US China AI Control Divergence Drives Trump Xi Summit Agenda on Model Safety

Read Time 6 minutes Tags AI Governance China US Policy AI Safety Trump Xi Agentic AI Public Trust In the US China tech race Beijing has underscored AI control from the start the US seems only now to be taking it seriously US President Donald Trump and Chinese President Xi Jinping will likely discuss AI at a summit in Beijing this week There is a striking difference in how eager the public in China and the US are for AI adoption This report is from this week CNBC The China Connection newsletter The big story A robotic voice warned me and others at a Hangzhou street intersection that a scooter driver did not have a helmet on even though I saw the rider wearing one Regardless the city and others are ploughing ahead in testing robot police officers The national cybersecurity regulator on Friday published guidelines for ensuring safe use of agentic AI It is a reminder that in the US China tech race Beijing has underscored AI control from the start the US seems only now to be taking it seriously As comparisons to the Cold War nuclear threat grow hopes are rising that Trump and Xi will talk about AI cooperation in Beijing this week Given concerns about the latest AI models we are willing to explore channels of deconfliction senior US officials told reporters in a briefing ahead of the planned summit The stakes are rising sharply US based Anthropic rolled out cyber focused Mythos to select clients in the last few weeks a model Chinese state media has noted for its unprecedented capabilities in cyberattacks Meanwhile the latest version of China open sourced DeepSeek model has weaned itself further off US chips Technical and policy analysis One Regulatory asymmetry defines negotiation baselines China national cybersecurity regulator published guidelines for ensuring safe use of agentic AI before broad enterprise deployment Agentic systems can plan execute multi step tasks and call external tools which raises control risk Beijing approach is top down compliance first The US has relied on voluntary commitments and NIST frameworks until now Trump is expected to unveil executive action on AI safety as soon as Monday That marks a pivot from laissez faire to mandated guardrails The summit will test whether the US can adopt enforceable rules without ceding model leadership Two Public adoption gap creates different risk profiles In the US a backlash over defense tech supplier Palantir manifesto last month reinforced fears about the dark side of AI Teachers warn about the impact of AI on literacy There is more concern about the technology in the US than in many parts of the world 50 percent of Americans are wary of it versus an average of 33 percent in other G7 countries according to Pew surveys In China fear of being displaced by AI has motivated people and businesses to adopt the tech more quickly while courts are ruling in favor of human workers After initially taking months to greenlight local ChatGPT alternatives Beijing has come around to encouraging adoption while issuing guidance on AI use in schools This divergence matters for safety talks A risk averse US public demands slower rollout and stronger safeguards A pro adoption Chinese public enables faster data collection and real world testing of agentic systems like robot police officers in Hangzhou That gives Beijing more operational telemetry on failure modes which is leverage in any safety dialogue Three Model capability and hardware decoupling are linked The latest version of China open sourced DeepSeek model has weaned itself further off US chips DeepSeek said its new model can use Huawei chips for inference while still training on Nvidia The US Anthropic Mythos model shows unprecedented capabilities in cyberattacks If China and the US get involved in an AI arms race then it is bad not just for both countries but for all humanity said Hai Zhao a director of international political studies at the Chinese Academy of Social Sciences The technical bridge is dual use control Both sides have frontier models that can automate vulnerability discovery Both sides are building domestic chip stacks to reduce supply chain risk The summit could produce channels of deconfliction similar to nuclear hotlines That requires shared definitions of red lines telemetry exchange for incidents and verification of model capabilities Four Education and research pipelines reinforce the split Hangzhou Zhejiang University and Shanghai Jiao Tong University surpassed Harvard to take the top spots in a ranking of universities by scientific performance Many of China high tech startups including DeepSeek have their roots in Zhejiang University The school alumni association has released a list of 10 companies such as quantum computing player Logistics Bits which it says could be the next DeepSeek Beijing mandate to achieve an AI penetration rate of over 70 percent in key industries by next year accelerates deployment In the US AI feels more behind the scenes said Hunter Roskom a student from the University of Wisconsin Madison at Zhejiang University In China it is integrated into daily life What to watch First scope of Trump executive action Will it mandate incident reporting for models above a certain compute threshold Will it require model cards with safety evals for agentic systems Second verification mechanism for AI cooperation A global treaty to regulate the use of AI in the military was proposed by Chinese scholars Cold War arms control relied on inspection This era may need cryptographic proofs of training data and continuous monitoring APIs Third enterprise impact If US and China align on safety baselines then multinationals can unify compliance If they diverge then firms must maintain two AI stacks with different guardrails and audit trails For risk and engineering leaders map your exposure now Classify models by autonomy level agentic versus chat Document chip dependencies for training and inference Prepare for jurisdiction specific safety logs and red team disclosures The Trump Xi summit is not just diplomacy It is the first draft of international AI control protocols What safety controls should be non negotiable for agentic AI Share your technical requirements in the comments