Anthropic warns China AI parity by 2028

Anthropic published a policy paper on May 14 laying out two possible 2028 outcomes for the U.S.-China AI race, arguing that Washington still has time to preserve a lead but only if it closes specific gaps now. The paper, titled “2028: Two scenarios for global AI leadership,” says access to advanced chips remains the central bottleneck in frontier AI development and that current U.S. export controls have been “incredibly successful.” Anthropic framed the warning in stark terms. The company said Chinese AI labs have stayed near the U.S. frontier through talent, “exploiting loopholes” in export controls and what it described as “large-scale distillation attacks” that extract capabilities from American models. It said the next few years will determine whether democratic governments or authoritarian ones shape the rules for advanced AI systems. (anthropic.com) ### What exactly did Anthropic publish on May 14? Anthropic’s research page lists the document as a policy publication dated May 14, 2026. The post presents two scenarios for global AI leadership in 2028 and says Anthropic expects “transformative AI systems” to have arrived by then. The company’s first scenario assumes the United States and its allies tighten export controls, disrupt distillation, and speed adoption of AI across democratic countries. (anthropic.com) The second assumes policymakers do not act, allowing Chinese firms to catch up to or overtake the frontier, according to the paper. ### Which pathways does Anthropic say could help China close the gap? (anthropic.com) Anthropic’s paper centers on compute. The company says the “most important ingredient” for developing advanced AI is access to the chips used to train models, and it argues that America’s edge comes from control over the most capable semiconductors and related infrastructure. The document points to three routes that could erode that advantage: loopholes around chip export controls, offshore or cross-border access to computing infrastructure, and distillation of U.S. model capabilities. (anthropic.com) Anthropic had already signaled similar concerns in a September 2025 policy update that tightened sales restrictions for entities linked to China and warned that such entities could use access to advance their own AI development through distillation. ### What is “distillation,” and why is it in the paper? Anthropic used the term “distillation attacks” to describe efforts to extract the behavior or capabilities of leading U.S. AI systems and reproduce them in another model. In the May 14 paper, the company says those attacks have helped Chinese labs build models “close in intelligence” to American ones. The paper does not present the issue as a theoretical future risk. (anthropic.com) Anthropic links distillation directly to current competition and argues that preventing unauthorized model extraction should sit alongside chip controls as a policy priority. That position is consistent with the company’s earlier public statements on restricting access from unsupported regions and subsidiaries tied to them. ### How does this fit with Anthropic’s broader China policy stance? Anthropic has spent the past year pairing commercial restrictions with policy advocacy. In September 2025, the company said firms controlled from unsupported regions such as China would be barred from using its services even through overseas subsidiaries, citing legal, regulatory and security risks. (anthropic.com) The May 14 paper extends that argument from customer access to national strategy. Anthropic says democracies should not “make it easier for the CCP to catch up,” while also saying it would support engagement with China on AI safety where possible. ### What happens next? May 14 is now the key reference point for this debate because Anthropic has put a date and a policy timeline on it. (anthropic.com) The paper says there is only a limited period to set the terms of competition before 2028, and its recommendations are published in full on Anthropic’s research site under the company’s policy section. (anthropic.com)