# The Co-Evolution of Conflict Aid and Terrorism Rhetoric Author: Charlie Carter Status: Working paper; undated ## Abstract Can foreign aid shape how states publicly frame terrorism? Foreign aid is a well-established instrument of statecraft, and growing evidence ties material assistance to changes in recipients’ public positions. Yet recent public criticism casts doubt on whether foreign aid advances donors’ counterterrorism interests, and we know little about whether conflict aid shapes how recipient governments frame terrorism in multilateral politics. I measure states’ terrorism rhetoric in the United Nations General Assembly from 1990 to 2012 by combining transformer-based embeddings of the local text around mentions of ‘terrorism’ and related terms with Bayesian dynamic latent-variable models, then estimate its co-evolution with conflict-aid networks using stochastic actor-oriented models. Joint tests support influence through aid ties across all three contested dimensions: recipients move toward donors’ positions on terrorism’s “root causes”, while larger inflows predict subsequent movement toward non-state-actor frames and more legalistic counterterrorism rhetoric, the latter largely irrespective of donor positions. The models support influence through established aid ties but not selection of partners based on rhetorical similarity. As major donors curtail conflict assistance, retrenchment may weaken one channel through which they shape multilateral counterterrorism debates. ## Links - [PDF (full text)](https://www.charlieacarter.com/papers/paper2.pdf) - [HTML landing page](https://www.charlieacarter.com/research/paper2.html) - [BibTeX](https://www.charlieacarter.com/research/paper2.bib) ## Version and source This source-derived Markdown corresponds to the published PDF with SHA-256 `a3e1c6ad6e6ee5b05cf7db1c84149de42b78deeed1254a52051736672d1cffd3`. Equations retain LaTeX notation; citation formatting and layout differ from the PDF. Use the PDF for page numbers. ## Introduction “Why pay terrorist organizations and certain countries to hate us when they’re perfectly willing to do it for free?” So asked Elon Musk on social media in February 2025, amplifying claims that the U.S. Agency for International Development had funnelled over \$160 million to “terrorist linked organizations” (Musk 2025). Days later, Chairwoman Marjorie Taylor Greene echoed the sentiment before the House Subcommittee on Delivering on Government Efficiency: “Has money through USAID been funneled to terrorists? Yes. …Taxpayer funds have literally been used to undermine U.S. interests and counter American foreign policy goals under the guise of foreign aid” (U.S. House of Representatives Committee on Oversight and Government Reform, Subcommittee on Delivering on Government Efficiency 2025). Within weeks, the Trump administration moved to dissolve USAID and halt the disbursement of nearly all foreign assistance. The premise underlying this dramatic policy shift invites empirical scrutiny. Focussing on the United Nations General Assembly’s (UNGA) annual General Debate, a site of marked and long-lasting contestation over multilateral counterterrorism policy, this paper examines whether conflict aid advances or hinders donors’ interests with respect to terrorism. Contrary to claims that aid begets no political influence, I find that established conflict-aid ties are associated with subsequent changes in recipients’ public rhetoric on terrorism. This relationship appears across the historically contentious dimensions of UN terrorism debates, although its form differs across issues. The common logic is that continuing aid relationships may connect both donor positions and aid exposure to subsequent changes in recipients’ expressed positions. If these relationships are present in contemporary international politics, a contraction of conflict aid could weaken one channel of influence over multilateral counterterrorism policies and debates. To understand why aid might matter for counterterrorism politics, it is necessary to first understand the nature of the contestation that has long characterised the UNGA’s treatment of terrorism. Previous studies have drawn attention to this many-decades-long contestation over terrorism’s causes, perpetrators, and remedies. Traditionally, explanations for its persistence have pointed to the essentially contested nature of the term “terrorism” itself. Terrorism, the argument goes, can be defined in many appropriate, but mutually exclusive, ways. The challenge posed to multilateral counterterrorism efforts, then, stems essentially from terrorism’s ambiguous meaning. On account of this ambiguity, “terrorism” can readily be used as a flexible, rhetorical cudgel in the rivalrous infighting of international politics (Saul 2006; Acharya 2009). Others are more sceptical of these purported semantic roots of international contestation over terrorism and view the contestation as being driven by a limited number of states with principally geopolitical motives. These states, disingenuous state-sponsors of terrorism and those that benefit from spoiling multilateral counterterrorism efforts, are held to sustain contestation at the expense of the broader UN membership (Ganor 2002). In this account, contestation over the meaning of terrorism and the appropriate international responses thereto is contrived: an entirely artificial barrier hindering progress on multilateral counterterrorism. In this paper, I contend that neither account is sufficient. The dimensions on which states’ public positions differ with respect to the issue of terrorism are well-defined and well-known. The differences between states on these dimensions are not contrived. Instead, they represent differences in expressed policy positions that are associated with the material, strategic, and attitudinal factors constituting states’ interests. One need not expect perfect correspondence between public and private beliefs for these differences to be politically meaningful. UNGA statements are repeated, attributable, and observable claims about what a government is prepared to defend before an international audience. They signal alignment and contest legitimate policy frames, even when they do not amount to international or domestic policy implementation. The distribution which these positions form may very well structure states’ expectations of their counterparts’ behaviour, even after accounting for any discounting of these signals on account of the possibility of misrepresentation. It’s through this channel that influence-bearing conflict aid may be consequential for multilateral counterterrorism policy. Specifically, I empirically assess the association between official development assistance (ODA) for peace, conflict prevention, and security (hereafter, “conflict aid”) and states’ terrorism rhetoric. I draw from relevant histories of contestation over the issue of terrorism in the United Nations to identify three key dimensions which have been the focus of sustained debates: whether to acknowledge and address terrorism’s “root causes” or to condemn and combat it without qualification (the ‘Causes’ dimension); whether counterterrorism should be a matter of more restrained law enforcement or of unbridled military force (the ‘Responses’ dimension); and whether to recognise and combat “state terrorism” alongside violence by non-state actors (the ‘Perpetrators’ dimension). I operationalise these dimensions using transformer-based embeddings of local terrorism contexts and Bayesian longitudinal latent-variable models to produce novel, substantively meaningful ideal-point measures of each state’s public rhetorical positioning, i.e., the language that they use in relation to the issue of terrorism in the United Nations General Assembly. I use a coevolutionary stochastic actor-oriented longitudinal network model (SAOM) to jointly model 1) changes in aid allocation, on the one hand, 2) changes in the rhetoric states use, on the other, and 3) the cross-process associations between them from 1990 to 2012. The results suggest that established conflict aid ties predict recipients’ subsequent rhetoric through a combination of donors’ positions and preferences, along with aid volume. The models yield significant joint Wald tests for the three-effect influence block on all three dimensions, though the relative importance of donors’ preferences