Working Papers
Under Review
Public opinion about international affairs is shaped by different cues. As large language models (LLMs) emerge as a new source of informational cues, whose foreign-policy orientations do they reproduce? We examine two sources of variation: developers’ national environments and the elite–public divide. Across three studies, we evaluate fourteen LLMs developed in the United States, China, and France using Kertzer et al.’s (2014) foreign-policy attitudes battery, seventy-nine Pew Research Center questions, and forty-three Chicago Council questions. We find substantial cross-national convergence in broad dispositions and specific policy judgments, marked by cooperative internationalism, sustained engagement, and selective support for military force. National differences arise mainly in assessments of China and Russia. Moreover, when U.S. elites and the public disagree, most models align more closely with elites. These findings suggest that LLMs may reproduce U.S. elite-oriented foreign-policy perspectives despite their developers’ geopolitical differences, potentially reinforcing existing asymmetries in opinion formation.
- IPSAS Lecture Series, Academia Sinica, November 2026 (scheduled)
- ISA–APSA International Security Studies Conference, 2026
- Text as Data (TADA) 2026, UC Berkeley
- AI Upskilling for Social Scientists, Oxford Martin AI Governance Initiative, Oxford Martin School, 2026
- Joint workshop of the Australian Centre on China in the World and the Department of International Relations, Australian National University
Why do leaders in crisis sometimes meet their adversaries in a third country rather than in either side’s territory? This paper argues that venue choice is a strategic tool of crisis diplomacy. When bilateral visits are politically costly, third-country settings provide a spatial workaround: they allow leaders to communicate directly while reducing the symbolic costs of appearing conciliatory. We theorize that third-country diplomacy operates through political cover, attributional ambiguity, and responsibility diffusion. Using an original dataset of China’s senior leader diplomacy from 1998 to 2023, we examine China–counterpart encounters involving the Party General Secretary and the Premier and distinguish between bilateral-site and third-country venues. The results show that military disputes are associated with a greater likelihood of third-country encounters relative to bilateral-site encounters. Third-country encounters are also followed by higher levels of China-initiated cooperative behavior, particularly after periods of military friction. The findings contribute to research on crisis de-escalation, audience costs, and face-to-face diplomacy by showing that leaders do not only “talk their way out” of political constraints; they can also “move their way out” of them.
- ISA, 2024
Many Americans favor tariffs on China, yet a leader who raises them loses approval. Cost explains this gap but leaves open whether voters wanted the tariffs or the stance. Does the public reward the stance, the costly escalation, or the concession? I argue that the public’s appetite for confrontation is real but cheap. It is for the feeling of a trade war, not the fight. In two survey experiments, the public rewards the conservative message and a larger concession but penalizes retaliation. Even voters who say they want tariffs are no more eager once the bill arrives. A concession won without retaliation satisfies them as much as one won with it. The efficient (most supported) strategy is rude liberalization: combative talk with quiet tariff easing (no retaliation). This implies that economic coercion finds no domestic audience to make its threats credible. The public restrains a leader’s tariffs, not his words.
- APSA, 2022
- ISA, 2022
- 7th Annual Young Scholars Forum on U.S.–China Relations, The Carter Center, 2022 (earlier version; honourable mention)
Working Papers
- Political Economy of International Organization (PEIO), University of Texas at Austin, January 2027 (scheduled)
WTO members notified 47,042 draft product regulations between 2010 and 2026, against roughly fifteen dispute requests a year, so almost none will ever be litigated. Foreign governments comment on them anyway, and some regulators change their rules. I ask what a regulator concedes, and when. I argue that scrutiny is rationed twice, by how much the regulator discloses and by who holds trade on the covered lines. What it extracts is graded by what granting would admit. A regulator therefore grants time on the record and moves substance outside it. Where the regulated industry is organised, notified rules run longer and draw fewer comments. Who comments follows disclosure and the commenter’s own stake, which outweighs a third member’s on the identical lines. The written answer addresses seven requests in ten. The change record matches one. On conformity assessment the answer engages 67 of 79 requests and the change record none. A comment raises the chance a rule is later marked modified from four percent to nine. Coded blind, those changes relax the rule in a third of cases against a seventh. They arrive a median of four years later, and two of sixty-seven credit a comment. A transparency obligation can compel disclosure but not informativeness, since the record that makes the duty enforceable is the one no regulator will write a concession into.
- APSA, 2026
- ECPR Joint Sessions of Workshops, 2026
- Green Trade Lab Workshop, 2025
How do rising powers promote their norms? Rhetoric and socialization leave nothing to show. Material inducement costs more than a rising power will pay. Legitimation needs assent: public statements of support, not belief. We argue that a rising power can manufacture the record of assent itself. It builds an initiative whose label and borrowed diplomatic vocabulary commit no speaker, and solicits assent only there, never for the security content that takes sides. Endorsement is cheap, and the record shows it as acceptance of all the initiative carries. Neither side has reason to examine the difference. We test the argument on China’s Global Security Initiative, rebuilding its endorsement record from 8,601 news articles with an ensemble of large language models that establishes who produced each claim and in whose words. Chinese investment predicts how much a state endorses, domestic accountability who speaks: autocracies supply sitting leaders, wealthy democracies former officials. Much of the record is Beijing repeating endorsements or asserting support in its own voice, the latter only for aligned states. Endorsers take the label, not the content. A count kept by its beneficiary is evidence of promotion, not acceptance: manufactured assent works, but only to the depth of a label.
- ISA–APSA International Security Studies Conference, 2026
- Workshop on Security Networks and Contested Orders, UC San Diego, May 2026