What happened
Caroline Zier, OpenAI’s deputy lead for national security, stated at a Hudson Institute event that the US must remain 'clear-eyed' regarding information exchanges with China during upcoming bilateral AI talks. Zier expressed skepticism that the US and China share a common understanding of AI risks, specifically regarding 'safety' and 'loss of control.' She also characterized unauthorized model —where actors use a 'teacher' model's outputs to train a 'student' model—as a significant threat to intellectual property, noting that such activities often strip away safety . The discussion occurred ahead of expected meetings between US Treasury Secretary Scott Bessent and Chinese Vice-Premier He Lifeng.
Caroline Zier, OpenAI’s deputy lead for national security, emphasized the need for a cautious approach to bilateral AI talks with China during a panel at the Hudson Institute. Zier, a former special assistant to the US secretary of defence, argued that the US should be highly aware of the information it shares, noting that the Chinese Communist Party likely interprets concepts like 'safety' and 'loss of control' differently than US counterparts.
The discussion addressed the ongoing issue of 'unauthorized ,' which Zier distinguished from authorized training. She stated that unauthorized distillation is frequently used to strip away safety and misappropriate protected intellectual property. This follows previous accusations from US intelligence and cybersecurity agencies targeting Chinese firms like Alibaba and DeepSeek, which Beijing has formally rejected as unfounded.
Sam Bresnick, a research fellow at Georgetown University’s Centre for Security and Emerging Technology, participated in the same discussion and expressed doubt regarding the effectiveness of a US-China AI 'hotline.' He cited a historical lack of rapid information sharing from the Chinese side as a primary reason for his skepticism.
The panel also touched upon the widespread adoption of Chinese open-weight models by Western startups. Michael Ortiz of Reflection AI noted that these models are highly capable and cost-effective, with estimates suggesting that nearly 80% of US AI startups are currently utilizing them. Both Zier and Bresnick warned that this trend creates a potential 'rip and replace' risk if the Chinese government were to restrict access to these technologies in the future.
Why it matters
The comments highlight the growing tension between the US and China as both nations attempt to establish a formal AI dialogue mechanism. While technical teams are working to implement a framework agreed upon in September, experts and industry leaders remain divided on the efficacy of these channels. The debate over 'model ' is central to this friction, as US officials and companies view the practice as a form of IP theft and a security risk, while Chinese entities have rejected these claims as attempts to maintain technological dominance. Furthermore, the reliance of Western startups on cost-effective Chinese open-weight models creates a complex dependency that complicates potential future restrictions or 'rip and replace' scenarios, posing long-term strategic risks for US-based AI infrastructure.
The tension between the US and China over AI development is increasingly focused on the intersection of national security and commercial competitiveness. The debate over model serves as a proxy for broader concerns about IP protection and the integrity of AI safety protocols.
The reliance of US startups on Chinese open-weight models creates a strategic vulnerability. While these models offer immediate cost and performance benefits, they introduce a dependency on foreign technology that could be subject to sudden policy shifts or state-level control by the Chinese government.
The lack of a shared 'risk vocabulary' between the two nations suggests that even if communication channels are established, they may fail to prevent misunderstandings or address the core security concerns raised by US officials. This divergence in perspective makes the prospect of meaningful, collaborative highly uncertain.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Observers should monitor the upcoming meetings between US and Chinese officials, specifically whether the proposed AI 'hotline' or communication channels yield any concrete operational agreements. Additionally, watch for further regulatory or industry-led actions regarding the use of open-weight models from Chinese labs by US startups, as concerns over long-term dependency and potential CCP-imposed controls on these models continue to grow. The effectiveness of the bilateral dialogue in bridging the gap between US and Chinese definitions of AI safety remains a critical uncertainty.
Monitor the outcomes of the upcoming APEC summit in Shenzhen, where US President Donald Trump and Chinese leader Xi Jinping are expected to meet. Any specific agreements regarding the AI dialogue mechanism will be a key indicator of whether the current diplomatic efforts are yielding results.
Watch for potential policy shifts or industry advisories regarding the use of foreign open-weight models. As the debate over dependency intensifies, US policymakers may face pressure to address the prevalence of these models in the domestic startup ecosystem.
Observe whether the Chinese government provides any further response to the ongoing allegations of industrial-scale , as this remains a primary point of contention that could trigger retaliatory measures.