OpenAI executive urges caution in China-US AI talks, warns of distillation - South China Morning Post

The Unspoken Risks in US-China AI Collaboration

Caroline Zier, OpenAI’s deputy lead for national security, warned U.S. officials to remain “clear-eyed” during upcoming bilateral AI talks with China, citing risks of “unauthorized distillation” and divergent definitions of AI safety. The remarks delivered at the Hudson Institute highlighted tensions between Washington and Beijing concerning intellectual property and model integrity issues.

Zier’s comments emerged ahead of a planned summit between President Trump and Chinese leader Xi Jinping, aimed at advancing a September framework for AI risk management.

The ‘Unauthorized Distillation’ Controversy

Unauthorized distillation—where malicious actors use outputs from a “teacher” model to train a “student” model—was flagged as a primary threat by Zier. She noted that Chinese labs frequently engage in this practice, stripping safety mechanisms and misappropriating protected intellectual property.

The U.S. government has previously accused Chinese firms like DeepSeek and Alibaba of “industrial-scale distillation,” a claim Beijing denies.

According to Michael Ortiz of Reflection AI, about 80% of U.S. AI startups utilize Chinese open-weight models. AI startups rely on Chinese open-weight models due to their cost‑effectiveness, creating a “strategic vulnerability” if access were restricted.

The Geopolitical Chessboard

The tension reflects broader U.S.-China competition over AI infrastructure. While both nations invest heavily in AI, their priorities differ: the U.S. emphasizes guardrails against model misuse, whereas China’s approach to AI infrastructure is described as a “different race.”

Source-provided image accompanying OpenAI executive warns of risks in US-China AI talks and model distillation — Photo: aiunderstanding.org

Sam Bresnick of Georgetown University’s Centre for Security and Emerging Technology echoed Zier’s concerns, questioning the efficacy of a U.S.-China AI “hotline” given historical delays in Chinese information sharing.