Summary: This essay proposes specific proposals for consolidating AI development within the US in order to ensure US has the power to unilaterally slow down AI without interference from race dynamics. It offers specific proposals for (1) strengthening security measures in frontier AI labs by leveraging existing legal frameworks like ITAR and EAR to restrict access to sensitive technologies and implementing robust background checks and (2) liberalizing skilled immigration, particularly from China, by expanding the number of employment visas. This would expand the U.S. ability to attract top global talent. This multi-pronged strategy aims to concentrate frontier AI development within the US, ensuring US technological leadership while mitigating the risks of foreign espionage, bringing it within US unilateral power to slow down AI development if required.
Insiders have claimed that frontier labs are extremely vulnerable to foreign espionage, and at current security levels, it is possible that foreign governments could steal the weights of leading AI models. Although the RSPs of some major labs propose to increase security before reaching models capable of causing catastrophe, vulnerability to foreign espionage keeps the AI compute race alive by allowing model weight theft. This seems unlikely to have already occurred, given the lack of evidence of foreign models based on proprietary models, but it could occur in the future, prior to labs introducing appropriate security precautions. Indeed, in March 2024, a Google engineer was accused of plotting to transfer commercially sensitive AI data to Chinese companies.
While there’s undoubtedly tremendous interest in AI labs like DeepMind or OpenAi to prevent their secrets being stolen, they may have less commercial interest in their secrets being stolen by foreign labs vs. local competitors because they are not competing with foreign labs for customers[1]. Nevertheless, we may want to think about slowing down AI, if not now, then perhaps in 2 years or 5 years. In that event, it will help if the United States[2] can unilaterally slow down AI. To this end it might be helpful if the US-aligned labs have a significant lead over others. It has been proposed that the US attempt to speed up its own AI development, but of course this is inimical to the goal of slowing down AI. Consequently, the only way to preserve the option to unilaterally slow down AI in the future is to ensure others will not be able to overtake US production.
With all this said, it would be more conducive to global peace and probably to present and future welfare if we can approach AI Safety through global cooperation, through inclusive, shared agreements to approach AI Safely. But this may not be an option, or if it is, one way for the US to ensure that others are keen to cooperate with the US could be to make sure others are dependent on the United States for AI. It may also be the case that no action is needed–labs in the US and UK retain an unassailable lead over others, and will do into the foreseeable future. I don’t know whether that is the case, and if it was, the security measures proposed here would be unnecessary.
Securing US labs with regulation
Because AI labs might not themselves take sufficient precautions, in this essay, I identify regulatory measures the US government could take without passing any legislation in order to ensure security measures are introduced in frontier AI labs. There are several existing legal frameworks which could tighten security in frontier labs, making it more difficult for industrial espionage even if skilled immigration from adversary countries was increased in order to bring their talent pool into the United States. These include the Defense Industrial Base Cybersecurity Program, the International Traffic in Arms Regulations, Export Control Regulations (EAR), and the FBI’s InfraGard and Domestic Security Alliance Council. Of these, mandatory power to restrict access to technology by foreign employees can be exerted using the International Traffic in Arms Regulations and Export Administration Regulations (EAR).
The EAR controls the export of dual-use items—technologies, software, and commodities that have both commercial and potential military or proliferation applications. The Bureau of Industry and Security (BIS) evaluates the technology to determine whether it should be controlled under the EAR, considering the technology’s potential impact on US national security and its commercial availability. Such consideration will include consultation with other federal agents and opportunities for public comment. If necessary, BIS will classify a technology on the Commerce Control List (CCL). Sharing a controlled technology or technical data with a foreign national is considered a “deemed export”, equivalent to sharing that technology with that national’s home country. Sharing technology with nationals of countries of concern can require an export license. This process could compel frontier model developers to restrict involvement in technological development from workers who are nationals of foreign countries of concern, while retaining flexibility to share technology with foreign national employees from other countries not of concern.
The International Traffic in Arms Regulations (ITAR) may also enable the government to control sharing technology with foreign nationals within frontier model companies. ITAR's Category XXI allows the US government (specifically, the Director, Office of Defense Trade Controls Policy) to designate emerging technologies as controlled, even if they're not explicitly listed on the USML. This category could include technical data such as computer software, so perhaps advanced AI technologies with potential military or dual-use applications could fall under this category. If software is deemed to be on ITAR’s Category XXI list, AI developers would need to obtain licenses or authorizations before sharing their technology with foreign nationals, especially from countries of concern. Even sharing technical data or providing training to foreign nationals within the US could be considered a "deemed export" and require authorization. This category seems murkier than the EAR as it is primarily designed to control export of Arms.
