Open-source AI models from China are having a moment. They are climbing industry benchmarks, attracting American developers, and causing growing alarm in Washington. Amid it all, US policymakers have begun considering measures that could effectively ban access to Chinese open models, echoing an earlier proposal by Senator Josh Hawley.
These ideas arise for different reasons and with varying degrees of seriousness. Taking such drastic action, however, would be imprudent and short-sighted.
There are legitimate debates about the security, provenance, and use of particular AI models developed in China. But proposals to categorically bar US access to open-source models developed there raise three broader issues, regardless of one’s perspective on any particular company or model.
Regulate conduct, not code
First, a ban could inadvertently undermine US competition. Open models are regularly used by American startups, scientists, and researchers. In a July 24 letter, Nvidia, Microsoft, Dell, and other large companies stated that open-weight models “strengthen competition and competition is what keeps the gains of AI broadly shared rather than concentrated in a few hands” and “[o]penness may be one of the most important paths to AI safety and security.” Small companies and startups made the same point. A July 22 letter signed by the prominent startup accelerator Y Combinator and 178 other founders and companies cited a survey of founders in its network, which found that nearly half of respondents used open-weight models to power their products. The coalition stressed that “[d]enying American startups access to models available abroad would stifle competition, entrench incumbents, and function as a tax on intelligence, raising their costs and narrowing their choices while foreign competitors retain access to the full global market.”
Open source is neither exotic nor new. The Cybersecurity and Infrastructure Security Agency calls it part of the foundation of critical infrastructure, as well as the backbone of the open Internet. The sensible answer is to secure that ecosystem, not sever parts of it.
Second is technology neutrality: the principle that federal law should apply even-handedly across different types of technology—including emerging technologies—and over time. If the Trump administration were to ban open-source AI models developed in China, how would that apply to an American company running such software on local servers? What about a Chinese model that is improved upon by a US startup, or a derivative later hosted in Europe?
Recent events underscore the point. According to preliminary accounts, an experimental model from OpenAI—a predominantly closed-source, US entity—caused a major cybersecurity incident when it “escaped” an evaluation environment and compromised the infrastructure at another company called Hugging Face, an open-source US company. Hugging Face then used a Chinese open-weight model on its own systems for forensic analysis after hosted models blocked the work. The episode shows that emerging risks do not map neatly onto the open-versus-closed or US-versus-Chinese divide.
Third, we have heard this sort of rhetoric before when it comes to technology bans. OpenAI’s head of strategic futures recently warned of “full AI communism.” In 2000, Microsoft’s then-chief executive officer said the open-source operating system Linux had “the characteristics of communism” because it was free. Washington has gone through similar cycles with strong encryption in the 1990s and Bitcoin around 2021: identify a misuse, characterize the technology itself as the threat, and reach prematurely for prohibition. But neither effort resulted in an outright ban.
The takeaway is not that emerging technology is harmless. It is that regulators and prosecutors should focus on conduct, not code.
The case for targeted enforcement
To be fair, Treasury Secretary Scott Bessent is right to point out that “open source is not open season on American IP,” and that there is a difference between “covert, industrial-scale distillation attacks” and genuine open-source development. None of this is to defend the Chinese government or any company that uses fraudulent accounts or proxy services to evade access controls and extract model outputs at scale. Those are legitimate concerns that warrant enforcement.
But “distillation” is not synonymous with theft. It is a somewhat elastic term that also encompasses training one model using outputs from another; the relevant concerns arise from how access was obtained and used.
Rather than prohibit a broad class of models, Washington has more targeted policy tools. It can enforce rules against credential fraud, prosecute trade-secret theft where applicable, and restrict prohibited entities’ access to advanced chips and large-scale computing. The federal government also has ample latitude to decline to buy particular AI products or prioritize US models without barring private companies from running publicly available weights.
The best response to capable Chinese open models is not to make US developers less capable. It is to build better, safer, more competitive US models—open and closed—and work with European and other democratic allies on shared evaluation and security standards. The European Commission’s new cybersecurity plan, which combines stronger model testing with the use of open-source models where appropriate, is a helpful guidepost.
The broader lesson here is that the urge to ban a technology can be attractive because it reduces complex policy problems to a simple act of prohibition. Bans may appear decisive and generate headlines, but often displace the painstaking work of policymaking on complex technical issues.
Effective AI policy will require something harder: enforcing existing US laws, negotiating new legislation, and setting clear security standards. Many constructive conversations lie ahead about how to regulate AI, both foreign and domestic. That work may feel less cathartic than a ban, but it is far more likely to deliver a sound national strategy.
JP Schnapper-Casteras is a nonresident senior fellow at the Atlantic Council’s GeoEconomics Center and a partner and general counsel of Collective Global, a venture capital firm with investments in the technology and AI sectors.
At the intersection of economics, finance, and foreign policy, the GeoEconomics Center is a translation hub with the goal of helping shape a better global economic future.
Further reading
Fri, Jul 24, 2026
Want AI you can trust? Start by building the right institutions.
Dispatches By Konstantinos Komaitis
Strengthening regional capacity on AI trust and safety could ultimately prove more valuable than establishing isolated national institutions.
Thu, Jul 23, 2026
Kenya wants to monetize its data. But at what cost?
Econographics By
As data becomes a strategic resource, Kenya wants to cash in. But without stronger safeguards on consent, privacy, and valuation, its data strategy could become a cautionary tale.
Fri, Mar 14, 2025
What is strategic about the new digital assets reserve?
Econographics By JP Schnapper-Casteras
To many on Wall Street and Main Street, this executive order on a strategic bitcoin reserve may still seem more like political maneuvering than sober monetary policy.
Image: Kimi K3, a new-generation AI model with 2.8 trillion parameters, is released by Moonshot AI at the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai, China. Source: Reuters.