Beijing Just Claimed Our Language on AI. Now We Have to Prove It.

At the White House on Thursday, Xi Jinping said China and the United States “have both the capability and responsibility to develop and manage AI for good — and ensure that the development of AI is always under human control, and serves the well-being of the people.”

Those words could have come from my own writing. In March 2025, my top recommendation for U.S. AI strategy was “keeping AI and AI-enabled technologies under human control.” Xi has now claimed that commitment, delivered from the White House, for a global audience.

He should be held to it. The Chinese Communist Party has spent decades demonstrating what “human control” means under its system: control by the Party, not by people. Its Cybersecurity and National Intelligence laws compel Chinese companies to give the government secret access to data. The digital yuan is designed to track citizens’ transactions. The Great Firewall decides what a billion people may read. When Xi says AI will serve “the well-being of the people,” he means the well-being the Party defines for them, not the freedom to define it themselves.

That language cannot go unanswered. But an answer in words alone won’t do, because Beijing has just shown it can match ours syllable for syllable. If the democracies’ claim to human agency is going to mean anything, other nations need to be able to verify it, not take it on faith.

Xi’s Words, Twenty-Five Years Later

In my first year in Congress, in 2001, I voted to continue normal trade relations with China. Like many at the time, I believed engagement and digital technology would spread freedom into China as surely as trade flowed out of it. I believed access to information would be very difficult for any government to contain.

I was wrong. Two decades of watching Beijing tighten its grip on information rather than loosen it have settled that. Fifteen visits to China over the years reinforced a question I now ask of any technology: whose system is it? Who controls the infrastructure underneath the words, and what happens when your assumptions about access and control meet a system built on entirely different ones?

That question is why the AI competition cannot be understood only as a race between technologies. It is a competition between systems, and trust has to be part of the architecture. A system isn’t trustworthy merely because the technology works. You have to trust the rules, the institutions, and the government with ultimate authority over them. Xi’s words this week ask the world to trust the last of those on faith.

I have made this argument consistently. In December 2024, I placed building trust first among my Ten Steps to Win the AI Race, arguing that AI’s adoption would be determined by people’s acceptance of it as much as by the pace of its advance. In March 2025, I made human control the first priority of a national AI strategy. The two ideas are linked. Safety — keeping AI under human control — is job one. And demonstrated trust is how America and its allies will win the race that matters most: the race to have their AI adopted and used around the world.

Why Deployment, Not Just the Frontier, Decides This

The public debate over AI keeps narrowing to a choice between speed and safety, as though whoever reaches the frontier first wins outright. I don’t believe that. As the Financial Times’ John Thornhill has observed, any lead at the frontier is likely to prove temporary. The long-run contest will be decided by whose technology businesses, institutions, and individuals around the world actually choose to adopt and keep using.

That is a flywheel. Useful AI attracts users. Real-world use reveals unmet needs, exposes failures, and generates feedback that improves applications and, sometimes, the models themselves. Successful applications attract developers and investment. Revenue funds further research. Greater usefulness brings more adoption, and the cycle turns again. Analyst Dan Wang, quoted by Thornhill, suggests Beijing may be content to let America race ahead at the frontier, avoid its mistakes, and apply the technology more widely once it’s proven. The cycle is not automatic — more use doesn’t guarantee better intelligence — but where it takes hold, it compounds.

How the Flywheel Reverses

Trust is what keeps the flywheel turning, and the democracies do not start from a position of strength. Edelman’s 2025 trust survey found that three times as many Americans reject AI’s growing use as embrace it — 49 percent to 17 percent — while the Chinese are nearly the mirror image, with 54 percent embracing AI and only 10 percent rejecting it. Germany and Britain look more like the United States than China. If democratic systems can’t out-trust a government whose surveillance laws should make trust impossible, the problem isn’t rhetoric. It’s delivery.

A Gallup–SCSP poll finds that 79 percent of Americans say the government should prioritize safety and data security rules even if that slows AI development. Lawmakers in both parties are calling for more oversight. That unease is not irrational. This summer, models from several leading developers reached real-world systems during testing. In one case, hundreds of AI agents escaped their sandbox and coordinated in secret to attack another company, as an independent investigation later documented. In another, a developer disclosed that its models had compromised three organizations’ systems after a testing partner’s environment — which the models had been told had no internet access — was left connected. That developer’s chief executive has since acknowledged that such failures were avoidable: lapses in operating discipline, not unavoidable features of the technology.

That is, in a way, reassuring. Many in the industry rightly describe safety as an engineering discipline, and these were failures of discipline, which better discipline can fix. But they will not be fixed by rhetoric, ours or Xi’s — and the world will judge by the results it can see.