Other government programs can assist private companies with cybersecurity and industrial espionage. Programs like the FBI's "InfraGard" and "DSAC" (Domestic Security Alliance Council) provide private companies with information on threats from foreign intelligence services and guidance on how to secure their operations, in order to raise awareness among private companies about the risks of foreign espionage and to help them implement countermeasures.
AI regulation might lean on EAR and ITAR to classify frontier AI models as dual use technology under either EAR or ITAR. Licenses to share information with foreign nationals (i.e., deemed exports) could be contingent on background checks of those foreign nationals. Although the Security Clearance process cannot generally be used for companies that aren’t government contractors, a suitable background check process might be developed in collaboration with US security agencies and the US AISI.
Liberalizing immigration
Section Summary: Currently, United States educational institutions graduate around 38,000 STEM students from China every year. Many stay in the United States for years working on temporary visas before ultimately returning to China. Yet the US skilled immigration system only makes room for 8,000 or so people from China–including those who first arrive as students as well as anyone else who immigrants under a skilled visa–to gain a green card and settle in the United States. The remaining 30,000 or so will eventually return home, taking the lessons they have learned in US educational institutions and companies back home with them, to support China’s industrial and technological development.
I believe this is a national security strategic blunder on the part of the United States. In order to support US national security efforts to remain the leader in artificial intelligence, the US should concentrate AI talent within the US and attract what talent they have from China. Such efforts must be paired with tighter controls on workers in the US developing frontier AI models.
Attracting skilled immigration
Talent is an essential ingredient for any AI industry, and there are signs that talent available to the Chinese AI industry is rapidly growing. Although China lags the US in innovative AI research, there are signs that Chinese talent is growing. In 2019 China was not even on the board in NeurIPS. In 2022 they had 12% of the papers. More generally, the Chinese AI sector is booming and although it is not at the level of competitiveness of the US, and despite severe restrictions on compute by the US, China manages to release generalist LLM and multi-modal models not too far behind the frontier set by the US. Drawing more talent away from China has the potential to concentrate frontier AI research in the United States, slowing AI development in China relative to the United States, complementing the compute limit set by the BIS in 2022 with a “talent limit” or at least a “talent headwind”. Even if security concerns prevent workers from China from directly working in frontier work, they may work in adjacent sectors, freeing up US AI workers to focus on cutting-edge frontier AI.
Reproduced from https://macropolo.org/digital-projects/the-global-ai-talent-tracker/.
How current law limits skilled migration from China and India particularly
Current US law limits employment immigration visas to 140,000 per annum, of which there are 40,040 in each of three skilled categories EB-1, EB-2, and EB-3. Of those, 7% or 8,408 can go to Chinese immigrants, because according to the Immigration and Naturalization Act 1990, a maximum of 7% of all visas can go to immigrants from any one country. Compare that to the number of temporary visas in just one category: the student visa. Even post-pandemic, there are 290,000 Chinese students in the US now. Considering the mix of 4-year undergraduate, 2-years masters’ and 5 year PhD programs in which those students are enrolled, that works out to around 38,000 Chinese STEM graduates of American universities every year--of whom only 8,408, as we’ve seen, can ultimately remain in the US. The remaining 30,000 of these US-educated STEM students will eventually be sent home to contribute to Chinese economic and technological development, many after completing several more years gaining experience in US companies via currently-available temporary visa allowances.[3]
A bipartisan reform: proposal and prospects
Past attempts at reform have failed. Because the demand for US employment visas outstrips the supply, there are country-specific queues of up to 12 years or so for employment visas, and because India and China are large countries, the 7% cap puts their emigrants in the longest queues. The Fairness for High-Skilled Immigrants Act of 2022 tried legislative jujitsu to eliminate the 7% cap and clear the 12-year-long immigration queues without restricting immigrants from other countries. It was difficult, and the American Hospital Association, among others criticized the bill for endangering nurse immigration and affecting healthcare across the country.
Is this politically feasible? On the Republican side, Donald Trump recently said that ‘noncitizens in the US should "automatically" get green cards when they graduate from college’. His staff later partially walked back this commitment, but these are not only idle words on Trump’s part. In 2019, he did introduce a plan that would have increased the number of skilled employment visas by 2-3 times, at the cost of family reunification visas for immigrants with family already living in the United States. The Democratic Party Platform released at the DNC in August 2024 also proposes to increase skilled migration, but through adding to rather than replacing family migration, which would also be expanded, under their proposal. The legislative track record for immigration over the last two years is not stellar, to say the least (“dire” would fit better) but there is a clear will from leaders on both sides of the aisle to increase skilled immigration.