varies across dimensions. Causes provides the clearest directional effect of donors’ positions: recipients are expected to shift in the direction of donors’ positions in proportion to their intensity. On Perpetrators, larger conflict aid inflows are the component most clearly associated with subsequent movement toward non-state actor frames, and more strongly so when donors already occupy those positions; at the highest volumes, that association still produces a significant push toward non-state framing even when donor positions are more state-oriented. On Responses, donors’ positions have essentially no independent association with subsequent rhetoric. Larger inflows are associated with movement toward legalistic and diplomatic frames irrespective of donor composition, a disciplining constraint on recipients even where donors themselves remain more repressive. Taken together, these are issue-specific expressions of a shared structure. Donor positions, aid volume, and their combination all recur, with different relative contributions depending on the issue at hand. The paper proceeds as follows. First, I identify and elaborate on the key rhetoric dimensions relevant to United Nations debates on terrorism, drawing on existing literature to map how these dimensions have historically shaped and reflected contestation among states. I then consider the motivations behind states’ decisions to provide conflict aid and the mechanisms through which aid might influence recipients’ policy positions and rhetoric. The paper next describes the data, textual measurement strategy, and co-evolutionary model, followed by the SAOM results and measurement-sensitivity checks. It concludes with the implications and limits of the dyadic findings. ## Literature Review ### Rhetoric Dimensions This section defines three recurring disputes in United Nations rhetoric on terrorism. Each is represented as a continuous dimension between two opposed positions, allowing states to occupy intermediate locations and to move over time. I establish the substantive meaning of the poles here and describe their textual measurement in the Methods section. #### Causes of Terrorism: Legitimating Root Causes vs. Outright Condemnation > “To achieve liberation and to secure their rights the peoples must confront reactionary violence-sometimes brutal, sometimes subtle, but always cruel. Thus to say the least it is unheard of that an item on terrorism should be suggested for inclusion in the agenda, as though that phenomenon did not have very clearly defined roots, as though its causes were not known to all, as though violence in the world today were a geometrical abstraction with neither explanation nor content.” > > *UNGA speech given by Cuba, 1972* > “Turkey has always condemned as criminal all acts, methods and practices of terrorism, regardless of their origins, causes or purposes. We believe that terrorism cannot be justified under any circumstances.” > > *UNGA speech given by Turkey, 2005* The first dimension of terrorism rhetoric that this paper identifies is one pertaining to terrorism’s causes. In the context of the United Nations, a recurring position advanced by a number of states is that terrorism is the natural consequence of some more egregious political circumstances, i.e., its “root causes”. These have included issues of colonialism, political domination, repression, racism, poverty, and disenfranchisement. It follows, in this line of argument, that terrorism is only a symptom of a deeper societal ill; counterterrorism efforts without concomitant attempts to bring about the end of these prior causes are akin to treating a symptom and not the sickness. This issue of terrorism’s root causes has preoccupied the United Nations’ attention since some of the earliest modern attempts at multilateral counterterrorism measures in the 1970s. Drawing upon the political and linguistic repertoires made available by the UN’s enshrinement of self-determination as a human right in resolution 1514 of December 1960, formerly colonised states and left-wing governments asserted that the factors which gave rise to national liberation movements, including alien domination and oppression, could give rise to terrorism (Mani 2004). Terrorism acted out in the course of struggles for self-determination, however, was accorded a legitimacy that terrorism emerging from different contexts was not because of the validity of the struggles which it served. Calls for such an exemption relied on the perceived moral and ethical character of these actors’ struggles against unjust forms of domination and the asymmetry in capabilities between foreign oppressors and the oppressed ((Jenkins and Johnson 1975, 2); (Saul 2006, 71); (Acharya 2009, 678)). Opponents of this argument adopted a more uncompromising stance, maintaining that terrorism is impermissible and illegitimate regardless of its motivating grievances (Saul 2006, 73). This view was most conspicuously advanced by former colonial powers and by those states which faced domestic insurgencies ((Dugard 1974, 74); (Friedrichs 2006, 73)). This pattern, while certainly normatively loaded, is also consistent with an interests-based account. In Friedrichs’s account, the struggle over definition turned on “power and ideology, namely the question of who defends and who challenges the political status quo”: the maintenance of national liberation violence “was in the real or perceived interest of many Third World regimes”, while “the Western status quo powers were not willing to accommodate” these claims (Friedrichs 2006, 78). Material exposure to political violence was likewise tied to support for condemnation: Britain embraced strong anti-terrorism measures chiefly because they would “criminalize external support for the IRA” while avoiding debate on the “underlying causes” of terrorism in Northern Ireland (Friedrichs 2006, 80). This issue of whether a distinction could be made between valid armed struggle and terrorism evolved, but remained consistently politically salient, over the following decades. At times a categorical distinction was made: terrorism and similar acts carried out in the course of justifiable armed struggle were fundamentally distinct and, out of principle, should not be conflated (Becker 2006, 90). Terrorism, then, is egregious violence against civilians without legitimating grievances. In contrast, the violence of liberation movements is legitimated by the nature of their fight and the need to use the means of resistance at their disposal. Alternatively, there were instances where it was taken as given that liberation movements did indeed use terrorism as a tactic. However, it was held that when these acts were committed in the course of armed resistance, while a regrettable choice of means, they stemmed from exceptional circumstances that had to be acknowledged in assessments of the violence’s legitimacy and illegality (Peterson 2004). In this view, terrorism may be an objectionable tactic that can nonetheless carry some degree of permissibility, depending on the context in which it is committed. Importantly, however, the literature suggests that support for the most severe interpretation of the root causes position declined over time, especially from the early 1990s, as a number of states warmed to the idea that “even persons or groups having morally justified ends need to employ morally justifiable means and exercise certain restraints in the conduct of their armed resistance” (Peterson 2004, 181). It was only in 1985, after the *Achille Lauro* seizure, that the General Assembly first condemned terrorism “unequivocally” as criminal “wherever and by whomever committed”, and resolutions continued to reaffirm the legitimacy of national liberation struggles until 1991 (Halberstam 2003, 575–76). The first resolution to drop that reaffirmation came in 1993, and the 1994 Declaration on Measures to Eliminate International Terrorism declared terrorist acts “in any circumstances unjustifiable, whatever the considerations of a political, philosophical, ideological, racial, ethnic, religious or any other nature that may be invoked to justify them” (Halberstam 2003, 576–77). Rhetoric of national liberation, self-determination, and calls to address terrorism’s root causes continued to be invoked in subsequent years with regularity, but the noted shift in attitudes represented a blow to the durability