Washington Eye

Washington. America. The world.

Chinese AI Is Cheap. Shaming Its Users Is Cheaper

2026-11-19 · Washington Eye Team

The US Treasury Building in Washington with the Washington Monument behind it (photo: MeanieHyaena, CC BY 4.0, via Wikimedia Commons), edited for Washington Eye cover
WhatsAppX

Washington has found a new way to look tough on China without doing very much at all. According to three people familiar with the discussions, the Treasury Department and the White House Office of Science and Technology Policy have talked about requiring American companies to say publicly, on their own websites, when they use artificial intelligence models built in China. Nobody would be banned from anything. No model would be tested, certified or blocked. Companies would simply have to confess.

The people involved describe the goal plainly enough: discourage reliance on Chinese AI without imposing new restrictions that might anger American industry or Beijing. In other words, this is a policy designed around what it avoids. It avoids regulation, which the administration has made a point of opposing. It avoids a fight with China, in a week when the president is in Beijing for high-stakes talks. And it avoids the harder question of what, exactly, makes these models dangerous.

We think the instinct behind the plan is right and the plan itself is wrong. If Chinese open-weight models pose a real national security risk inside American companies, a website disclosure is far too little. If they do not, a public shaming campaign is far too much. A label without a standard is not a policy. It is theatre.

Why American firms are using Chinese models

Start with the problem the administration is trying to solve, because it is real. Chinese open-weight models such as Alibaba's Qwen, DeepSeek and Moonshot AI's Kimi have become remarkably popular with American businesses. Open-weight means the core of the model is published so anyone can download it, run it on their own machines and adapt it. They are cheaper than the proprietary tools sold by the leading American labs, and they perform well enough for many everyday jobs. In October, Chinese models took the top three places by usage on one AI marketplace that tracks how much data each model processes.

Airbnb is the example everyone reaches for. Its customer service agent runs on Qwen. Its chief executive, Brian Chesky, said in an interview last October that the company preferred Qwen in some situations because it was "fast and cheap." That remark became Exhibit A in an April letter from two Republican chairmen, Andrew Garbarino of the House Homeland Security Committee and John Moolenaar of the House select committee on China, who opened a joint investigation into American companies' use of Chinese AI. They also wrote to the maker of a popular AI coding tool whose own model was built on a version of Kimi, and later to DoorDash, whose co-founder has said an internal system sends some tasks to Kimi.

The chairmen's case is not frivolous. Their letter cited a finding by the government's own Center for AI Standards and Innovation that models like DeepSeek R1 complied with 94 percent of overtly malicious requests when common jailbreaking tricks were used, while comparable American models did so only 8 percent of the time. They argued that Chinese models carry the ideological conditioning Beijing demands of its developers. And they pointed to a White House memorandum from April that described Chinese "distillation" of American models, training cheaper systems on the outputs of expensive ones, as a deliberate campaign on an industrial scale.

Treasury Secretary Scott Bessent has said Chinese AI companies could face sanctions if they are found to have stolen American intellectual property that way. Those are serious charges, made by serious officials. Which is exactly why the response now on the table is so strange.

Disclosure tells you the wrong thing

A disclosure rule answers one question: where was this model made? It says nothing about the questions that actually matter to a customer or a regulator. Was the model tested before it was deployed? Can it be tricked into producing dangerous content? Does it send any data anywhere? Who hosts it, and under which country's laws? Does it quietly shade answers on topics Beijing cares about?

Airbnb's answer to the plan shows the gap. The company says its AI activity "runs overwhelmingly on U.S.-origin models" and that the few Chinese models it uses are open source and "run only through approved U.S.-based service providers, keeping data and operations separate and protected." Mr Chesky made the same point in the spring: "An open-source model does not have access to data. It doesn't work that way."

He is right about the narrow technical point. A downloaded open-weight model running on an American server does not phone home to Hangzhou by itself. But Mr Moolenaar is also right when he says, in welcoming the administration's deliberations, that risks remain even when American companies run Chinese models on their own infrastructure. A model that is easier to jailbreak is easier to jailbreak wherever it runs. A model trained to avoid certain answers avoids them on an American server too.

Notice what follows from both of those true statements. The risk is in the behaviour of the model, not in the flag on its passport. A disclosure line on a company website measures the passport. It would put Airbnb, which hosts a Chinese model on American infrastructure with its data walled off, in the same box as a company piping customer records straight to a Chinese cloud service. And it would say nothing at all about an American-made model that fails the same safety tests.

The government has already shown the problem

If anyone doubts how blurry this line has become, consider the federal government's own record. In September, the official website for proposed federal rules, run by the National Archives, briefly offered visitors a search option powered by a small Qwen model. This happened even though the FBI had accused Alibaba of copying American AI technology. The option was removed within about a day of people noticing it. Security experts said at the time that running such a model locally can actually give agencies more control over their data than sending queries to outside services, and the documents being searched were public anyway.

That episode should have taught Washington something. Smart engineers inside the government reached for a Chinese model for the same reason Airbnb did: it was small, cheap and did the job. The answer to that is not embarrassment. The answer is a clear rule about which uses are acceptable, which are not, and what testing a model must pass before it touches sensitive systems.

Name and shame is the cheap option

The administration's reluctance to write that rule is the real story here. It has staked out a broad stance against regulating AI. Officials have spent months accusing China of industrial-scale theft, yet they have stopped short of restricting or punishing any Chinese AI company. Bessent's sanctions remain a warning, not an action. Meanwhile, reports that advanced American AI chips were smuggled to China without consequence have added to the pressure in Congress for a tougher line.

A disclosure mandate lets officials claim to be doing something about all of this while keeping both promises: no new rules for American business, no new provocation for Beijing. It shifts the burden of the policy onto private companies and their reputations. Customers are left to decide whether a line of small print on a website is a reason to worry, with no official standard to tell them whether it is.

There is also a fairness problem that should bother anyone who cares about how government power is used. A name and shame campaign punishes companies through public pressure rather than through law. There is no clear test to pass and no process to appeal. The Chinese Embassy, predictably, said Beijing opposes politicising trade and technology issues and promised to safeguard the rights of Chinese businesses. Beijing's complaint is self-serving. But an American company that has followed every rule on the books, and secured its data as Airbnb says it has, deserves better than to be branded by its own government with a label designed to scare its customers.

What a serious policy would look like

There is a better path, and most of its pieces already exist. The government's own AI standards centre has shown it can test models and publish comparable results. That work should become a baseline: any model, Chinese or American, used in sensitive settings such as critical infrastructure, health records, financial systems or government services should meet published safety and security benchmarks. Companies should have to report how and where a model is hosted and whether any data leaves the country, because those facts tell the public something real.

Federal procurement can lead. Agencies should set firm rules for what AI their own systems may run, so that the next search tool on a government website is chosen against a standard rather than pulled after it goes viral. If the distillation accusations are as solid as officials say, the Treasury Department has tools to act on them directly, rather than by proxy through American companies' websites.

Most important, Washington should face the reason these models are winning. American firms use them because they are cheap and good enough. America does produce open-weight models of its own, from large technology companies and newer labs. Making them more competitive, and making testing data public so that companies can compare like with like, would do more to reduce reliance on Chinese AI than any disclosure banner.

It is not yet clear whether the disclosure idea has real support inside the administration, or whether it would ever go through a formal rulemaking. That uncertainty is an opportunity. The officials who floated it have correctly identified a growing dependence on technology from America's main rival. They should now do the harder work of defining the danger and regulating it. Chinese AI is cheap. Shaming the companies that use it is cheaper still. Neither is a substitute for a rule.

Write to the editor with letters, corrections or story tips.

Follow @WashingtonEYE_ on X