For decades, the Department of Commerce has kept a little-known list of technology that are forbidden from being freely marketed to foreign countries due to national security concerns. Any corporation that wants to sell such a technology overseas must ask for authorization, giving the government authority and control over what is sold and to whom.
Tensions between the United States and China are currently being heightened by these export restrictions. They have evolved into the main strategy used by the US to stifle China’s advancement of AI: The government restricted China’s access to the computer chips required to run AI last year, and it is currently in talks to increase those restrictions. The plan resembles a form of economic warfare, a semiconductor analyst told.
The battle lines might soon go beyond chips. People acquainted with the situation claim that Commerce is considering imposing a new ban on a broad range of general-purpose AI programmes as opposed to merely physical components. Experts outlined serious stakes, despite the fact that there is still much to be determined regarding the measures’ implementation and even if they will roll out at all in the end. The restrictions, if implemented, may increase tension with China while eroding the U.S.’s ability to innovate in the field of artificial intelligence.
So-called frontier models are of special interest to Commerce. The term, which has been adopted into the Washington lexicon by some of the very businesses that are attempting to create these models (Microsoft, Google, OpenAI, and Anthropic), describes a type of “advanced” artificial intelligence with flexible and wide-ranging uses that may also develop unexpected and dangerous capabilities. Frontier models do not currently exist, in their opinion. However, a significant white paper from a group of researchers that included researchers from the majority of those tech companies and was released in July makes the case that these models might come from the further improvement of large language models, the technology that underpins ChatGPT. The same prediction abilities that enable ChatGPT to generate sentences may be sufficiently developed in their next generation to make customized misinformation, develop recipes for innovative biochemical weapons, or permit other unforeseen abuses that could endanger public safety.
This is a quite different issue from the use of artificial intelligence to create autonomous military systems, which has been one of the driving forces behind restricting the export of computer chips. Frontier model threats are vague and linked to theories about how new skill sets might “emerge” in AI programmes. The authors of the report contend that it is time to think about them right now. Frontier models have the potential to do significant harm as soon as they are developed and used. The creation of a licensing system that compels firms to obtain authorization before they can release, or possibly even build, frontier AI, is one of the solutions the authors propose in their 51-page document to get ahead of this issue. According to the authors, it is crucial to start acting practically to restrict frontier AI right away.
When ChatGPT was released, policy makers shared the same dread that many have since then: an inability to understand what it all means for the future. At that moment, the white paper was released. In its voluntary AI commitments, a set of rules for top AI companies designed to assure the safe deployment of the technology without surrendering its purported benefits, the White House made use of some of the wording and phrasing from the report shortly after it was published. The Frontier Model Forum was later established by Microsoft, Google, OpenAI, and Anthropic as a business organization to do research and provide suggestions on “safe and responsible” frontier-model development.
According to Markus Anderljung, one of the white paper’s main authors and a researcher at the Centre for a New American Security and the Centre for the Governance of AI, the purpose of the document was simply to promote timely regulatory thinking on a topic that had recently come to the forefront of his and his team’s minds. He said that since AI models are developing quickly, it is necessary to be proactive. He doesn’t know what the next generation of models will be able to do, but he is particularly concerned about a system where choices about what models are made available to the public are solely left up to these private firms, he added.
The four private businesses at the center of discussions about frontier models, however, might benefit from this type of regulation. Meta, which also creates general-purpose AI programmes but just announced a commitment to at least part of them being made available for free, is conspicuously missing from the group. The business models of the competing companies have been put in jeopardy as a result, as they rely in part on their ability to charge for the same technology. Convincing regulators to regulate frontier models could limit Meta’s and other companies’ ability to continue sharing and developing their best AI models through online open-source communities; if the technology needs to be regulated, it would be preferable to do so under conditions that are beneficial to business.
The tech corporations at the center of this discussion were fairly silent when contacted for comment. According to a Google DeepMind spokesman, the business thinks that safety must be prioritized in order to innovate responsibly. As a result, it is collaborating with peers in the industry through the forum to promote research on both short- and long-term hazards. Anthropic thinks that models should be evaluated before being deployed in any way, whether it be commercial or open-source, and that coming up with the right tests is the most crucial issue that needs to be addressed by the government, business, academia, and civil society. Brad Smith, the president of Microsoft, has previously emphasized the importance of government involvement in advancing safe, responsible, and reliable AI development. An inquiry for comments was not answered by OpenAI.
In order to completely integrate concepts for the governance of the models with national-security concerns, the fascination with frontier models has now met with growing anxiety over China. Members of Commerce have met with experts over the past few months to discuss controlling frontier models and whether it would be possible to keep them out of Beijing’s reach. According to a government representative, the department routinely evaluates the environment and modifies its policies as necessary. A further specific request for comment was refused by her.
The fact that the white paper was able to gain traction in this manner reveals the unstable dynamic at play in Washington. Emily Weinstein, research fellow at Georgetown’s Centre for Security and Emerging Technology and who has since joined Commerce as a senior adviser, claims that the tech sector has been aggressively asserting its power and that the panic over AI has made policy makers especially receptive to their messaging. This, together with worries about China and the forthcoming election, is leading to new and perplexing policy ideas on how to precisely characterize and address the AI-regulatory dilemma. According to Weinstein, certain members of the administration are clinging on anything they can because they want to take action.