and effectuality of the coalition that had once emerged around this position. In an episode which hints at such a shift, the US formally rejected, in its 1993 response to the Secretary-General, the request in UNGA resolution 46/51 of December 1991 to convene an international conference to define terrorism and differentiate it from national liberation struggles, noting that such a conference would prove “non-productive” and “undercut the international community’s consensus” that specific acts proscribed by individual anti-terrorism conventions were “unacceptable whatever the rationale, context, or professed cause” (United States of America 1993). #### Responses to Terrorism: Legalistic vs. Repressive Counterterrorism > “Our efforts to fight international terrorism must be carried out with full respect for human rights and fundamental freedoms. It is imperative to strike the right balance between, on the one hand, preventive and swift action against terrorists and, on the other, adequate safeguards for the individual, and not least for those who have been unjustly listed.” > > *UNGA speech given by Denmark, 2005* > “All civilized nations are in this struggle together and all must fight the murderers. We are determined to destroy terror networks wherever they operate...” > > *UNGA speech given by the United States, 2004* The second dimension this paper identifies is one related to the responses to terrorism that states may adopt. At one extreme, a state might turn to repressive, kinetic, military-driven means of deterring or countering terrorists. At the other extreme, states might emphasise approaches which are traditionally the preserve of domestic law enforcement: prosecution, criminalisation, and intelligence-gathering. This decision ultimately hinges on whether counterterrorism is to be treated as criminal activity or as a military campaign and “lies at the very heart of the definition of modern terrorism” (Ganor 2018, 142). While this question may appear as an analytical or pragmatic one at first glance, it is fundamentally one of legitimacy and morality (Ganor 2018, 142), making it germane to the high politics of the United Nations and its fundamental focus on governing the use of force in international politics. Whether states frame terrorism as a criminal matter or a military one bears on the actors that may be involved, the rules to be followed, the legitimacy accorded to offensive and defensive operations, and the degree of force that is permissible against would-be terrorists (Ganor 2018; Boyle 2018; Crelinsten 2018). Each state’s particular security situation may lend itself to a more or less military-driven counterterrorism regime. And indeed, many modern counterterrorism organisations rely on a combination of police and military to capitalise on their unique capabilities and relative strengths (Boyle 2018). But a state’s decision to adopt either a military or law enforcement approach to counterterrorism affects, and is affected by, their relations with other states. Excessive use of force in counterterrorism operations may harm a state’s international status or its relations with its neighbours (Ganor 2018, 141) or engender backlash in their or their allies’ domestic publics (Abrahms 2007, 244–46). Formal models of proactive counterterrorism suggest that offensive measures can degrade terrorists’ operational capacity while also generating grievances and recruitment benefits that increase the risk of spectacular attacks (Rosendorff and Sandler 2004). Law enforcement approaches to terrorism may be more likely to facilitate international cooperation on counterterrorism efforts (Kruglanski et al. 2007, 109) by casting terrorism as an affront to public order and the rule of law. On the other hand, a law enforcement approach to counterterrorism may be seen by allies as too light-touch, especially when the tactics of terrorist groups enjoy wider support from the population of which they are a part (see Kruglanski et al. 2007, 110). These approaches are also more likely to be covert and less visible to domestic publics, meaning that they may fail to convince observers of the effectiveness of a government’s counterterrorism strategy even when the chosen strategies are more effective than alternative, more provocative strategies (Kydd and Walter 2006). In such cases, law enforcement style efforts to apprehend criminals may not have the *de facto* public support and legitimacy on which its efficacy traditionally relies. The Global War on Terror (GWoT) and the United States’ framing of its counterterrorism efforts as constituent parts of a broader “war on terrorism” might readily, and rightly, be identified as a political process where movement and changing positions were engendered on this Responses dimension. It has been documented that the “war” rhetoric of this period was accompanied by practical changes, with a number of states according a greater role for their militaries in their formal counterterrorism efforts at the expense of a more traditional policing model (Boyle 2018). However, even the “war on terrorism” rhetoric relied on calls for international collaboration, the application of relevant legal frameworks, and adherence to the rule of law to generate support. These elements are historically components of restrained, law enforcement styles of counterterrorism. This underscores the need to robustly measure the observable features of states’ rhetoric during, and in relation to, the GWoT, and to resist deductively ascribing to GWoT politics a universal push to repressive, militaristic forms or frames of counterterrorism. #### Perpetrators of Terrorism: State Terrorism vs. Nonstate Terrorism > “We do not and cannot condone the killing of innocent people, but how can we equate the desperate acts of a few individuals driven by ‘misery, frustration, grievance and despair’, to quote the Secretary-General’s words, with the brutal, deliberate and cold-blooded acts of indiscriminate killing perpetrated by the armed forces of a State?” > > *UNGA speech given by the United Arab Emirates, 1972* > “In the battle against violation of human rights, the authorities whose task it is to maintain public order have often exceeded the limits established by law. I am referring to a situation in which, in the context of its activity to contain the spreading tide of terrorism, legitimate authority exceeds its jurisdiction and expands its activity to a point where it can be termed a violation of the law. This gives rise to the idea that the concept of **State terrorism** must be added to that of terrorism pure and simple.” > > *UNGA speech given by Dominica, 1981* The third rhetoric dimension this paper identifies is one related to the type of actor that commits or facilitates acts of terrorism. While the traits of actors engaging in terrorism are varied, the dimension here specifically contrasts terrorism committed by *non-state* and *state* actors. Although many maintain that terrorism is necessarily committed by non-state actors, and that the expression “state repression” already identifies similarly illegitimate or excessive forms of violence committed by agents of the state against civilians, there has also been an opposing view expressed in the United Nations General Assembly that states are just as capable of committing terrorism as non-state actors. Generally the rhetorical features of this argument include the contentions that (1) states can commit terrorism, (2) the incidence of state violence matches or exceeds that of non-state terrorism, and (3) state terrorism gives rise to popular resistance and can produce non-state terrorism as a backlash effect (Saul 2006). “State terrorism”, as it is used in this context, refers to acts of excessive violence perpetrated by state agents against civilians or persons under their political authority. The salience of this rhetoric dimension in the early UN-led debates on international definitions of terrorism is attributable to a few factors. For one, recently decolonised states, and especially those that attained their decolonised status through violent means, represented a newly powerful political bloc in the United Nations at the time of the debates, the early 1970s (Becker 2006, 91). These states were opposed to the efforts of powerful Western states to criminalise and delegitimate acts of political violence that formed part of the repertoire of