For nearly four decades, NeurIPS has been the cathedral of machine learning — the place where a researcher from Peking University and one from MIT might co-author a best paper, eat the same conference lunch, and argue over the same whiteboard. On March 23, that era ended with a sentence.
The 2026 NeurIPS Main Track Handbook, published that day, stated for the first time in the conference's history that it must comply with U.S. sanctions and trade restrictions — specifically the Treasury Department's OFAC Specially Designated Nationals list — and cannot provide any services, including peer review or publication, to affiliated individuals. It was a legally defensible clause. It was also a match dropped into dry grass.
Within 72 hours, the China Computer Federation had urged every Chinese researcher to refuse submissions, reviewer roles, and area chair positions. Senior Tencent researchers publicly declined conference roles. A NeurIPS reviewer since 2020, Jiang Nan of the University of Illinois Urbana-Champaign, declined for the first time. By Thursday, the China Association for Science and Technology formalized the rupture: all funding applications for NeurIPS 2026 were halted immediately, and papers accepted at the conference would no longer count as representative works in any CAST-affiliated project evaluation.
The message was unambiguous. China's most powerful scientific institutions were withdrawing their labor and their legitimacy from the world's premier AI venue simultaneously.
The NeurIPS Foundation says it had no choice. U.S. law prohibits any American-registered entity from providing services — including peer review, editorial work, or publishing — to individuals or institutions on the SDN list. With over 870 Chinese-affiliated targets currently designated, among them Huawei, SenseTime, Megvii, and Hikvision, the legal exposure is real. The foundation is not wrong that compliance is mandatory.
But compliance does not excuse opacity. The conference dropped a blunt clause into its handbook without publishing a detailed implementation memo: which lists apply, how "affiliation" is determined, what happens to papers from mixed international collaborations, whether appeals exist. That ambiguity became policy by default — and in a high-stakes political environment, ambiguity is always read at maximum severity.
Crucially, there is a material factual distinction that has been lost in the outcry. OFAC records show that several of the most-discussed Chinese companies — including Huawei, SenseTime, Megvii, and Hikvision — appear as non-SDN entities under a separate executive-order program, not as full SDN designees. Whether NeurIPS is applying a broader internal interpretation of "affiliation with an SDN," or whether the public discourse has outrun the actual text, remains publicly unresolved. That is the difference between a narrow legal-compliance rule and a de facto nationality filter. NeurIPS owes its community a precise answer.
The Chinese response, meanwhile, is not merely defensive. China's scientific establishment has spent years documenting what it calls its "key links at both ends are controlled abroad" problem — a dependency where Chinese research is published abroad on foreign platforms and, in effect, imported back at foreign prices. The CCF's threat to strip NeurIPS from its prestigious recommended-conference directory, the list that Chinese researchers rely on for promotions and grant evaluations, is not just nationalist outrage. It is infrastructure-building. Beijing has been waiting for a forcing function to justify accelerating domestic journals, domestic conference brands, and domestic evaluation systems. NeurIPS has provided one.
The spillover risk extends well beyond a single conference. ICML and ICLR are both organized by U.S.-registered nonprofits. CVPR operates under IEEE, a U.S.-based institution. The same legal exposure exists before any parallel policy language has been published. The fracture, if it spreads, will not stop at NeurIPS's registration desk.
Both sides are underestimating the second-order damage. Washington may believe it is enforcing sanctions narrowly. But when the enforcement surface reaches peer review, it accelerates exactly the parallel scientific institutions it fears. Beijing may believe it is defending academic principle. But a blanket boycott of the world's highest-bandwidth research feedback loops also diminishes the science it is trying to protect.
What happened this week is not a conference dispute. It is the moment the U.S.-China AI conflict entered the paper-review pipeline — the coordination layer of science itself. Chips can be embargoed. Conferences can be boycotted. But a global scientific commons, once fractured, is not easily rebuilt.
Sources: NeurIPS 2026 Main Track Handbook (contains the sanctions compliance clause) — https://neurips.cc/Conferences/2026/MainTrackHandbook South China Morning Post — https://www.scmp.com/tech/article/3348006/ai-rift-widens-china-urges-boycott-top-us-conference-over-sanctions-ban