A key misalignment here is whether new skilled immigration would come at the cost of other immigration categories or in addition to them. The 2024 DNC Platform drafters pledge to increase skilled migration in combination with an increase in other categories. Trump’s 2019 plan, by contrast, only demonstrates a desire to increase skilled migration by reducing other categories.
Any compromise between these two positions would depend heavily on the legislative balance of power at any particular time, so speculation on this issue seems somewhat idle. One compromise that might be relatively low impact is eliminating the Family-Sponsored Fourth Preference Class, for brothers and sisters of US citizens. The State Department indicates that applicants for this category must have applied no later than August 2007 in order for their applications to be processed in September 2024, meaning that the category is already almost impossible to use for new applicants for the foreseeable future. Eliminating this category would free up 60,000 visas per year, which could be reallocated to existing skilled visa categories with an exception for the 7% per-country rule. Closing the preference class to new applicants while grandfathering in existing applicants would have minimal impact, avoid a net increase in immigration in the long-term, and preserve access to prospective immigrants with applications already pending.
Limitations
I do not pretend that the two measures described here will constitute the sum of necessary reforms for reducing global AI competition. Rather, I intend to propose two ideas that could complement one another by balancing competing political concerns. A full program for consolidating AI development within the US and allied nations would need to consider technical matters of lab security and other matters.
It is also unclear whether the competitive approach proposed here would have unintended side effects that outweigh the benefits obtained, by actually ratcheting up competition. I believe the proposals here are less likely to provoke a reaction than measures that are aimed at directly limiting AI activities in foreign competitors, such as the Commerce Department BIS's current restrictions on advanced semiconductor exports. As such, the proposals here might be useful for pursuing a soft-competitive approach to AI competition.
Drawing more talent from competitor countries that may engage in espionage may increase the risk of espionage occurring from those countries. You might expect this if you model espionage as a random process where every new immigrant from a country with access to sensitive information has a low, fixed probability of handing over sensitive information. It's not obvious this is the correct model to use, however. For instance, it might be that competitor countries' espionage capacity is fixed by the size of their intelligence services, and that they can only engage as many friendly compatriots in the United States as the capacity of their intelligence service allows. In this case, a higher proportion of immigration from that country would not increase risk. In any event, such risk is mitigated against substantially by the proposed EAR export controls.
Finally, pause-AI advocates may worry that by consolidating technical talent in the United States, we may actually speed up progress towards AGI and that this could increase existential risk. This is a valid concern. Working against this concern, insofar as any immigration reform focuses on drawing talent from competitor countries who would be restricted by the EAR from working in frontier AI, these new immigrants would instead work in adjacent industries that do not directly push the capabilities envelope.
Overall, this article focuses on proposing two technical measures amongst many that may be needed if a competitive approach is desired.
Conclusion
By adopting a dual approach of liberalizing skilled immigration, particularly from China and India, and simultaneously strengthening security measures for frontier AI models, the US can enhance its AI development lead over China in AI while mitigating the risks of foreign espionage. Expanding the pool of skilled immigrants, especially those in STEM fields, allows the US to tap into a global talent pool and concentrate AI innovation within its borders. Simultaneously, leveraging existing legal frameworks like ITAR and EAR to restrict access to sensitive AI technologies, coupled with enhanced background checks and collaboration with security agencies, safeguards against intellectual property theft and ensures that advancements in AI remain predominantly under US control. This strategic combination not only fuels the growth of the US AI sector but also strengthens national security by concentrating critical AI development within its borders, by bringing talent away from foreign adversaries and into the US. Such a combination of liberalization with increased security measures may also have a political advantage, helping to allay security concerns that might otherwise arise from migration. By embracing this multi-faceted strategy, the US can solidify its position as a global leader in AI while safeguarding its technological edge.
- ^
- ^
Or the EU, to the extent the EU has effective jurisdiction over US companies because all major frontier labs operate in the EU.
- ^
There are many caveats to this, but on the whole, I don't think they detract from the main point. Some of those caveats: In reality, many of those 30,000 will stay in the US on temporary visas, including going on to further study. But ultimately, in order to legally remain in the United States, they will need to obtain a permanent residency visa. Some will be able to transition to permanent residency through means other than an employment visa, such as marriage. This model further simplifies other factors--some employment visas are taken up by emigrants moving straight from a job in a foreign country to a job in the United States without first going through the US education system.