national liberation movements and the weaker parties in asymmetric conflicts. Moreover, these states argued that far more flagrant breaches of appropriate conduct in war and conflict were often perpetrated by state agents, especially when serving to suppress the activities of national liberation movements or oppressed peoples (Levitt 1986; Becker 2006). In addition, this rhetoric dimension mapped onto the re-intensification of the rivalries of the Cold War in the early 1980s, with the Soviet Union and the politically leftist governments of the Global South defining state terrorism “largely as use of force by the U.S., Israel, and South Africa against leftist revolutionary movements, Palestinians, and anti-apartheid activists, respectively” (Peterson 2004, 179). State opponents of the “state terrorism” concept have at times objected to it directly. For instance, an enduring objection is that the acts which have most frequently been associated with state terrorism are already proscribed under international law, humanitarian law, human rights law, and criminal law (Saul 2006, 207–8). Otherwise, these states have mostly signalled their objection to the state terrorism frame by neglecting to speak about terrorism in those terms, instead confining their discussion of terrorism to violence committed by non-state actors. Finally, the dimension has also expressed itself in relation to the issue of “state-sponsored terrorism”; transnational networks of cooperative ties between states and nonstate actors residing in rival countries, including those engaged in violence or contentious politics, are common features of contemporary statecraft that can exacerbate interstate tension and even drive escalation in rivalrous state-dyads (Maoz and San-Akca 2012; Kinne et al. 2026). It is true that labelling a state a sponsor of terrorism might be expected to “accomplish little” if the target state is already subject to penalties for other forms of deviant international conduct, or if their domestic politics and regime characteristics insulate them from such critiques (Byman 2023, 114). However, successfully drawing a connection between a state and nonstate actors engaging in illegitimate or contentious activities might nonetheless be stigmatising or invite material penalties, including sanctions or foreign aid suspensions (Byman 2023, 102), so invoking state culpability in UNGA discussions on terrorism can be seen as attempts at “naming or shaming” (Terman 2023; Brandt and Kotajoki 2026) or as attempts to steer multilateral policymaking toward the transnational relations underlying such conduct. ### Foreign Aid, Donor Interests, and Political Influence > “No matter how much we may wish to support the global war on terrorism, Palau, like many other developing nations, will not be able to participate meaningfully in this vital campaign without the assistance of the developed nations. Therefore, to those leading the combat against terrorism, Palau says: ‘Please, give us the tools we need to make a substantial contribution to this fight and render practical support for your efforts.’ ” > > *UNGA speech given by representative of Palau, 2002* Having traced the salient dimensions on which terrorism rhetoric has been contested in UN politics, I now review the theory and empirical findings on why states give foreign aid, how they decide to whom aid should be allocated, and what political effects aid provision has for donors and recipients. This discussion establishes the theoretical basis for expecting aid to affect states’ terrorism rhetoric and clarifies why these mechanisms should operate across the rhetoric dimensions described above. I begin with how donors allocate aid and what motivates these decisions. Although donors routinely frame aid provision in philanthropic terms, the social scientific literature has long emphasised its strategic character (Morgenthau 1962; Alesina and Dollar 1998; Findley 2018; Adhikari 2019). This does not mean that aid donors are strictly self-serving; aid provision is more likely when it meets both donor and recipient interests (Hoeffler and Outram 2011). It does mean, however, that donors often allocate resources with the expectation that doing so will advance their foreign policy goals. The features that constitute a donor’s “interests” are contextual, but the aid literature has identified several recurring strategic considerations that bear on allocation decisions. The strategic logic of foreign aid provision is long-standing. During the Cold War, both superpowers used aid to shore up recipients’ bloc allegiance, and the United States gave unconditional aid more readily to neutral and non-aligned states than to dependable allies, essentially buying alignment at the margins (Kanbur 2006; Lee 2022). Self-interested allocation did not end with bipolarity. If anything, the post-Cold War period was a “watershed” (Dunning 2004, 410). With this transition, donors could more credibly threaten to reduce or revoke aid when recipients failed to deliver desired policy changes because they no longer needed to retain recipients within a bloc at all costs. Geopolitical context thus conditions the credibility of donor demands, but aid’s utility for purchasing political concessions persists across it (Dunning 2004). #### Mechanisms of Political Influence: Aid-for-Policy and Social Influence I now turn to the mechanisms through which this influence may shape terrorism rhetoric, and the policy positions espoused through this rhetorical medium, in the United Nations General Assembly. I consider two mechanisms: explicit quid-pro-quo exchanges of policy concessions for aid, and social influence processes operating through the networks that aid ties create and maintain. Both mechanisms can matter: aid can alter the material costs and benefits of adopting a donor’s frame, and it can also change the beliefs, professional norms, and policy assumptions that make such a frame appear appropriate. These mechanisms need not operate exclusively and the evidence I present is consistent with their joint operation. I address aid-for-policy mechanisms first before turning to social influence. The aid-for-policy mechanism rests on a logic of reciprocity, offering a quid-pro-quo explanation of aid’s leverage over recipient politics (De Mesquita and Smith 2009). In this view, foreign aid is a resource granted in exchange for specific, negotiated policy concessions. These deals may take the form of explicit political conditionalities, as in cases where ex-ante conditions on aid are made public through national selectivity criteria or official documents (Molenaers et al. 2015). They may also be obscured from public view while being explicitly negotiated during aid bargaining, or communicated implicitly but clearly throughout the same process ( see (Yarhi-Milo 2013)). Applied to terrorism rhetoric, the mechanism implies that recipient states may adopt a donor’s preferred framing because doing so helps secure, maintain, or expand a valuable aid relationship. There is empirical support for the contention that states use such strategies to shape the counterterrorism policies and rhetoric of international counterparts. The European Union has repeatedly paired material ties with counterterrorism demands, for instance conducting trade and antiterrorism negotiations with Iran in parallel while serving as Iran’s largest trading partner (Gregory 2005, 112–13). The Bali bombing, and the 9/11 attacks on the United States that preceded it, similarly led to considerable material pressure and incentivisation from the United States, Australia, and others. The Indonesian police were “flooded with foreign aid” for counterterrorism purposes; funds from the US’s Anti-Terrorism Assistance Program led to the creation of a special counterterrorism police division, Detachment 88, while funds from Australia enabled the creation of the Jakarta Centre for Law Enforcement Cooperation (JCLEC) and the strengthening of police investigative capacity (Laksmana and Newell 2020). And yet, Jakarta continued to criticise the US “war on terror” frame, illustrating that aid can pull recipient rhetoric toward donors, but not necessarily override domestic constraints (Capie 2004; Febrica 2010). The pattern recurred among Indonesia’s neighbours, with the linkage sometimes only inferred and sometimes stated outright. Progress on free-trade agreements with the