Anthropic and the Export Control Dilemma

A considerable part of my professional life has involved export controls, first on Senate staff and then as under secretary of commerce for export administration. In the process, I learned firsthand that the field is like quicksand—once you get sucked in, it’s very hard to get out. My solution after leaving government was to concentrate on other things, which is why this column discusses export controls only occasionally. Sometimes, however, an issue comes up that illustrates the fundamental dilemma of export control policy, and I feel compelled to get back into the quicksand and, before sinking beneath the surface, spell out once again the intractable policy problems the issue presents.

Such is the case with Anthropic and the administration’s recent imposition of export controls on its most advanced AI models. Others have written about how that happened and how it might be resolved—in particular, see this excellent CSIS piece by Kate Koren, Kevin Kurland, and Aalok Mehta. I am not going to repeat all that. Instead, I will dredge up some history to put the issue in a context that may be helpful.

Government officials making export control decisions walk a fine line between under-controlling and over-controlling. The consequences of the former are obvious: The bad guys get stuff we don’t want them to have because it might give them military capabilities that would come back to bite us. The consequences of the latter are equally dire but more complicated to explain. In brief, over-controlling creates two risks. One is kneecapping our own industry—cutting off sales that are a major source of the revenue that high-tech companies need to develop next-generation products. I have said for years that the U.S. high-tech community’s biggest dilemma is that China is simultaneously their best customer and their biggest threat. They need the revenue from sales to China, but they know that those sales are helping China catch up technologically with the United States and enabling it to compete more effectively.

The other risk, which is illustrated in the Anthropic case, is that too-strict U.S. controls will encourage buyers to switch to other alternatives and other countries to develop those alternatives. They may be inferior to the U.S. tech, but they’re better than nothing. We have seen this movie before: night vision equipment, commercial communications satellites, high performance computers, encryption technology, semiconductors, and semiconductor manufacturing equipment, to mention a few. The eternal dilemma is how the United States can defend its national security and simultaneously maintain its leadership in critical technologies.

China “hawks” (in the past, Soviet Union hawks) argue that the correct path is to strictly control U.S. high tech to make sure nobody else gets it. Their underlying assumption is that U.S. technology is inevitably better than everyone else’s, always will be, and that any downsides to simply keeping it at home don’t matter much. The other side of the debate (doves?) argues that that assumption is wrong, and that overly strict controls encourage the development of competing technologies that may well end up surpassing ours. Their alternative is to “flood the zone”—aggressively export the U.S. tech stack worldwide and make it the global standard so everyone will buy U.S. and not buy Chinese.

The dilemma for government is that both sides have got it wrong, and there may be no “right” answer. The hawks discount the harm that over-controlling does to U.S. industry, and ultimately to U.S. national security, if it stimulates the creation of better foreign technology. The doves ignore the likelihood that massive exporting means U.S. high-tech products will inevitably end up, one way or another, in China and Russia, either because U.S. companies or the government have not done sufficient due diligence on their customers or because criminals succeed in diverting exports to unintended destinations. In the good old days when the United States had no hostile near-competitors, leakage still occurred but did not make much difference. Now, with an able, aggressive competitor nipping at our heels—particularly in AI—leakage can make a lot of difference.

The dilemma is magnified by the unprecedented challenge AI models present. Whether one believes AI is an existential threat to humanity or not, it seems clear that uncontrolled, unregulated access to the technology is dangerous. One proposed solution to the Anthropic case, and probably the most likely outcome, is that the company will increase cyber defense tools and testing that will convince the government the technology is effectively protected. The conceit of the high-tech community is that they can always fix it, but experience suggests that is wrong. There is never a permanent, foolproof fix, and those who think that one exists are deluding themselves.

Sadly, this column ends on a depressing note. Over-controlling does not work; under-controlling does not work; tech patches won’t work—but there is no clear better choice. So, I fall back on the marathon analogy. If you are in one, there are only two ways to win: run faster or trip the other guy. This column has explained the limitations of tripping the other guy. That leaves us with running faster, which has always been the better answer.

Author’s Note: I retired from CSIS on March 29, 2026. I plan to continue writing this column and participating in the Trade Guys podcast, so please continue to read and listen. However, my CSIS email address will no longer be working, so if readers or podcast listeners want to contact me directly, they should do so at [email protected].

William A. Reinsch is a senior adviser (non-resident) and Scholl Chair emeritus with the Economics Program and Scholl Chair at the Center for Strategic and International Studies in Washington, D.C.