A World Divided Over Artificial Intelligence

In November 2023, a number of countries issued a joint communiqué promising strong international cooperation in reckoning with the challenges of artificial intelligence. Startlingly for states often at odds on regulatory matters, China, the United States, and the European Union all signed the document, which offered a sensible, wide-ranging view on how to address the risks of “frontier” AI—the most advanced species of generative models exemplified by ChatGPT. The communiqué identified the potential for the misuse of AI for “disinformation” and for the kindling of “serious, even catastrophic” risks in cybersecurity and biotechnology. The same month, U.S. and Chinese officials agreed to hold talks in the spring on cooperation over AI regulation. These talks will also focus on how to handle the risks of the new technology and ensure its safety.

Through multinational communiqués and bilateral talks, an international framework for regulating AI does seem to be coalescing. Take a close look at U.S. President Joe Biden’s October 2023 executive order on AI; the EU’s AI Act, which passed the European Parliament in December 2023 and will likely be finalized later this year; or China’s slate of recent regulations on the topic, and a surprising degree of convergence appears. They have much in common. These regimes broadly share the common goal of preventing AI’s misuse without restraining innovation in the process. Optimists have floated proposals for closer international management of AI, such as the ideas presented in Foreign Affairs by the geopolitical analyst Ian Bremmer and the entrepreneur Mustafa Suleyman and the plan offered by Suleyman and Eric Schmidt, the former CEO of Google, in the Financial Times in which they called for the creation of an international panel akin to the UN’s Intergovernmental Panel on Climate Change to “inform governments about the current state of AI capabilities and make evidence-based predictions about what’s coming.”

But these ambitious plans to forge a new global governance regime for AI may collide with an unfortunate obstacle: cold reality. The great powers, namely, China, the United States, and the EU, may insist publicly that they want to cooperate on regulating AI, but their actions point toward a future of fragmentation and competition. Divergent legal regimes are emerging that will frustrate any cooperation when it comes to access to semiconductors, the setting of technical standards, and the regulation of data and algorithms. This path doesn’t lead to a coherent, contiguous global space for uniform AI-related rules but to a divided landscape of warring regulatory blocs—a world in which the lofty idea that AI can be harnessed for the common good is dashed on the rocks of geopolitical tensions.

CHIPS ON THEIR SHOULDERS

The best-known area of conflict related to AI is the ongoing duel between China and the United States over global semiconductor markets. In October 2022, the U.S. Commerce Department issued its first comprehensive licensing regime for the export of advanced chips and chip-making technology. These chips are needed to manufacture the devices that can run the cutting-edge AI models used by OpenAI, Anthropic, and other firms on the technological frontier. The export controls apply not just to U.S. companies but to any manufacturer that uses such U.S. software or technology; in practice, Washington’s export-control regulations have a global remit. In August 2023, China countered with its own export controls on the rare minerals gallium and germanium—both necessary components for manufacturing chips. Two months later, the Biden administration toughened its earlier regulations by expanding the range of covered semiconductor products.

Tit-for-tat competition over semiconductors is possible because international trade law under the World Trade Organization does not sufficiently constrain governments from instituting export controls. The body has rarely addressed the issue in the past. And since former U.S. President Donald Trump neutered the WTO’s appellate body in 2018 by blocking the appointment of new members, there has been little prospect of new formal rules that can be credibly enforced by an authoritative global institution. As a result, these salvos in the chip war between China and the United States are eroding free trade and setting destabilizing precedents in international trade law. They will likely work as a complete substitute for such law in the near term, guaranteeing lower levels of trade and greater geopolitical strains.

But the chip war is just the most high-profile front in the gathering contest over AI’s necessary components. A second zone of conflict concerns technical standards. Such standards have long undergirded the use of any major technology: imagine trying to build a railroad across the United States if every state had a different legally mandated gauge for train tracks. The rise of the digital era has seen the proliferation of various kinds of standards to enable the production and purchase of complex products around the world. The iPhone 13, for example, has nearly 200 parts sourced from more than a dozen countries. If these disparate elements are to work together—and make an object that can communicate with cell towers, satellites, and the Internet of Things—they have to share a set of technical specifications. The choice of such standards has profound effects. It determines whether and how innovations can find commercial uses or achieve market shares. As the German industrialist Werner von Seimens said in the late 1800s, “He who owns the standards, owns the market.”

The chip war between China and the United States is setting destabilizing precedents in international trade law.

At present, a series of little-known bodies such as the International Telecommunication Union, the International Electrotechnical Commission, the International Organization for Standardization, and the Internet Engineering Task Force negotiate technical standards for digital technology in general. Based in Geneva and operating as nonprofits or as UN affiliates, these bodies play a major role in setting the terms of global digital trade and competition. Members of these institutions vote on standards by majority rule. To date, those forums have been dominated by U.S. and European officials and firms. But that is changing.

In the last two decades, China has increasingly taken on leadership roles in the technical committees of several of these bodies, where it has unstintingly promoted its preferred standards. Since 2015, it has integrated its own technical standards in the projects of its Belt and Road Initiative, a vast global infrastructure investment program. As of 2019, it had reached 89 standardization agreements with 39 countries and regions. In March 2018, China launched yet another strategy, “China Standard 2035,” calling for an even stronger Chinese role in international standard setting and demanding greater civil-military coordination within China on the choice of standards. Predictably, some industry analysts in the United States have responded by calling for Washington to combat “more proactively . . . Chinese influence over standard-setting bodies.”

