As Donald Trump and Xi Jinping prepare to meet, the real groundwork for a potential artificial intelligence détente has been laid not in the Oval Office or Zhongnanhai, but in the seminar rooms of think tanks and university campuses. An informal network of “Track Two” dialogues has spent the last year convening American and Chinese scientists, policy analysts, and former officials to navigate the shared existential risks of advanced AI systems — from biosecurity threats to loss of human control — creating a rare channel of communication that official diplomacy has struggled to replicate.
The Shadow Architecture of AI Governance
While headlines focus on export controls on high-bandwidth memory and the compute arms race, a parallel track has been operating in the margins. These dialogues — hosted by institutions ranging from the Brookings Institution and the Center for a New American Security to Tsinghua University’s Institute for AI International Governance — operate without the binding authority of treaties. Their currency is trust, built over years of academic collaboration before geopolitics froze formal channels.
Participants describe a shift in tone. Early meetings in 2023 were dominated by mutual suspicion; Chinese delegates arrived with strict talking points on US containment strategies, while Americans pressed hard on intellectual property theft. But as frontier models demonstrated unexpected capabilities — and unexpected failures — the conversation migrated toward technical safety standards, evaluation protocols, and the terrifying prospect of “unaligned” systems escaping human oversight.
From Compute Caps to Kill Switches
The agenda has matured rapidly. Recent sessions have moved beyond abstract principles into the weeds of technical governance: watermarking synthetic content, standardised red-teaming methodologies, and the feasibility of international compute thresholds that could trigger mandatory safety audits. One participant, a senior researcher at a leading US lab who requested anonymity to speak freely, characterised the evolution as “moving from ‘should we talk’ to ‘how do we verify a model won’t engineer a pandemic’.”
Chinese counterparts have signalled surprising openness on verification regimes. At a closed-door workshop in Geneva last autumn, delegates from the Shanghai AI Laboratory and the Beijing Academy of Artificial Intelligence presented a framework for mutual model evaluation that bore striking resemblance to proposals circulating in Washington. The convergence suggests a shared recognition that the alternative — an unregulated race to the bottom — serves neither capital.
The Trump-Xi Variable
The upcoming leader-level meeting injects new urgency. The Trump administration’s first term saw the collapse of nearly every bilateral science and technology agreement; advisers now signal a willingness to compartmentalise AI safety from the broader tech war. Whether that compartmentalisation survives contact with Congress — where the Select Committee on the Chinese Communist Party views any engagement as capitulation — remains the central political question.
Track Two veterans are under no illusions. “We’re not drafting treaties,” said a frequent participant from a West Coast university. “We’re building the vocabulary and the personal relationships so that when the political window opens — however briefly — the technical solutions are already on the table.” That window may be narrower than anyone hopes. The next generation of models, currently training on clusters measured in hundreds of thousands of GPUs, will not wait for diplomatic calendars.
Why it Matters
The Track Two process represents the only functioning shock absorber between the world’s two AI superpowers; without it, a misinterpreted model deployment or a runaway autonomous agent could trigger a crisis that no hotline can de-escalate. These conversations have already produced a shared technical lexicon for risk that official channels rejected, proving that scientists can outpace politicians when the stakes are existential. If the Trump-Xi summit yields even a modest joint statement on AI safety standards, it will be because the vocabulary was written years earlier in rooms without flags.