United States was tied to Singapore’s backing for the Iraq war and Thailand’s early support for the war on terror (Capie 2004; Chambers 2004). In the Philippines and Trinidad and Tobago, by contrast, US officials made the quid pro quo explicit, promising increased aid and warning of reduced investment, respectively, should antiterrorism legislation fail to pass (Whitaker 2007). Foreign aid might also induce a donor’s desired policy outcomes absent a quid-pro-quo exchange. Constructivist foreign aid scholarship argues that states can use aid to diffuse norms, ideas, or values that justify or naturalise the policy positions they would like recipients to adopt. This social influence mechanism operates through attitudinal change within target segments of the recipient state, including domestic publics, civil society, policy technocrats, diplomats, or policymakers. If these actors internalise supportive ideational frameworks, their influence or authority can translate micro-level attitudinal changes into shifts in recipient behaviour and public rhetoric. Such social influence strategies might target rank-and-file professionals, such as bureaucrats or active-duty military personnel, to affect aggregate recipient behaviour. Research on foreign military training finds that such programmes socialise trainees into donor-favoured norms. Graduates of US training programmes have been shown to endorse democratic-governance, human-rights, and liberal civil–military norms more strongly than personnel trained elsewhere (Fábián and Boutton 2024; Grewal 2022; Joyce 2022). Because the counsel of military elites carries weight with both leaders and publics (Jost et al. 2025; Jost and Kertzer 2024), such attitudinal shifts might alter a recipient government’s policy positions to the extent that those revised attitudes are reflected in elite counsel. Similar attitudinal revisions have been noted among non-military elites: US grants to Russian scientists following the collapse of the Soviet Union—established to contain proliferation risks by dissuading scientists from selling their expertise to authoritarian states or terrorist organisations—shifted recipients’ professed allegiances “away from a Soviet orientation to a new Western orientation”, an attitudinal effect that Western grants produced and Russian grants did not, and which was not reducible to employment provision alone (Ball and Gerber 2005). Similar programmes may target domestic publics, policymakers, diplomats, NGOs, and other civil society stakeholders in the hope of translating micro-level attitude change into macro-level policy shifts (Cronin 2004; Banks et al. 2015). Aldrich’s natural experiment study in northern Mali (2014) suggests that USAID programming against violent extremism affected the norms and civic behaviour of targeted populations in Mali, albeit without modifying their high-level, abstract beliefs about the nature of US action against terrorism. Complementing these grassroots efforts, donors also engage actors closer to government through policy and political dialogues. These mechanisms allow donor and recipient technocrats, diplomats, and policymakers to convene around policy priorities and reforms, providing a responsive means to steer the aid relationship. While these forums can be consensus-based, or aimed at spurring reform that is sensitive to both donor and recipient views, the political conditionality literature recognises that they can also be used to relay and enforce donor expectations without mandating hard, explicit triggers for aid disbursement or suspension (Molenaers et al. 2015). With greater relevance to the social influence mechanism, these forums can also realise donor priorities through regular channels of sustained technical cooperation and dialogue (Molenaers et al. 2015). Crucially, the two mechanisms can interact. A state may first adopt a donor’s framing for instrumental reasons and later internalise it. The constructivist “spiral model” holds that norms adopted to secure or regain aid can become genuine through subsequent argumentation (Risse and Sikkink 1999, 34), while positive inducements can durably realign a leader’s domestic coalition in ways that outlast the original bargain (Nincic 2006, 2010). Instrumental rhetorical concessions may thus induce genuine shifts in belief, though not necessarily in the direction a donor had envisioned. #### Aid and Theoretical Expectations Material inducements take many forms (foreign aid, preferential trade arrangements, sanctions), but this paper assesses a specific subset: official development assistance earmarked for peace, conflict prevention, and security (indexed by OECD and IATI purpose code $152$). Although traditional military assistance cannot be designated as official development assistance (see OECD 2023, for full enumeration of exemptions, exclusions, and edge cases), aid under purpose code 152 nonetheless includes activities that can empower recipients relative to domestic opponents and would-be agents of political violence. These include certain forms of material assistance from the aid provider’s military, activities supportive of democratic control of the military system, some training of military and police personnel, the provision of policing rather than military equipment, financing for civil policing functions, technical training for security sector reform, some training of law enforcement agencies in intelligence gathering, activities enhancing civil society’s role in security, prevention of violent extremism programming, programmes for the reintegration of ex-combatants and child soldiers, technical cooperation to counter the proliferation or availability of small arms and light weapons (SALW), and peacekeeping (OECD 2023). Three theoretical reasons justify the focus on this subset of aid. First, conflict aid establishes direct channels of professional communication between the security establishments of donor and recipient states. Unlike health or infrastructure aid, assistance for conflict prevention and security involves technical cooperation between military officers, intelligence officials, and police forces. These are the professional networks through which norms regarding appropriate counterterrorism conduct and conflict management are most likely to diffuse (Cronin 2004). Second, conflict aid also has an international-coordination rationale: unilateral counterterrorism measures can impose negative externalities on other states, and effective counterterrorism requires states to act in concert (Rosendorff and Sandler 2004). Donors can use conflict aid to build the relationships, capacities, and policy alignment such concerted action requires. Third, in the logic of quid-pro-quo bargaining, conflict aid offers especially credible leverage for issue linkage. A donor can plausibly condition security assistance on a recipient’s adoption of specific security policies or rhetoric, whereas conditioning food aid or medical assistance on such concessions would strain credibility. Deviant behaviour by the recipient on matters of conflict and security directly interferes with the policy outcomes for which the donor allocates conflict aid. Conflict aid can therefore provide especially credible leverage when donors seek rhetorical concessions on terrorism. The preceding arguments generate two dyadic expectations. First, conflict aid provides material leverage and creates channels through which donors’ expressed positions can influence relevant actors in recipient states. Once an aid tie exists, donors’ positions should therefore predict the direction of subsequent change in recipient rhetoric, with the relationship conditioned by aid exposure and volume. Second, if donors attend to rhetorical positioning when choosing partners, conflict-aid ties may be more likely where donor and recipient rhetoric is dissimilar rather than similar, i.e., disassortative mixing. Assortative mixing would concentrate aid among like-minded partners whereas disassortative mixing would concentrate it among partners whose rhetoric diverges, i.e., those who ostensibly stand in the way of the realisation of the donor’s preferred policy objectives. Thus, successful influence across disassortative ties might be expected to produce effects of relatively greater political consequence than would ties to actors that already espouse the donor’s preferred framing. These expectations