Anthropic has released a comprehensive paper analyzing the competition between the United States and China in the field of artificial intelligence. Titled '2028: Two scenarios for global AI leadership,' the document spans over 2,000 words and outlines two potential futures for the global AI landscape by the year 2028, when transformative AI systems are expected to have arrived.
Scenario One: Democratic Leadership Maintained
In the first scenario, the United States successfully defends its compute advantage. Policymakers have acted to tighten export controls further, disrupt China's distillation attacks, and accelerate the adoption of AI among democracies. In this world, democratic nations set the rules and norms around AI. This scenario also offers the best opportunity for engaging with China on AI safety, a goal Anthropic supports to the extent possible.
Scenario Two: Authoritarian Triumph
The second scenario depicts a world where America chooses not to act. Policymakers have not tightened loopholes regarding the Chinese Communist Party's access to compute, allowing Chinese AI firms to quickly catch up and even overtake the United States. In this future, AI norms and rules are shaped by authoritarian regimes, and the most capable models enable automated repression at scale. Anthropic warns that this authoritarian triumph would be built on American compute.
The Importance of Compute Advantage
The paper emphasizes that access to advanced computer chips, or compute, is the most critical ingredient for developing AI. Since the most capable chips are produced by American companies, U.S. export controls currently limit China's supply. These controls have been largely successful, but Chinese labs have managed to stay close through talent, exploitation of loopholes, and large-scale distillation attacks that illicitly extract innovations from American companies.
Threat of Authoritarian AI
Anthropic argues that if authoritarian regimes like the Chinese Communist Party lead in AI, the technology could be used for unprecedented repression and military advantage. The CCP already uses AI for censorship, surveillance, and hacking. Frontier AI systems could automate these activities at a scale far beyond human capabilities.
Policy Recommendations
To secure democratic leadership, Anthropic recommends that the U.S. and its allies act now to tighten controls on advanced compute access to Chinese labs, disrupt distillation attacks, and accelerate AI adoption in democracies. The window of opportunity to lock in a 12-24 month lead in frontier capabilities may not remain open for long.
Four Fronts of Competition
The paper identifies four fronts in the AI competition: intelligence (model capabilities), domestic adoption, global distribution, and resilience. Intelligence is deemed the most important, as it drives adoption and distribution. However, if China integrates near-frontier AI more effectively into its economy and security apparatus, it could overcome an intelligence deficit.
Current State of Play
Democracies currently hold a substantial lead in compute, thanks to innovation from companies like NVIDIA, AMD, and TSMC, as well as bipartisan U.S. export controls. However, Chinese labs have remained competitive by smuggling chips and accessing offshore data centers, as well as through distillation attacks. Anthropic notes that China's own AI leaders confirm the impact of export controls and the critical need for U.S. chips.
Opportunities for Engagement
Anthropic supports international AI safety dialogue with Chinese experts when possible, but stresses that productive engagement is most likely when the U.S. maintains a large capabilities advantage. The paper concludes that a lead for democracies is essential to avoid authoritarian AI leadership and protect national security interests.