She continued, The conversations at Commerce are particularly symbolic of this dynamic. Export restrictions on frontier models may be seen as a logical continuation of the department’s past chip-export prohibitions, which really set the stage for focusing on AI at the cutting edge. However, Weinstein referred to it as a weak strategy, and other AI and tech-policy specialists issued similar cautions.
The choice would be an aggravation against China, further fracturing the already strained relationship. Beijing has taken a number of apparent retaliatory actions since the chip export curbs were announced on October 7 of last year, including prohibiting the export of several metals used in chip manufacturing and banning items from the American chip manufacturer Micron Technology. Many Chinese AI researchers have voiced extreme annoyance and regret about having their work—on topics like drug discovery and picture generation—turned into collateral in the U.S.-China tech competition. Most of them said they didn’t perceive themselves as resources of the state, but as global citizens who are advancing technology on a worldwide scale. Many people continue to aspire to work for American firms.
Traditionally, online collaboration among AI researchers has been a routine occurrence. While larger software companies, like those in the white paper, have the resources to create their own models, smaller businesses rely on open sourcing—sharing and expanding on code made available to the public. Smaller developers would have fewer options than ever to create AI goods and services if academics were prohibited from sharing their code, and the AI industry’s biggest players, who are already lobbying Washington, would see their influence grow. “If the export controls are broadly defined to include open-source, that would touch on a third-rail issue,” says Matt Sheehan, a fellow at the Carnegie Endowment for International Peace who focuses on China-related global technology concerns.
The extent to which this cross-border cooperation enhances rather than diminishes American leadership in artificial intelligence is frequently overlooked in discussions. The U.S. and China are each other’s top partners in the development of this technology as they produce the majority of AI researchers and research in the globe. They have drawn inspiration from one another’s work to progress the field and a wide range of applications far more quickly than any one could alone. Unlike the transformer architecture that serves as the foundation for generative AI models, which was developed in the United States, one of the most popular algorithms, ResNet, was developed by Microsoft researchers in China. With the open-source model Llama 2 from Meta, this pattern has persisted. In one recent instance, Sheehan observed an old acquaintance in China who operates a medical-diagnostics business post on social media about how much Llama 2 was assisting his work. Export restrictions on frontier models might therefore “be a pretty direct hit” to the enormous network of Chinese developers that build on American models and subsequently provide their own research and breakthroughs to American artificial intelligence development, according to Sheehan.
However, there are still questions over the technical viability of such export controls. It’s practically impossible to identify exactly which AI models should be banned because the foundation of these limitations rests solely on fictitious risks. As the previous wave of regulations demonstrated, any standards might also be easily avoided, whether by China boosting its own innovation or by American businesses coming up with workarounds. Last year, the California-based chipmaker Nvidia unveiled a less potent chip that precisely met the technical requirements of the export regulations and was able to continue selling to China within a month of the Commerce Department’s announcement of its blockage on potent semiconductors. Since then, Bytedance, Baidu, Tencent, and Alibaba have all placed orders for approximately 100,000 of Nvidia’s China chips to be delivered this year, as well as more for future deliveries. These deals, according to the report, are estimated to be worth $5 billion.
Restricting the company’s exports to China, according to an Nvidia spokeswoman, would have a huge, negative effect on the U.S. economy and its position as a global leader in technology. The company’s chips, the spokesperson added, are essential for speeding beneficial applications around the world. Controlling particular applications, like frontier-AI models, would be a more targeted move, according to the business, with less unexpected repercussions. A request for comment was not answered by Bytedance, Baidu, Tencent, or Alibaba.
Fixating on AI models could occasionally be a diversion from dealing with the main issue: Weinstein argues that getting the supplies and machinery needed to actually synthesize the armaments rather than figuring out a recipe is the bottleneck for developing novel biochemical weapons. It would be difficult to overcome the issue by limiting access to AI models.
The four firms promoting frontier-model legislation may have another advantage, according to Sarah Myers West, managing director of the AI Now Institute. By raising the spectre of potential threats, regulators divert their focus from the current drawbacks of their current business models, such as privacy issues, copyright violations, and job automation. The idea that this technology carries significant dangers, so we don’t want it to fall into the wrong hands.
People exaggerate how much this is in these firms’ best interests, said Anderljung, adding that as an outside collaborator, he cannot fully understand their perspectives. A regulator may very easily inform a corporation that it is not permitted to use the technology even after it has invested $1 billion in establishing a model. He stated, he doesn’t think it’s in any way clear that would be in the interest of companies. He continued by saying that such regulations would involve a “yes, and” scenario. They wouldn’t in any way take the place of the requirement for other sorts of AI regulation on the hazards of present models. He remarked that it would be unfortunate if the focus on frontier models overshadowed those other conversations.
However, according to West, Weinstein, and other people, this is exactly what is taking place. Even a few years ago, West informed that the field of AI safety was far more varied. Now? We are not discussing how these systems affect workers or how they affect the labor market. We’re not discussing environmental issues. It makes sense why the landscape of policy ideas would collapse under strain, undermining the foundation of a robust democracy, when resources, expertise, and influence have concentrated so strongly in a small number of firms and policy officials are steeped in their own cocktail of worries.