This is not the first time technical standards have become ensnarled in geopolitical tensions. In August 2019, U.S. sanctions on the Chinese telecommunications giant Huawei led China to establish its own energy-efficiency standards that were incompatible with Western ones. The result was a fracturing of technical standards for managing how large data centers, which are central to the digital economy, work. In the AI context, markets separated by different technical standards would slow the diffusion of new tools. It would also make it more difficult to develop technical solutions that could be applied globally to problems such as disinformation or deepfake pornography. In effect, the problems that great powers have identified as important to jointly address would become harder to solve.

Divisions over AI-related technical standards have already emerged. The EU’s AI Act, for example, mandates the use of “suitable risk management measures.” To define this term, the act looks to three independent standard-setting organizations that would develop and promulgate context-specific standards regarding AI safety risks. It is telling that the three bodies specified in the legislation to date are European, not the international ones mentioned above. This seems a quite conscious effort to distinguish European regulation from its U.S. and Chinese counterparts. And it promises the Balkanization of standards pertaining to AI.

THEIR DARK MATERIALS

Geopolitical conflict is not just shaping a new international regulatory landscape for the physical commodities that make up AI. It is also sharpening divides over the intangible assets needed for the technology. Again, the emerging legal regime entrenches a divided world order in which broad-based, collective solutions are likely to fail.

The first important intangible input of AI is data. AI tools such as ChatGPT are built on massive pools of data. To succeed, however, they also need more targeted batches of data. Generative AI tools, which are able to produce paragraphs of text or extended video based on brief prompts, are incredibly powerful. But they are often unsuited to highly specific tasks. They must be fine-tuned with smaller, context-specific data sets to do a particular job. A firm using a generative AI tool for its customer-service bot, for example, might train such an instrument on its own transcripts of consumer interactions. AI, in short, needs both large reservoirs of data and smaller, more bespoke data pools.

Companies and countries will therefore invariably compete over access to different kinds of data. International conflict over data flows is not new: the United States and the EU repeatedly clashed over the terms under which data can cross the Atlantic after the EU’s Court of Justice struck down, in 2015, a safe harbor agreement that had allowed companies to move data between servers in the United States and Europe. But the scale of such disagreements is on the rise now, shaping how data will flow and making it harder for data to cross national borders.

The emerging legal regime regarding AI will frustrate broad-based, collective solutions.

Until recently, the United States promoted a model of free global data transfers out of a commitment to open markets and as a national security imperative—a more integrated world, officials believed, would be a safer one. Washington was aggressive in its use of bilateral trade deals to promote this vision. In contrast, European law has long reflected greater caution about data privacy. For their part, China and India have enacted domestic legislation that mandates, in different ways, “data localization,” with greater restrictions on the flow of data across borders.

Since AI swept to center stage, these views have shuffled. India recently relaxed its prohibition, suggesting that it will allow greater data flows to other countries—thus giving it greater sway over the terms of global digital trade. China also seems to be easing its localization rules as its economy sputters, allowing more companies to store data outside China’s borders. But startlingly, the United States is moving in the opposite direction. U.S. politicians who were worried about the social media app TikTok’s relationship with the Chinese government pressured the company to commit to limiting data flows to China. (TikTok, by its own admission, has honored this commitment somewhat inconsistently.) In October 2023, the U.S. trade representative announced that the federal government was dropping the country’s long-standing demands at the WTO for the protection of cross-border data flows and prohibitions on the forced localization of data. If Washington maintains this path, the world will have lost its principal advocate of free data flows. More data localization would likely ensue.

Finally, global competition is starting to emerge over whether and when states can demand the disclosure of the algorithms that underlie AI instruments. The EU’s proposed AI Act, for instance, requires large firms to provide government agencies access to the inner workings of certain models to ensure that they are not potentially damaging to individuals. Similarly, recent Chinese regulations regarding AI used to create content (including generative AI) requires firms to register with authorities and limits the uses of their technology. The U.S. approach is more complex—and not entirely coherent. On the one hand, Biden’s executive order in October 2023 demands a catalog of disclosures about “dual-use foundation models”—cutting-edge models that can have both commercial and security-related uses. On the other hand, trade deals pursued by the Trump and Biden administrations have included many provisions prohibiting other countries from mandating in their laws any disclosure of “propriety source code and algorithms.” In effect, the U.S. position seems to demand disclosure at home while forbidding it overseas.

Even though this kind of regulation regarding algorithms is in its infancy, it is likely that countries will follow the path carved by global data regulation toward fragmentation. As the importance of technical design decisions, such as the precise metric an AI is tasked with optimizing, becomes more widely understood, states are likely to try to force firms to disclose them—but also to try to prohibit those firms from sharing this information with other governments.

THINGS FALL APART

In an era of faltering global resolve on other challenges, great powers had initially struck an optimistic note in grappling with AI. In Beijing, Brussels, and Washington, there seemed to be broad agreement that AI can cause potentially grave harms and that concerted transnational action was needed.

Countries are not, however, taking this path. Rather than encouraging a collective effort to establish a clear legal framework to manage AI, states are already engaged in subtle, shadowy conflicts over AI’s material and intangible foundations. The resulting legal order will be characterized by fracture and distance, not entanglement. It will leave countries suspicious of one another, sapping goodwill. And it will be hard to advance proposals for better global governance of AI. At a minimum, the emerging regime will make it more difficult to gather information and assess the risks of the new technology. More dangerously, the technical obstacles raised by the growing legal Balkanization of AI regulation may make certain global solutions, such as the establishment of an intergovernmental panel on AI, impossible.

A fragmented legal order is one in which deeply dangerous AI models can be developed and disseminated as instruments of geopolitical conflict. A country’s efforts to manage AI could easily be undermined by those outside its borders. And autocracies may be free to both manipulate their own publics using AI and exploit democracies’ free flow of information to weaken them from within. There is much to be lost, then, if a global effort to regulate AI never truly materializes.