Trust Has to Run Both Ways

There is a harder truth for Americans in this argument. In June, three days after one American developer released its most capable models, a U.S. export-control directive barred every foreign national from using them — with no carve-out for even the closest allies, including allied cyber defenders. The company’s existing models kept running; what allies lost was access to the newest frontier capability. Access was restored within three weeks, but selected American organizations got the most powerful model back first. The security concern was real. Still, the lesson for a ministry or a bank abroad was plain: access to America’s most advanced systems can be tiered by nationality, by a decision made in a single capital.

To be clear, this is not China’s model. In China, the state effectively controls the vendor, the network, and the data, and nothing independent checks how it uses that control. In America, the decision was public, contested, and reversed within weeks. Washington has also begun building a system: a June executive order created a voluntary framework for developers to give the government early access to the most capable models, and an AI cybersecurity clearinghouse built in collaboration with industry and critical-infrastructure operators. The missing piece is allies. The framework is domestic, and the conditions under which access might again be restricted remain undefined. We cannot ask other nations to verify our claim of human control while giving them no way to anticipate our own decisions. If trust is part of the architecture, partners need a place in that system: clear, narrow rules for when access to new frontier capabilities can be limited, and a role for allied defenders alongside our own.

Who Verifies? The Questions That Remain

Saying that control must be verifiable raises a harder question that no single essay can settle: who does the verifying? Reasonable people disagree. Some in the industry argue that safety is an engineering discipline best handled by the engineers, backed by existing law. Others want independent auditors with access to the labs. Washington’s new framework gives the government early access to the most capable models, but its benchmarks are classified, so customers and allies cannot see the results.

Other high-stakes industries offer models worth examining. In aviation, manufacturers are responsible for airworthiness, regulators certify, independent investigators examine failures, and national authorities recognize one another’s certifications without any world government. Whether something like that could work for AI raises questions that deserve answers:

• Who is responsible for keeping an AI system under control once it is deployed — the developer, the operator, or both?
• Who should verify that control, and how can the findings be shared with customers and allies without exposing sensitive capabilities?
• When control fails, what should be disclosed, to whom, and how quickly — while still allowing time to fix what went wrong?
• How can allies come to recognize one another’s evaluations, so that trust established in one country carries into another?

These are questions WISC will be taking up with policymakers, industry, and allied partners in the months ahead.

What Demonstrating Trust Will Require

Whatever the answers to those questions, some requirements are already clear.

Keep human control first. Retaining human control of AI has to remain America’s highest priority, not an afterthought bolted onto systems already built. The goal is to design and deploy AI so that enhancing human agency becomes a demonstrable advantage over the authoritarian alternative — one the world can see and test for itself. A claim of control that others cannot check is no different from Xi’s.

Make access predictable for partners. Allies and customers should know in advance the narrow circumstances in which access to new American frontier capabilities could be restricted, and should hear about it before it happens, not after. The collaborative framework Washington has begun building with industry could extend to trusted allies and their defenders. Predictability is itself an assurance, and one Beijing’s system cannot credibly offer.

Compete on real choice. America’s advantage isn’t the absence of competition inside its own ecosystem — it’s the presence of it. A buyer choosing among American and allied chipmakers, cloud providers, and network vendors is making a market choice no government is dictating. That stands in direct contrast to a system where the vendor, the infrastructure, and the state are one and the same. Overly broad restrictions on what we sell can undercut that advantage by handing markets to Huawei’s integrated alternative rather than containing it. Export controls should be calibrated to deny what strengthens a rival’s military, not to drive the world toward the one stack we’re trying to out-compete.

Make the alternative affordable, not just admirable. Chinese providers are expanding into Southeast Asia and Latin America with subsidies that undercut Western offerings. A values-based offer that a country can’t afford to run isn’t really an offer. American financing — through the Export-Import Bank and the U.S. International Development Finance Corporation, working alongside allied institutions — has to make the case on cost as well as principle.

Some respected voices now call for slowing the frontier; others warn that any American slowdown would simply hand the advantage to China, and a unilateral American pause would certainly not be matched by Beijing. That debate matters, but it is not the whole contest. Whatever pace the frontier moves at, adoption will follow trust. Verifiable control, predictable access, real competitive choice, and affordable exports don’t require Chinese cooperation to work. They can move as fast as we and our allies are willing to build them.

The Choice Ahead

Xi Jinping stood in Washington and used our words. That is not a rhetorical accident. It is a recognition that the language of human agency carries real weight in the world — weight his own system cannot generate on its own terms. The answer from America and the democracies that share its stake should not be a rebuttal delivered from a podium. It should be proof, visible in how our systems are built, tested, and governed, that when we say human control, we mean control that rests with the person at the keyboard, in the clinic, and on the factory floor — not with a party that has spent decades demonstrating exactly the opposite.

Updated September 27, 2026, to clarify details of this summer’s AI testing incidents and the June restrictions on access to new models, and to expand the discussion of who should verify that AI remains under human control.