are formalised below. **H1:** Through established conflict-aid ties, donors’ rhetorical positions will predict the direction of subsequent change in recipients’ terrorism rhetoric. **H2:** Conflict-aid ties are more likely to form and persist where donor and recipient terrorism rhetoric are more dissimilar rather than similar. Finally, while “informational disparities and a lack of leverage” allow recipients of US foreign counterterrorism aid to pursue their own goals under the guise of using aid for its intended purposes (Byman 2019, 628), a state’s public response to its donors’ terrorism rhetoric is observable in the United Nations General Assembly. If foreign aid is directed toward winning rhetorical concessions and instituting a preferred framing of terrorism, then donor states should be able to evaluate whether this goal is being met and, if not, terminate or reduce aid. The relationship between aid and rhetoric should therefore be observable in the longitudinal network data. ## Research Design This paper’s research design makes three contributions which address difficult inferential challenges presented by this paper’s research question. First, I employ a co-evolutionary stochastic actor-oriented model (SAOM) to jointly model changes in foreign aid networks *and* state rhetoric over time. This framework permits the estimation of interdependencies between network change and behavioural change, including the effect of prior aid ties on subsequent rhetoric and the effect of prior rhetoric on subsequent aid ties. By modelling these processes simultaneously, I can assess whether the joint process is consistent with aid operating as a channel of influence while accounting for the modeled selection process and observed determinants of aid-tie change. Second, I distinguish and model three aspects of recipient exposure: the direction implied by donors’ rhetorical positions, donor rhetoric weighted by each donor’s contribution, and total incoming conflict aid. A separate term in the aid-network equation uses the product of donor and recipient rhetoric scores to ask whether ties are assortative or disassortative given donor and recipient estimated positions. Respecting network structure and allowing for each donor’s individual position and aid amounts to exert independent effects is essential: each recipient can be connected to donors that disagree with one another and who contribute different amounts, and thus might carry different degrees of influence. Third, I combine transformer-based representations of local terrorism contexts with a Bayesian robust dynamic latent-variable model (Reuning et al. 2019) to estimate longitudinal positions in states’ terrorism rhetoric. Qwen3-Embedding-0.6B maps each local context window around a mention of “terrorism” or a related term into a dense semantic space. These embeddings place semantically similar rhetoric near one another in the vector space, although individual coordinates have no substantive interpretation. For each rhetorical dimension, I define a bipolar axis as the difference between the mean vectors of exemplars at its opposing poles, then score each local context by projecting its embedding onto that axis. This procedure extends the logic used to position words along such axes as rich–poor in static embedding spaces (Kozlowski et al. 2019). Because vocabulary associated with a pole may appear in contexts that reject as well as endorse it, I use Qwen3-Reranker-0.6B to rank candidate contexts by their correspondence to a written definition of each pole and construct the axis from the highest-ranked contexts. The resulting projection scores serve as noisy indicators of states’ latent positions. States contribute unequal numbers of local contexts and, in some years, none; moreover, any one context provides limited evidence about the position that generated it. Conventional dynamic models address such noise by shrinking estimates toward a smooth trajectory, but this smoothing can attenuate abrupt shifts of substantive interest. Reuning et al. (2019) develop a robust model for punctuated-equilibrium processes, in which traits remain stable for extended periods before changing sharply. Terrorism rhetoric plausibly follows this pattern. Adapting their model, originally designed for discrete ratings, to continuous text-derived scores constitutes the paper’s principal measurement contribution. I describe the construction in [Data](#subsec:data) and [Methods](#subsec:methods). I also validate the measures descriptively against outcomes that played no part in their construction. For nine counterterrorism instruments opened after 1990, accession is less likely both for states whose average rhetoric sits farther from the instrument’s position and in years when a state sits farther from that position than its own average; the association holds on all three dimensions. A state one standard deviation farther from an instrument has about 16% lower accession odds, and on Responses a one-unit move toward the instrument is associated with about 70% higher odds. Estimates are precise for the average-distance terms on Causes and Responses and for the movement term on Responses; the remaining three point the same direction but remain short of conventional precision, because deviations from a state’s long-run average are measured with far more error than the average itself. Separate state-year models confirm that contemporaneous terrorist attacks and fatalities are associated with Responses and Perpetrators but not Causes. Reported in the supplementary appendix, these checks show that the measures track political content correlated with a binding legal commitment and with lived exposure to terrorism; they establish neither a causal effect of rhetoric nor an identification strategy for aid. ### Data I create three cross-sectional longitudinal datasets of 195 states spanning the years 1990-2012, yielding 23 annual waves and 22 inter-wave periods. Each of these datasets has two dependent variables: conflict aid networks and measures of states’ rhetoric in the UNGA. These are at the network-year and state-year levels of analysis, respectively. The network-level dependent variable is identical across the three datasets: an adjacency matrix of directed conflict foreign aid ties, where “directed” indicates that each tie has a donor $i$ sending aid to a recipient $j$. A state’s absence from the network in any given year, which conceptually corresponds to missingness due to entry or exit from the international system but not non-receipt of aid, is recorded and handed to the SAOM estimation routine as network composition change. This network is plotted to visualise its structure at various time points in this paper’s Appendix. The state-level dependent variables are unique to each dataset and each correspond to one of the three rhetoric dimensions I have described. These roughly relate conceptually to the causes of, responses to, and perpetrators of terrorism. The UNGA General Debate is a particularly suitable forum for this analysis. It has been a noted site of sustained contestation over multilateral counterterrorism policy, providing substantive relevance to the research question (Peterson 2004). Although focussed plenary or committee meeting notes might afford richer treatment of terrorism by individual states, the General Debate offers greater temporal breadth and observation regularity, which are critical for identifying robust associations with patterns of aid allocation over time. The forum also offers near-universal coverage of recognised states, annual regularity, and wide discretion for representatives to determine the content of their statements on terrorism (see United Nations General Assembly 2003). These features contrast with more focussed meetings, which may imply a substantive or regional focus and reward more technical, less political contributions. The dependent aid network is constructed from the AidData Core Research Release (Version 3.1) record of bilateral official development assistance commitments (Tierney et al. 2011; AidData 2017), restricted to commitments under OECD/IATI purpose code 152 (peace, conflict prevention, and security). After aggregating purpose-code 152 records to donor–recipient–year cells, negative aggregate sector commitments are set to zero before any logging or tie construction. For each year I record a directed tie from