The White House is fixated on China copying U.S. AI. Experts say that’s the wrong threat

The White House appears preoccupied with Chinese entities replicating U.S. artificial intelligence models, but experts argue this focus is misplaced. Recent statements from the White House suggest the agency believes Chinese companies developed sophisticated methods to distill U.S. AI models, including Anthropic’s Fable, to create their own models.

However, the White House has not provided any evidence to support these claims, and the Office of Science and Technology Policy (OSTP) has not responded to inquiries about this assertion.

Distillation is a standard practice within the AI industry, where companies pay for access to pre-trained models and learn from their outputs. Critic Nathan Lambert, founder of the posttraining research team at the Allen Institute for AI, argues the distinction between routine model development and "adversarial distillation" is often blurred.

Attackers might try to bypass API rules, manipulate models to reveal sensitive information, or obscure the identity and purpose of the requesting organization. However, merely using an API as a paid service is considered legitimate.

Lambert contends most of the evidence does not support the claim that Chinese firms have engaged in large-scale distillation of U.S. models, nor does it suggest this practice has significantly narrowed the gap between Chinese and American labs. He notes officials may have received undisclosed classified briefings on this matter.

Furthermore, Lambert points out that distillation may be losing relevance as companies increasingly rely on reinforcement learning and other posttraining techniques. The White House's heightened focus on distillation may be part of a broader strategy to contain Chinese advancements, rather than a response to a concrete security threat.

Kristy Loke, a fellow at the MATS AI safety research program, says this focus reflects a desire to view China's progress as contingent on American technology rather than evidence of Chinese labs innovating independently. By emphasizing distillation, the narrative suggests that dominant U.S. labs discovered the sole viable path to advanced AI, potentially obscuring other competitive approaches. This framing strengthens the position of companies advocating for government protection from emerging rivals.

Critics argue that the political focus on distillation exaggerates the ease with which open-weight models can be turned into cyberweapons. While downloading a model and disabling its safeguards to replicate the capabilities of closed systems can be challenging and costly, requiring substantial resources and access to cloud capacity, the more immediate concern should be the autonomous behavior of the most advanced models developed by U.S. companies.

Lambert argues policymakers are overreacting to potential harms from model copying, while U.S. AI firms are making significant strides in improving security.