2026-03-28 | An event analysis and governance breakdown for AI practitioners
On March 23, 2026, NeurIPS (Conference on Neural Information Processing Systems, one of the most important academic conferences in AI) included a sanctions compliance clause in its 2026 submission handbook, claiming that U.S. law required it to reject submissions from sanctioned institutions. Within four days, the China Computer Federation (CCF) called for a boycott, the China Association for Science and Technology (CAST) announced it would stop funding Chinese scholars’ attendance and exclude NeurIPS papers from its evaluation framework, and NeurIPS itself issued an apology and revised the clause.
The significance of this event extends beyond submission rules. It is a concrete case study of what happens when U.S. domestic legal jurisdiction, the governance decisions of a nonprofit foundation, and a global academic market heavily dependent on Chinese talent and paper supply collide. If you care about the infrastructure of AI research—who decides which research gets published, reviewed, and recognized—this case deserves careful examination.
The Main Track Handbook published on March 23 contained the following passage (original text, since revised):
The NeurIPS Foundation, like any entity operating within U.S. legal jurisdiction, is required by law to comply with U.S. sanctions and trade restrictions. Under these regulations, providing ‘services’ (which includes peer review, editing, and publishing) to individuals representing sanctioned institutions is prohibited. Consequently, we are unable to accept or publish submissions from these institutions.
The clause included a link to the Office of Foreign Assets Control (OFAC) Sanctions List Search, a general-purpose screening tool maintained by the U.S. Department of the Treasury.
As a 501(c)(3) nonprofit registered in California, the NeurIPS Foundation is indeed subject to U.S. law. The problem lay in two key areas of overreach, analyzed in detail below.
The speed and intensity of the response exceeded the typical pace of academic disputes.
On March 25, CCF issued an overnight statement calling on Chinese researchers to refuse to submit to NeurIPS 2026 and to decline roles as reviewers and Area Chairs. This response pattern closely mirrored CCF’s reaction in 2019 when IEEE restricted Huawei employees from reviewing papers.
On March 26, CAST announced two measures: suspension of attendance funding applications for NeurIPS 2026, and exclusion of NeurIPS accepted papers from recognition as representative achievements within the CAST framework. The second measure carries deeper consequences. In China’s academic evaluation system, CAST’s achievement recognition directly affects tenure reviews, grant applications, and institutional rankings. Excluding NeurIPS from this recognition framework effectively reduces the career return for Chinese researchers who submit to NeurIPS.
On March 26, NeurIPS posted a preliminary response on X with vague wording and no explicit apology. On March 27, it published a formal statement acknowledging that the handbook had linked to a screening tool whose coverage far exceeded its legal obligations, attributing this to a miscommunication between the foundation and its legal team.
What specifically changed? The original clause pointed to the OFAC Sanctions List Search, a general-purpose tool covering multiple lists, including the SDN List (Specially Designated Nationals—the most stringent sanctions list, where inclusion means asset freezes and a total ban on transactions with U.S. persons) and the Non-SDN CMIC List (Chinese Military-Industrial Complex list, which includes Huawei and SMIC). The critical distinction: the Non-SDN CMIC List’s legal effect is limited to prohibiting U.S. persons from trading publicly listed securities of these companies—entirely unrelated to academic services. The original clause conflated these two categories, effectively bringing entities like Huawei and SMIC within the scope of potential restrictions.
The revised clause explicitly narrowed the restriction to the SDN List and added a key statement: “NeurIPS will consider submissions from institutions and individuals categorized as Non-SDN.” The link was also changed from the general screening tool to the SDN list entry point. This represents a substantive narrowing of the compliance scope, explicitly excluding Non-SDN entities like Huawei and SMIC from restrictions.
However, NeurIPS did not abandon sanctions compliance itself. The revised handbook still retains the statement that “the NeurIPS Foundation operates under U.S. legal jurisdiction and must comply with U.S. sanctions and trade restrictions,” and still refuses submissions from entities on the SDN List. NeurIPS’s official narrative frames this as “correcting an error,” not “withdrawing a policy.”
A more accurate description is partial retreat: NeurIPS pulled back from an overly broad compliance posture that was legally indefensible to a narrower posture closer to the boundary of its statutory obligations. It acknowledged that the implementation was flawed, but did not acknowledge that the compliance direction itself was problematic.
This is also why the Chinese side did not accept the apology at face value. From the Chinese perspective, the issue is not merely the breadth of the implementation, but the fact that a conference claiming to be international embedded U.S. unilateral sanctions law into its global governance rules. The clause revision addressed the technical overreach but did not respond to the challenge on governance legitimacy.
The legal basis is the International Emergency Economic Powers Act (IEEPA, 50 U.S.C. § 1701 et seq.), enforced by OFAC. The core rule: U.S. persons (including U.S.-registered entities) may not engage in any transaction with entities or individuals on the SDN List, including providing services.
The SDN List is the most stringent sanctions list; inclusion means asset freezes and a comprehensive transaction ban. The Non-SDN CMIC List’s legal consequences are limited to a securities trading prohibition. The general-purpose screening tool linked in the original clause presented both categories together—this was the most critical technical error in the incident.
Even when limited to the SDN, a deeper question persists: does academic peer review, editing, and publishing constitute a prohibited “service” under OFAC? No court has directly ruled on this question.
The closest precedent is the 2024 settlement in GPE v. OFAC. In the settlement, OFAC determined that, under specific conditions, allowing sanctioned individuals to speak at public conferences does not constitute a prohibited “service,” but explicitly limited this finding to multiple preconditions (no financial transactions, no technical assistance, etc.) and noted it was a narrow, fact-specific determination. Peer review involves editorial coordination, reviewer assignment, and iterative feedback on revisions—whether this exceeds the scope of “speech” and constitutes a “service” remains a legal gray area.
IEEPA also includes an informational materials exemption that protects the import and export of information and informational materials from prohibition. However, this exemption covers the dissemination of completed materials, not the provision of services during the creation process. Whether peer review—fundamentally a service process—falls within this exemption is equally unsettled.
When IEEE restricted Huawei employees from reviewing papers in 2019 after Huawei was placed on the BIS Entity List, the legal regime involved was BIS export controls (EAR), not OFAC economic sanctions. The EAR contains an explicit “fundamental research exception,” and in 2019, the U.S. Department of Commerce resolved the restriction by clarifying that peer review does not constitute a controlled “service” under the EAR.
In 2026, OFAC has not issued a similar clarification. NeurIPS’s clause revision was the result of the foundation itself repositioning its compliance scope within a legal gray area, not a change in the legal constraints themselves. This means the 2026 revision is more fragile than the 2019 resolution: it depends on the NeurIPS Foundation’s own judgment about the scope of its legal obligations, and this judgment could shift again if OFAC’s enforcement posture changes.
The NeurIPS Foundation, as a U.S.-registered entity, is indeed subject to OFAC jurisdiction and indeed cannot provide services to SDN-listed entities—this is a hard legal constraint. But the two core problems in the original clause were not legally mandated: linking to a general screening tool that covered Non-SDN lists, and the vague definition of “affiliated with sanctioned institutions.” Both instances of overreach were conservative choices made by the foundation’s legal team in a gray area.
This also answers the core question many have asked: if this was required by U.S. law, how could NeurIPS change it? The answer is that the parts NeurIPS modified were never required by law in the first place. The hard legal boundary—no services to SDN-listed entities—was never withdrawn.
CCF’s boycott call and CAST’s countermeasures drove NeurIPS to issue a formal apology and revise its clause within four days. This speed depended on several specific leverage points China holds within the NeurIPS ecosystem.
Paper supply. Mainland Chinese institutions (Tsinghua, Peking University, Shanghai Jiao Tong, CAS, etc.) collectively account for roughly 12% to 15% of NeurIPS accepted papers. If the scope is broadened to Chinese-background researchers (those who completed undergraduate education in China and later work abroad), The Economist’s sample analysis from March 2026 estimated that approximately half of NeurIPS 2025 papers involved researchers with a Chinese academic background. These two metrics should not be conflated: the former captures the weight of mainland Chinese institutions; the latter captures the weight of Chinese-trained talent in global AI research. Together, they explain why China’s response created real pressure on NeurIPS.
Reviewer labor force. NeurIPS 2025 had over 20,000 reviewers and more than 1,600 Area Chairs. This review system is essentially a global volunteer labor infrastructure. NeurIPS does not officially publish reviewer nationality breakdowns, but publicly available proxies indicate that researchers from Chinese institutions or with Chinese backgrounds have a substantial presence in the reviewer pool. CCF’s call to refuse reviewer and AC roles directly threatened the conference’s peer review capacity. A reviewer shortage cannot be mitigated the way a submission decline can (by adjusting acceptance rates)—it directly degrades review quality.
Evaluation framework. CAST’s exclusion of NeurIPS papers from its achievement recognition changed the career incentives for Chinese researchers to submit to NeurIPS. In China’s academic evaluation system, which institutions recognize a paper as a “representative achievement” directly affects grant applications, tenure reviews, and institutional rankings. When NeurIPS papers no longer count, the same effort invested in a CAST-recognized conference yields higher career returns.
It is important to note that the effectiveness of these leverage points does not equate to “the Chinese government changed NeurIPS’s policy.” CCF is an academic society and CAST is a science and technology organization; their countermeasures carry real influence within China’s academic system but are not administrative orders. According to NeurIPS’s own statement, the direct cause of the revision was an internal miscommunication; China’s response provided external pressure and urgency, but the factors driving the revision also included criticism from non-Chinese researchers, legal scholars’ analyses, and the foundation’s assessment of reputational risk. The more compelling narrative is this: in an academic market where China contributes a large share of papers and reviewer labor, any rule change perceived as discriminatory will trigger a sufficiently large backlash.
Is NeurIPS an American conference or an international conference? The answer is both—and this coexistence is itself the source of the dispute.
At the legal level, the NeurIPS Foundation is a 501(c)(3) nonprofit registered in California, and OFAC’s sanctions rules have jurisdiction over it regardless of where attendees come from. At the academic level, NeurIPS’s submitters, reviewers, and accepted papers span dozens of countries, and its academic reputation rests on the collective participation of a global research community.
A critical asymmetry exists here: Chinese researchers’ contribution rights to NeurIPS (papers, reviews, attendance) far exceed their representation in governance. The 2025 NeurIPS Foundation Board of Trustees was drawn primarily from institutions in the United States and allied countries (Salk, Apple, Stanford, Google, Meta, Mila, KAIST, etc.), with no board members affiliated with mainland Chinese institutions on the public roster. The reality of NeurIPS is this: its legal shell and governance shell are highly Americanized, while its talent and paper market is highly globalized. This mismatch is unremarkable in normal times, but the moment rules intersect with geopolitics, it is immediately exposed.
This is also not unique to NeurIPS. Virtually all major AI conferences operated by U.S. nonprofits (ICML, ICLR, etc.) face similar jurisdictional constraints. The NeurIPS 2026 incident provoked a stronger reaction partly because the clause’s wording exceeded the scope of peer conferences’ existing policies, and partly because it occurred against the backdrop of accelerating U.S.-China technology decoupling.
Academic conferences are jurisdictionally bound infrastructure. NeurIPS, ICML, and ICLR function as the distribution and certification systems for global AI research. Whose papers are accepted, who is invited to review—these decisions shape researchers’ career trajectories and the perceived legitimacy of research directions. But the organizations operating these systems are subject to the laws of specific nations. When geopolitics enters the legal framework, the international character of academic infrastructure faces a stress test.
Legal constraints and governance choices are two different things. In this incident, the parts genuinely mandated by law and the parts representing the foundation’s own overcompliance were presented as one and the same. Distinguishing between the two is essential for evaluating similar events. When an organization says “the law requires us to do this,” it is worth asking: where is the legal boundary, and does the implementation exceed it?
A trust deficit has formed. As of publication, CAST’s countermeasures have not been withdrawn. The Chinese academic community remains skeptical about whether NeurIPS will again expand its compliance scope in the future. For practitioners who rely on international academic conferences as channels for research dissemination, whether China-led AI conferences and journals will attract more attention and submissions against this backdrop is a trend worth watching.
This analysis is based on the following primary materials:
The article distinguishes three types of claims: verifiable facts (sourced from the primary materials above), analytical judgments based on those facts (marked as analysis), and open questions without settled answers (marked as gray areas or unresolved).