donor $i$ to recipient $j$ when the cleaned raw annual commitment amount is at least US\$100,000 in constant 2009 US dollars, yielding the binary adjacency matrix used as the network dependent variable. This binary tie is built from the raw commitment amount before GDP normalisation. These are annual commitments rather than disbursements or flows actually received. This cutoff is an operational rule rooted in some existing precedent and relied upon to binarise the aid network for compatibility with the SAOM framework. Functionally, it removes a low-value tail of aid commitment ties while preserving nearly all substantive commitment volume (99.75%) and most observed actors (37 of 40 donors, 151 of 157 recipients). Carlitz and Ziaja (2024, 183) use the same cutoff to identify notable donor engagement within an aid sector, while Ziaja (2020) uses the same numerical convention in a dyadic measure. The nonnegative annual commitment amount is retained separately for the influence statistics introduced in the Methods section as a GDP-scaled dyadic covariate, ``` math w_{ijt}=\frac{A_{ijt}}{\mathrm{GDP}_{jt}/10^6}, ``` where $A_{ijt}$ is the annual commitment in constant 2009 US dollars and $\mathrm{GDP}_{jt}$ is recipient $j$’s total GDP, computed as World Bank GDP per capita in constant 2015 US dollars multiplied by population. Aid and GDP therefore use different reference years. Dividing by $\mathrm{GDP}_{jt}/10^6$ produces comparative aid exposure normalised by recipient economic size; it is not a literal percentage or share of GDP. The model applies the stated log and inverse-hyperbolic-sine transformations to this covariate; it is not the tie-defining variable. The supplementary appendix writes the same annual quantity as $w_{ij}(t)$ when placing it in the continuous-time model. The rhetoric measures are derived from the English-language General Debate transcripts in the UN General Debate Corpus (Jankin et al. 2017). The analysis window runs from 1990 to 2012, bounded at its upper end by the data quality and coverage of the AidData series. The wider exemplar corpus, including post-2012 material, defines the semantic axes; SAOM outcomes and inference are restricted to 1990–2012. While AidData Research Release 3.1 extends to 2013, activity coding coverage declines substantially from 48–67% in 1999–2012 to only 33% in 2013. This decline is methodologically consequential because real conflict aid ties that are not activity-coded would be treated as non-ties rather than missing data when constructing the network dyads. Although the SAOM can accommodate missing ties through network-based multiple imputation, excessive misclassification of actual conflict aid relationships risks the introduction of bias into the network structure. I therefore restrict my analysis to 2012, the terminal year with sufficient activity coding coverage. The conflict-aid networks are sparse and static prior to 1990, providing limited data for a co-evolutionary design. Years post-1990 introduce greater conflict aid activity and compose a period in which counterterrorism politics reshaped both aid priorities and UNGA rhetoric. Additional sources of data are collected to fully specify the model, including primarily by eliminating evident sources of potential confounding but also by improving model precision and including effects of substantive, but secondary, interest. These include whether a state was colonised (nodal, state-level covariate) (Hensel and Mitchell 2007), whether one state had colonised another state (directed, dyad-level covariate) (Hensel and Mitchell 2007), a state’s GDP per capita in constant 2015 USD (World Bank 2025), logged geographical distance in kilometers to other states (Schvitz et al. 2022), the Varieties of Democracy “polyarchy” measure of democracy (Coppedge et al. 2021; Pemstein et al. 2020), the logged asylum-seeker/refugee flow volume between states per annum (directed, dyadic measure) (United Nations High Commissioner for Refugees 2023), the logged number of asylum-seeker/refugees incoming/departing each state per annum (nodal, state-level covariate) (United Nations High Commissioner for Refugees 2023), the logged number of terror attacks recorded within a state over the preceding year (National Consortium for the Study of Terrorism and Responses to Terrorism 2021), whether a state is recorded as being party to an active state-based conflict in a given year (Gleditsch et al. 2002; Davies et al. 2024), and the number of battle deaths experienced by a state’s forces across all recorded conflicts in a given year (Gleditsch et al. 2002). ### Methods This paper jointly models changes in the foreign-aid network and state terrorism rhetoric from 1990 to 2012 with a co-evolutionary stochastic actor-oriented model (SAOM). The model treats annual observations as snapshots of an underlying continuous-time process in which states can change their outgoing aid ties and their rhetoric. Method-of-moments estimation chooses parameters so that model-implied network and behaviour statistics match their observed counterparts. Each parameter corresponds to a specified process of tie or rhetorical change and is interpreted through conventional statistical inference. Relative to alternative model families, the SAOM is particularly appropriate for this paper’s data and research question. Extensions to exponential random graph models (ERGMs) are powerful and commonly-used alternatives for longitudinal network analysis, but they are designed to study changes to network ties and generally treat actors’ behaviours, traits, attitudes, or commitments as exogenous. Where researchers *have* sought to endogenize and explain actors’ behaviour with these models, they have typically done so by adding artificial nodes to the network that serve as stand-ins for categorical behaviours or traits (for example, see Leifeld 2013). An edge linking actors to these behaviour nodes is then taken to represent an actor’s engagement in, or association with, the represented behaviour. This approach works well for measuring changes in endogenous actor-level measures that can be treated as coarse categorical variables or binary states. For example, textual topics in party manifestos or discrete policy positions have been modelled using this framework. However, it risks eliminating meaningful variation from variables whose substantive value lies in finer, continuous gradations. The base SAOM treats behaviour as distinct from the network, with its own dynamics and rates of change, but still handles its representation in traditional, categorical form. The continuous co-evolutionary extension used here (Niezink and Snijders 2017; Niezink et al. 2019) appropriately separates behaviour from the network and handles continuous measures by default. An alternative model family that is worth consideration, the temporal network autocorrelation model, does readily handle continuous actor-level dependent variables but, unlike the SAOM, cannot disentangle selection from influence (Duxbury 2023). The structural effects included in the aid model include indegree-popularity, square-root-weighted outdegree-activity, and the anti in-isolate effect. Indegree-popularity captures whether existing recipients of many aid ties are more attractive targets for new aid ties. Outdegree activity captures whether states which already send aid to many states are more likely to form new outgoing aid ties, with the square-root weighting yielding diminishing returns to additional ties. The anti in-isolate effect describes whether states are likely to send aid to a state which does not yet receive any; in the results table it is labelled *No-Donor Recipient Avoidance*. These effects are independent of those described by state- and dyad-level covariates and instead describe tendencies driven by properties of the wider network. To transform UN General Debate text into yearly measures of states’ positions on terrorism, I develop a text-as-data extension of Reuning et al.’s (2019) “robust dynamic latent variable” model. Reuning et al. devise this model to recover a smooth, persistent latent trait of actors from noisy and varied time-series cross-sectional data, while also accommodating rare abrupt changes after long periods of stability or modest change. Its original application is to discrete manifest variables, such as democracy levels and judicial decisions, that are characterised