How to Beat the Cheaters: A Free Market Framework for Dealing with Chinese AI

---

Issue Areas

---

Chinese AI has taken the U.S. by storm. No longer the sole purview of college students cutting corners on their homework, Chinese models now permeate prominent American companies, including Mozilla and Coinbase. Chinese firms claim they can approximate American performance, all while emphasizing their much lower costs. Yet the recent Chinese achievement in AI stems not from innovation but from a well-known proclivity of the CCP: cheating. Washington should not let models built on state-sponsored, industrial-scale fraud flood the American market. 

LLMs like Kimi, Alibaba’s Qwen, and DeepSeek’s V4 are trained using a process called distillation, in which a “student” model learns by prompting a more sophisticated “teacher” model rather than digesting the raw data, significantly cutting costs. That allows Chinese firms to make their models open weight or open source, meaning users can run them locally for pennies on the dollar. What is more, Chinese models benefit from extensive CCP support, be it through subsidies or Beijing’s National AI Investment Fund. The result: artificially low prices to lure consumers, even as their makers feign transparency. 

Determining the legality of distillation is the courts’ responsibility, and they are already weighing the matter. But litigation takes time, and the Chinese threat is imminent. Using another AI to “validate” your own is one thing; creating thousands of fake accounts to generate millions of training responses is another. Such “industrial scale” distillation, as Anthropic has called it, breaches terms of service and likely amounts to fraud in pursuit of trade secrets. And when the distillers are backed by a strategic rival, the stakes rise from civil liability to national security. 

Extolling Chinese “openness,” as the country’s president, Xi Jinping, did in a speech last month, is a bit like praising a bank robber because he shares some of his loot. Faced with the massive technological theft, there is growing pressure on Washington to adopt a sledgehammer approach: preemptively sanction Chinese developers and erect overt or hidden—but no less real—barriers to open models.

That would be a grave mistake. The White House is already implementing review processes for closed models, and industry giants like Anthropic have publicly argued for mandatory safety testing that encompasses open technology. Mandatory review could cripple the smaller developers who rely on open models and amount to regulatory capture, with a “closed” duopoly gatekeeping the field. The cost to innovation and consumers would be huge: as entrepreneurs from Silicon Valley giants to small business owners have pointed out, open models will be vital to bringing down costs and democratizing AI usage in the long run. 

Nor should Chinese competition frighten the U.S. if conducted fairly. In the early 1700s, the Whig lawyer Henry Martyn — a pioneering political economist and champion of Britain’s North American colonies — noted that importing cheaper products from overseas was often “the cause of the invention of arts and mills, and engines to save the labour of hands in other manufacturers” in his homeland of England. Chinese competition has forced American firms like Anthropic and OpenAI to change strategies and adjust prices; in the long run, it will push them to innovate in how they train models. 

Yet turning a blind eye to Chinese cheating would amount to the U.S. burying its head in the proverbial sand. If American firms are the ones turning garbled data into lucid models only for China to repackage them at lower prices, why make notoriously costly models in the first place? Were the White House to look the other way, it would be rightfully assailed by American AI developers just as Martyn was by his fellow Whigs, who mocked his blind faith in free overseas trade in a world where Spanish armadas and French corsairs regularly preyed on English ships. 

Washington should ditch the sledgehammer in favor of a scalpel. Given the breakneck cycle times of AI model development, the White House should issue an Executive Order (EO) that explicitly bans the “industrial-scale” distillation Chinese developers engage in, with the goal of eventually sending it to Congress for codification. The ban is all the more pressing because current intellectual property (IP) law requires human input to qualify for protection, potentially giving Chinese firms legal leeway even as they steal American technology.

The longer Chinese firms selling stolen technology operate without consequence, the deeper their models embed themselves in American software, making decoupling all the more painful. Moreover, Chinese embedding is not just a threat to fair U.S. competition, but to national security as well: there is already evidence that Chinese AI models deliberately produce lower- quality code when prompted by American users, with the goal of degrading U.S. software supply chain resilience. Who knows what other “sleeper cell” features might be lurking under the hood? 

Under the EO, American firms could seek redress against Chinese models engaged in industrial distillation, as evidenced by similar prompting patterns, shared payment sources, and coordinated timing across hordes of fake accounts. The typical remedy of civil fines would only extract a paltry sum from Chinese companies’ wallets. Instead, the Treasury should enforce OFAC sanctions: offending Chinese models would be banned from sale or use in the United States, and their users would be barred from contracting with the federal government or any entity receiving federal funds. 

Far from dulling the blade against China, the EO would ensure that sanctions are not imposed arbitrarily and that American companies receive justice under the rule of law. The AI threat from China is real. But meeting it will require surgical precision, lest the U.S. undermine the very innovation that led to its AI preeminence in the first place.