Trump-Xi Dinner Tests the Politics of an AI

OpenAI CEO Sam Altman plans to attend President Donald Trump’s Sept. 24 state dinner for Chinese President Xi Jinping, CNBC reported, alongside Nvidia CEO Jensen Huang. The seating chart will put two incompatible approaches to artificial intelligence within canapé range of the two governments capable of shaping the global contest.

Altman has endorsed slowing the improvement of frontier AI systems so safety work can catch up. Huang says developers should “run as fast as you can,” pausing when they lose confidence in a product. Trump has called warnings about runaway AI a “hoax” and treats speed as essential to defeating China. Chinese state-run newspaper Global Times, meanwhile, characterized demands for a slowdown as an attempt to contain Chinese development.

A dinner invitation does not turn any of those positions into policy. It does reveal the problem inside the slowdown campaign: once companies ask competitors in other countries to move together, they are asking for diplomacy. OpenAI cannot negotiate reciprocal obligations with Beijing, however impressive its fundraising deck may be.

What a “Slowdown” Would Require

Anthropic CEO Dario Amodei has outlined three steps: place independent evaluators inside AI companies, coordinate standards among democratic countries, then pursue broader global coordination. Altman has made the governmental dependency explicit. In a social media statement quoted by IGN, he said companies could act themselves at first, but would need government help with international coordination.

The sequence makes sense. A laboratory can hire an outside evaluator and delay its own model. A club of laboratories can agree on voluntary tests. Neither can ensure that a rival company, foreign laboratory or newly funded entrant accepts the same cost. Coordination needs agreed thresholds, access for evaluators and consequences when somebody ignores the rules. Otherwise, “slowdown” describes a preference rather than a system.

The industry has begun working on the voluntary layer. OpenAI executive Chris Lehane said OpenAI, Anthropic and Google DeepMind were developing frontier AI standards “with or without government support.” Amodei said Anthropic was discussing better checks with the rest of the industry. Yet Altman also asked the public to trust AI companies to do the right thing, telling a San Francisco conference that OpenAI would slow or stop if safety failed to stay ahead of capabilities.

That promise contains its own audit problem. The company building the model would also decide whether its safeguards were adequate and whether commercial pressure had become dangerous. Independent evaluation attempts to separate referee from player, although the available proposal leaves crucial details unresolved: who chooses the evaluators, what they can inspect, what counts as failure and whether a failed test prevents release.

The Acceleration Camp Has a Safety Argument Too

Huang’s strongest case goes beyond “go fast and hope.” He argues that safety is an engineering responsibility, existing laws already govern product reliability and functionality, and a company should withhold a release whenever it lacks confidence. He says new AI laws and regulations are unnecessary.

Trump and administration officials supply a separate geopolitical argument. They contend that limiting American development could weaken the United States in its competition with China. Those positions overlap in opposing new regulation, but the documented rationales differ: Huang emphasizes developer responsibility and existing product law, while the administration emphasizes national competition.

Meta CEO Mark Zuckerberg offers a related incentive argument. He says laboratories face significant liability if their models cause harm, giving them commercial reasons to invest in alignment. Meta delayed its Muse AI technologies over safety and security concerns, a concrete example of a company withholding a product without an external command.

Pace occupies the slogans; verification occupies the policy dispute. Zuckerberg points to liability and Meta’s internal decision as evidence that incentives can work. Amodei wants outsiders checking whether companies have done enough. One case of voluntary restraint cannot establish how every laboratory will behave under every competitive deadline. A demand for independent evaluation also leaves open whether evaluators can reliably measure capabilities that developers themselves say they do not fully understand.

Critics have reason to inspect who benefits from the slowdown campaign. In objections reported by The Street and summarized by Slashdot, investor Michael Burry argued that fast competition benefits incumbents and that danger warnings function as hype ahead of initial public offerings. AI researcher Timnit Gebru, according to Wired in the same summary, has argued that extinction rhetoric can divert attention from present harms such as autonomous weapons. Those objections do not disprove future risks. They identify two design hazards: safety rhetoric can double as financial promotion, and spectacular scenarios can crowd out harms already facing workers and the public.

Washington and Beijing Hear an Industrial Contest

Trump has said the United States leads China by about a year and worries that slowing down would surrender that advantage. Vice President JD Vance described frontier laboratories requesting regulation as a possible “Trojan horse.” From Beijing’s side, language about Western safety standards can resemble an effort to lock in a Western lead, particularly when the debate also encompasses chip exports and open-weight models.

Treasury Secretary Scott Bessent has nevertheless left a narrow opening. In a statement to Axios quoted by Gizmodo, he said the United States was open to discussing “shared risks,” avoiding a split between the two countries’ AI systems, and both open- and closed-weight models.

That language falls far short of a joint safety regime. “Shared risks” could cover anything from cyber misuse to loss of control, while avoiding bifurcation could mean preserving commercial interoperability rather than slowing development. Still, it offers a rare administration framing of AI risk as an area for discussion with China rather than solely as ammunition in a race.

The diplomatic bind follows from the incentives. Washington fears that restraint could transfer advantage to China. Beijing can interpret American-backed restraint as containment. Companies then cite the absence of international coordination as the reason unilateral action is difficult. Each actor can describe continued acceleration as a response to everybody else, a machine for manufacturing inevitability out of individual choices.

Earlier Warnings Did Not Supply a Mechanism

Calls for coordination predate this week’s argument. Amodei called for global coordination after OpenAI released GPT-2 in 2019, and Elon Musk made a similar appeal in 2023. More than 1,000 AI workers signed a letter in July 2026 asking the US government to control AI research for safety and security, according to Tom’s Hardware.

That history changes how this latest burst of agreement should be read. Executive unity can elevate an issue, but recurring appeals do not supply thresholds, inspectors or enforcement. The current campaign advances the conversation by naming independent evaluators and stages of international coordination. It remains incomplete at the points where costs and authority begin.

Sen. Richard Blumenthal has suggested one domestic precedent: expert review before release, analogous to the Food and Drug Administration’s role in pharmaceuticals. He also invoked social media as the warning case, where lawmakers struggled to respond after algorithms and business models had become entrenched. His proposal supplies a principle, objective review by delegated experts, rather than a full institutional blueprint for AI. The FDA comparison cannot decide who evaluates a model used across borders or what evidence should block deployment.

Those unanswered questions offer a practical way to judge the Sept. 24 summit. A photo of Altman, Huang, Trump and Xi would show access. Evidence of policy would look different: a named US-China working group, reciprocal model testing, agreed capability thresholds, protected access for independent evaluators or even a timetable for negotiating them.

A state dinner can open a channel, but the AI slowdown campaign will remain ceremonial until somebody puts inspection rights and consequences on the menu.