- Trump and Xi will discuss artificial intelligence when Xi arrives next Thursday for his first state visit to Washington since 2015, alongside Taiwan, Iran and trade. Treasury Secretary Scott Bessent, who leads the US AI track, meets Xi top economic official He Lifeng in New York this weekend to prepare.
- Bessent said the US remains the AI leader and is open to discussing shared risks and avoiding a split between the two systems, adding that talks will cover both open and closed weight models. American labs including OpenAI and Anthropic keep their strongest models closed behind paid access, while Chinese developers such as DeepSeek release open-weight systems anyone can download.
- Chinese minister of state security Chen Yixin set out Beijing view of AI risk in a party publication this week, naming political security first. He warned the technology could let hostile forces mass-produce rumours, conduct cognitive warfare and provoke unrest, and prescribed tighter party control over the internet and over data rather than technical safeguards.
- The 2023 attempt at a formal AI dialogue under the Biden administration is the cautionary precedent. Beijing assigned its foreign ministry rather than a technical body, which US officials involved said limited the substance of the exchange.
What Happened?
The two governments approach AI safety from incompatible starting points. Washington debate centres on models escaping human control, autonomous weapons and tools that could help a rogue actor build a weapon, with Trump playing down the need for regulation while leading American executives call for a slowdown. Beijing frames the risk as a threat to the political order. Chris McGuire, formerly of the National Security Council under President Biden and now at the Council on Foreign Relations, said China has consistently rejected serious collaboration on AI safety and is unlikely to accept anything slowing its development while its capabilities sit only months behind the leading US labs. Signals from Beijing are mixed. Party ideologue Wang Huning appeared to accept the value of a bilateral discussion when he met American business and think tank representatives last month. Chris Slevin of the US-China Economic and Security Review Commission, who visited Beijing in July, said officials spent roughly half of an hour-long foreign ministry meeting on AI, asking how the US was organising the talks and whether chip export controls would be included. Requests to visit DeepSeek and Moonshot were refused. Slevin said Chinese officials were markedly more self-assured than three years ago and stressed that the two countries had reached rough parity.
Why It Matters?
Bessent reference to open and closed weight models is the most revealing line, and it indicates the talks are about distribution rather than safety. China releases its best models freely so they spread worldwide and undercut American containment, which is a commercial and geopolitical strategy rather than a research philosophy. Any agreement constraining open-weight release would remove China principal instrument for competing where it cannot match US capability, and Beijing will read it that way. Gary Rieschel of Qiming Venture Partners identified the asymmetry that could prove most consequential: Chinese officials believe the Great Firewall protects them, but that system was built to control what a billion-plus users can see, not to repel intrusions into networks. A model that probes for weaknesses on its own is not a website that can be blocked. That gap means the two sides are not only defining safety differently but defending against different threats, and Beijing may be underprotected against precisely the risk Washington worries about. The Cold War analogy that Robert Hormats, a National Security Council member under Kissinger, offers is instructive but incomplete. Nuclear arms control worked partly because warheads are countable and verifiable. Model weights are copyable files, so verification has no equivalent mechanism, which makes any AI agreement structurally weaker than the arms treaties it is being compared to.
What Next?
The Bessent and He Lifeng meeting in New York this weekend is the first concrete event, and whether AI produces any joint language there will indicate how much has been pre-negotiated before Thursday. The summit itself carries a crowded agenda, so AI competing against Taiwan, Iran and trade for time is itself a measure of priority. The clearest signal to watch afterwards is whom Beijing designates to lead any dialogue, since assigning a technical body rather than the foreign ministry would mark a real change from 2023 and anything else repeats a process both sides already know produces little. Also watch whether chip export controls enter the agenda, because Chinese officials have been probing that question specifically, and any linkage between AI safety discussions and export relief would tell you what Beijing actually wants from the process. A communiqué announcing a dialogue is the easy outcome and should not be mistaken for progress.
Source: The Wall Street Journal
