The Information Machine
FollowingDay 1307 sourcesupdated 21 Nov 2026

Chinese distillation of US AI models

The gist

Latest: OpenAI Publishes Distillation Research as US-China AI Flashpoint Intensifies

Distillation allows foreign actors to extract frontier AI capabilities while bypassing the safety measures built around them, posing both security and safety risks. US labs, policymakers, and researchers are actively debating how much technical countermeasures can accomplish and what policy interventions are warranted.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

OpenAI published research warning that distillation can reproduce powerful AI capabilities without the safeguards built around them, and described a layered defense combining model training, safeguards, detection, and enforcement. The findings are intended to sharpen OpenAI's defenses and support safeguards across the industry. The backdrop is a documented pattern of Chinese entities using distillation of US AI models to train domestic rivals, which Reuters described as a significant US-China geopolitical flashpoint. The Trump administration is considering requiring US companies to disclose publicly when they use Chinese AI models, with the approach framed as 'name and shame' pressure rather than new restrictions. House investigators sent letters to Anysphere, whose Cursor coding tool was built on a version of Moonshot AI's Kimi, and to DoorDash, which routes some internal tasks through Kimi. Airbnb runs its customer service agent on Alibaba's Qwen model but contends its AI usage is overwhelmingly on US models and that Chinese models are routed through approved US-based providers. Rep. John Moolenaar, chair of the House select committee on China, argues that hosting Chinese models on US infrastructure does not eliminate associated risks. Reuters separately reported that a US government website deployed an Alibaba model to search proposed federal regulations. Anthropic, OpenAI, and Google began sharing distillation threat intelligence through the Frontier Model Forum in April 2026, though the depth of cooperation remains constrained by uncertainty over what antitrust rules permit competitors to share. RAND concluded that technical counter-distillation measures could in theory raise the cost of distillation or degrade model quality, but are unlikely to change the underlying incentives driving the practice. Analysis published by Zvi Mowshowitz characterizes China's AI advantage as primarily fast-following and distilling US models rather than independent frontier research, with China lacking the chips, funding, and apparent desire to push the frontier independently in the near term. HSBC cut Zhipu's price target by two-thirds, citing in part the strengthening of anti-distillation measures by leading US AI labs. Former National Security Advisor Susan Rice rejected the argument that competition with China justifies accepting existential AI risk. China's top CCP theoretical journal published what was described as its strongest and most urgent language on existential risks, warning AI could threaten human survival.

How it developed
20 November 2026added 21 Nov

OpenAI published research on November 20, 2026 warning that distillation can reproduce powerful AI capabilities while bypassing the safety measures built into source models, and described a layered defense combining model training, safeguards, detection, and enforcement.

OpenAI said the findings are intended to sharpen its own defenses and support safeguards across the industry, against a backdrop Reuters described as a significant US-China geopolitical flashpoint over Chinese entities training domestic rivals by distilling US AI models.

19 November 2026

Analysis published by Zvi Mowshowitz characterized China's AI advantage as primarily distillation-based; former NSA Susan Rice rejected the 'but China' argument for accepting existential AI risk.

18 November 2026

Trump administration reported to be considering disclosure requirements for US companies using Chinese AI models

10 September 2026

Cloud Security Alliance published research on industrial-scale model distillation as a national security problem.

16 July 2026

Institute for Law & AI published guidance on practical defenses against unauthorized distillation, including pattern detection and industry coordination.

Sources
2 more sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy