A well-meaning regulatory effort in Congress intended to ensure American AI dominance over competing Chinese models may have the opposite effect.
The United States has spent three years building an increasingly elaborate system of export controls meant to keep advanced artificial intelligence out of adversary hands. A new analysis by Clark Asay, a law professor at Brigham Young University, argues the policy is now producing the opposite of its intended effect: rather than isolating China, poorly calibrated restrictions are steering allied governments and legitimate commercial users toward Chinese AI models, ceding ground Washington will struggle to reclaim once it is lost.
The current regime dates back to October 2022, when the Commerce Department’s Bureau of Industry and Security imposed the first broad controls on high-performance chips and chipmaking equipment bound for China, later tightened through further rules in 2023 and 2024. The most ambitious iteration came in the Biden administration’s final week in office, when Commerce published the Framework for Artificial Intelligence Diffusion on 15 January 2025.
That rule sorted the world into three tiers: eighteen close allies, including the other Five Eyes members, most major NATO partners, and key semiconductor-ecosystem states such as Japan, South Korea and Taiwan, received near-unrestricted access to advanced chips. A large middle tier faced computing caps and case-by-case licensing requirements. A bottom tier of adversary states, including China and Russia, was effectively barred from imports.
The Trump administration rescinded the framework on 13 May 2025, two days before it was due to take effect, arguing it was too burdensome for industry and too damaging to relationships with allies, without immediately replacing it with a settled alternative.
The debate has not ended with that rescission. Pending legislation known as the Remote Access Security Act is designed to close a separate gap: the ability of adversary states to reach controlled American computing infrastructure remotely through cloud services rather than physically importing chips. The concern behind it is legitimate, since restrictions on physical hardware exports can otherwise be circumvented entirely through cloud access. Asay’s analysis argues the bill’s scope, as currently drafted, is broad enough to burden allied governments and ordinary commercial customers who present no meaningful diversion risk, alongside the adversaries it targets.
Why breadth becomes its own strategic cost
The argument is not that export controls are illegitimate. Keeping frontier AI out of hostile militaries is treated as an uncontroversial goal. The problem, according to the analysis, is that restrictions calibrated too broadly do not simply inconvenience the adversaries they target. They also push allied governments and legitimate commercial users toward alternatives, including increasingly capable Chinese models, when compliance costs or legal uncertainty make American platforms harder to rely on.
That matters more than a single transaction suggests because AI adoption carries strong switching costs. Companies and governments that build workflows, staff training and infrastructure around a given platform tend not to migrate away from it once committed.
The providers who capture users during this formative period of AI adoption are likely to retain them for years, and the reliability of access, not just raw model capability, is a significant factor in that decision. A foreign government or company that adopts an American model but faces the risk of sudden, unpredictable restrictions has a rational incentive to hedge toward a supplier whose access it can count on.
The gap that adversaries have already exploited
This dynamic has already played out in miniature. Despite years of expanding chip restrictions, Chinese AI developers, most visibly DeepSeek, have produced models assessed by industry observers as highly competitive, built using chips acquired before controls tightened, undercutting the assumption that hardware restrictions alone would meaningfully slow Chinese progress.
Separate technical assessments of the export control regime have described hardware-centric restrictions as leaky proxies: controlling physical chips does little to stop algorithmic optimisation, curated training data or software-level workarounds that allow capable models to be built on less-restricted infrastructure, meaning the compliance burden falls disproportionately on legitimate users while determined adversaries find alternative paths.
Part of the difficulty is structural. AI is inherently dual-use in a way that is harder to bound than most controlled technologies. A model trained to detect illegal fishing vessels can, with modest adaptation, identify naval vessels instead. Restricting the latter application can degrade the former, and no licensing regime has yet found a reliable way to separate the two at the point of export.
That ambiguity is part of why the diffusion framework’s tiered approach drew criticism even from close allies, some of which, including Latvia, Estonia and Israel, found themselves excluded from the most favourable tier despite their alliance status, a categorisation problem that fed resentment without clearly improving security outcomes.
The prescription is narrower and more predictable controls, targeted at demonstrated risks rather than broad categories of countries or use cases, paired with a willingness to treat American reliability, not just capability, as a strategic asset in its own right.
The counterargument, that some disruption to allies and commercial users is an acceptable price for denying transformative technology to adversaries, is treated seriously rather than dismissed, but the underlying assessment is that restrictions which are broad, unpredictable or poorly targeted risk accelerating the development of non-American alternatives faster than they slow adversary capability, shrinking the global footprint of American technology in the process.
The foundational choices about which AI ecosystem the world builds around are being made now, not in some hypothetical future in which enforcement can be perfected. Governments and companies making platform decisions this year are unlikely to revisit them soon after, given the switching costs involved. Whether Washington settles on a durable, narrowly targeted export control framework before that window closes, or continues cycling between broad restriction and abrupt rescission as it has for the past three years, will shape which country’s AI ecosystem becomes the default for the parts of the world Washington has not yet lost.
