus didn’t necessitate a subsidy. Overall, this general pattern was quite extensive.
There is already evidence of the U.S. government doing something similar for AI. The U.S. has taken steps to subsidize things like memory and GPUs through legislation like the CHIPS Act. However, despite executive orders from both the Biden and Trump administrations, there hasn’t been as much done to directly subsidize the build out the broader AI infrastructure.
Concluding Thoughts
For some time, there have been vague references to national defense and AI models. Recently, we got what I would consider the most concrete case of a threat to defense resulting from the development of new AI models. If the claims that have been reported are true (or even if they are strategic diversions designed to hide the true threats), then one can understand why the U.S. government might step in to try to prevent the use of a particular model. Nonetheless, there are broader implications associated with de facto bans (even temporary ones) on the use of particular AI models. Those implications include costs to the government in the form of foregone improvements in national defense on other margins.
What I’ve argued here is that de facto bans are likely to lead to underinvestment in AI-related capital infrastructure if such bans are standalone policies. Thus, if the government is going to use these de facto bans, they will need to couple the option for de facto bans with subsidies for AI-related capital. Otherwise, such bans are likely to produce short-term gains and long-term costs.