U.S. and China Open AI Safety Talks Ahead of Trump’s State Visit to Beijing
Why It Matters
The prospect of a formal emergency communication channel between Washington and Beijing on artificial intelligence represents one of the most consequential technology-policy developments in years. With AI systems now capable of penetrating financial networks, government databases, and critical infrastructure, the absence of any bilateral framework leaves both nations — and the global economy — exposed to scenarios neither side fully understands.
The discussions, spurred in part by the debut of a powerful new AI model, signal a notable shift in how the Trump administration views the risks of unchecked AI development, marking the first time since taking office that the White House has signaled openness to structured AI dialogue with China.
What Happened
Preliminary discussions have taken place ahead of President Trump’s state visit to China this week aimed at reviving a bilateral emergency channel for AI governance, officials confirmed. The conversations were prompted largely by shared concern over Mythos, a new model released by Anthropic that has been described across the technology industry and within government as possessing unprecedented capabilities as a cyberweapon — one capable of infiltrating digital communications, government databases, financial institutions, and healthcare systems.
A senior administration official told reporters that the White House is exploring a dedicated communication channel for AI issues, similar to frameworks the U.S. maintains with China in other high-stakes domains. The official acknowledged that the format and formality of any such channel remain undecided.
The backdrop stretches back several years. In 2023, at a summit in Woodside, California, President Xi Jinping and his foreign minister signaled support for establishing a bilateral AI safety channel. That led to formal talks in Switzerland in 2024, which produced little progress — Chinese negotiators dismissed American concerns about AI risk as largely theoretical and redirected the conversation toward U.S. export controls, which Beijing views as an effort to suppress Chinese technological development.
Despite the friction, both sides reached a limited agreement in November 2024 in Peru, committing to keep AI systems out of the command and control of nuclear weapons — a baseline accord that set a precedent for further engagement.
By the Numbers
- 2023: First high-level U.S.-China AI safety talks held in Woodside, California
- 2024: Switzerland talks ended without a substantive framework; Peru accord reached on nuclear AI restrictions
- 5 major U.S. technology firms — OpenAI, Google, Anthropic, Microsoft, and Meta — are actively racing toward artificial general intelligence (AGI)
- April 2026: Anthropic’s Mythos model released, triggering alarm in both Washington and Beijing over its potential as a cyberweapon
The Strategic Divide
The two countries approach AI development from fundamentally different strategic premises. American firms are competing to be first to achieve AGI — broadly defined as a model capable of performing any intellectual task a human can — operating on the theory that the first to cross that threshold will gain a self-reinforcing technological advantage that compounds over time.
China’s approach, by contrast, is state-directed and sector-specific. Chinese companies are training AI to improve targeted industrial and commercial applications rather than pursuing a single generalized model. “The Chinese believe there is no single race, but multiple races,” said Scott Kennedy, a senior advisor on Chinese business and economics at the Center for Strategic and International Studies.
That divergence complicates any arms-control-style agreement. Each side would arrive at the table with different definitions of what an AI “limit” even means — and deep suspicion about the other’s motives. “Right now, there is almost no support from U.S. policymakers to engage in formal discussions on AI governance with China,” said Aalok Mehta, director of the Wadhwani AI Center at CSIS, adding that the dominant American view treats AI as a winner-takes-all competition.
Zoom Out
The renewed interest in bilateral AI safety talks echoes earlier Cold War-era arms control diplomacy, where adversaries with incompatible ideologies nonetheless established communication channels to reduce the risk of catastrophic miscalculation. AI-enabled surveillance tools are already reshaping law enforcement operations domestically, illustrating how rapidly the technology is being integrated into high-stakes systems with limited regulatory oversight.
The challenge is compounded by the fact that neither government fully controls the private-sector actors driving AI development. Anthropic, the company behind Mythos, is a private firm — and its model’s release appears to have accelerated government urgency on both sides more than any diplomatic initiative had managed to do.
What’s Next
Whether a formal announcement will emerge from Trump’s state visit to China this week remains uncertain. Officials indicated the discussions are still exploratory, focused on whether any channel should be formal or informal, and what its scope would cover. Former National Security Advisor Jake Sullivan, who oversaw earlier AI diplomacy under the Biden administration, said he urged the incoming Trump team to continue the dialogue — and noted that administration posture has shifted meaningfully in recent weeks. Any eventual framework would face significant verification challenges and political resistance from voices in Washington who argue that engagement on AI governance cedes strategic ground to a competitor.