by such dynamics. However, its features are equally well-suited to recovering ideal-point estimates from continuous text-derived measures, where observations are irregular, noisy, intermittently missing, and occasionally exhibit sharp about-faces stemming from genuine policy changes. The Student-$t$ measurement and innovation priors simultaneously down-weight one-off noisy observations and permit rare, sharp shifts to register as such without the excessive smoothing exhibited by common alternatives. Implementing this extension required adapting the model’s measurement equation and estimation code to accept continuous inputs. Formally, the adapted model is a Bayesian state-space specification in which both the measurement and transition equations carry Student-$t$ priors in place of the Gaussian priors used in a standard linear-Gaussian state-space model. I first extract a twelve-token local context window around every mention of terrorism-related terms such as “terrorism” and “terrorist” in each speech. Qwen3-Embedding-0.6B maps each window to a dense transformer-based text embedding (Zhang et al. 2025). For each rhetoric dimension, I construct a candidate pool of within-corpus contexts to exemplify the extreme poles of the dimension. Candidate exemplars are filtered for relevance with Qwen3-Reranker-0.6B and ultimately selected according to a diversity constraint, which prevents either pole from being dominated by contexts with near-duplicate phrasing or from a single country, year, or session. The selected embeddings are then averaged into positive and negative pole centroids. Each local context receives a rhetoric score equal to the signed projection of its transformer-based embedding onto the normalized centroid-difference axis, computed as the cosine similarity between that embedding and the pole contrast and standardised within each dimension. This technique follows an emerging precedent in the text-analytic methodological literature and text analysis work in International Relations (Stoltz et al. 2025; Knapp 2026). Positive scores therefore correspond to root-cause/grievance language on Causes, legal or judicial response language on Responses, and non-state perpetrator language on Perpetrators; negative scores correspond to condemnatory framing, force or military response language, and state-repression/terrorism framing, respectively. The resulting projection scores are aggregated to state-years and treated as manifest indicators whose expectation equals the state’s latent position on the relevant rhetorical dimension, with Student-$t$ measurement error. The latent trajectories evolve as a random walk with Student-$t$ innovations that allow occasional large moves. Degrees of freedom of the Student-$t$ priors are fixed at four, following the recommendation of the original authors and Gelman et al. (2014, 319), for the $t$ distributions governing the size of both measurement and evolution shocks. This modification generalises the original application in Reuning et al. (2019) to the continuous embedding inputs used here, yielding estimates that are sensitive to genuine rhetorical reversals while down-weighting one-off noisy observations. This formulation also allows pre-1990 observations to inform the posterior trajectories even if they fall outside the primary 1990-2012 analytic window on which the SAOM is estimated. I use posterior medians from the DLVM as the state-year rhetoric ideal point measures in the primary SAOM. This is a defensible summary of each state’s latent rhetoric position, but it necessarily compresses the full Bayesian posterior distribution implied by the measurement model and smooths state-level behaviour trajectories. The estimation routine requires a realised behavioural trajectory composed of state-year point estimates rather than a full posterior distribution over dependent-variable values. I therefore treat the posterior-median specification as the primary model and assess measurement sensitivity with a 30-draw conditional coefficient-sensitivity exercise: the network component remains fixed while the behavioural coefficients vary. Because draw-specific rhetoric operands are not fully propagated through the network statistics, this exercise is conditional rather than a full propagation of measurement uncertainty. These results are reported below and visualised in the supplementary appendix. Table 1. Extreme local-context exemplars, by rhetoric dimension and pole | **Dimension** | **Pole** | **State** | **Year** | **Score** | **Context excerpt** | |:---|:---|:---|---:|---:|:---| | *Causes: high = grievance/root-cause framing; low = criminal/no-justification framing* | | | | | | | Causes | High | ROU | 2004 | 3.47 | the struggle to end *terrorism* must also address its underlying causes, including poverty … | | Causes | High | SYR | 2003 | 3.26 | success in addressing the root causes of *terrorism*, which are often found in poverty, ignorance, and injustice | | Causes | Low | SVK | 1998 | -3.38 | rejected *terrorism* and strongly condemns all acts, methods, and practices of *terrorism* as criminal and unjustifiable, wherever and by whomsoever committed … | | Causes | Low | SRB | 2013 | -3.11 | a comprehensive convention on international *terrorism*; Serbia strongly condemns *terrorism* in any shape or form, without regard to who the perpetrators … | | *Responses: high = legal/judicial framing; low = force/military framing* | | | | | | | Responses | High | CRI | 2016 | 4.71 | bring to justice those who commit such acts; the fight against *terrorism* must be framed in strict observance of international law, including human … | | Responses | High | MWI | 2009 | 4.16 | human rights, the rule of law, and the fight against *terrorism* in all its manifestations | | Responses | Low | LKA | 2007 | -3.28 | we launched military operations only to exert pressure on *terrorists*, in order to convince them that it would not be possible for … | | Responses | Low | LBY | 2016 | -3.11 | our limited resources, we are waging a relentless war to eradicate *terrorism*; all states have witnessed the courage of our armed forces | | *Perpetrators: high = non-state actor framing; low = state-actor/repression framing* | | | | | | | Perpetrators | High | AUS | 2017 | 3.36 | being challenged by non-state actors, *terrorists*, and violent extremists; the extreme islamist ideologies driving it must be confronted and defeated | | Perpetrators | High | NLD | 2005 | 3.19 | to take up the fight against radical ideologies that turn people into *terrorists*; as we all know, terrorism is a cross-border threat | | Perpetrators | Low | YEM | 1988 | -3.45 | condemnation of the acts of aggression of that regime and the *terrorism* and repression practised by it against the South African people … | | Perpetrators | Low | UGA | 1988 | -3.44 | under the inhuman practices of racism, repression, deprivation, and state *terrorism*; for the last three years, the South African regime has stepped … | *Notes:* Scores are standardised projection scores derived from Qwen3 embeddings under the rerank-diverse exemplar specification. Excerpts are drawn from the scored local context window; the target terrorism term is italicised where re-inserted. Ellipses indicate truncation at window boundaries. The behaviour component of the model utilises a stochastic differential equation (SDE) model to characterise the dynamics of state rhetoric in continuous time, conditional on network and state-level covariates. The SDE model includes baseline structural terms alongside the theoretically specified covariates and effects. An intercept term, feedback term, and the remaining beta coefficients define the expected equilibrium value of each state’s behaviour, *ceteris paribus*. The feedback term, specifically, models how the value of one’s behaviour influences its own derivative (Niezink et al. 2019). This feedback parameter is concerned with the value of one’s behaviour relative to its conditional equilibrium or expected value; the absolute value of behaviour is less important than whether one’s behaviour departs sharply (or not) from the value that would otherwise be expected. Positive feedback indicates tendencies of extreme behaviours to compound or veer further from the expected value, whereas negative feedback indicates tendencies of extreme behaviours to regress towards the mean or self-stabilise. Network and behaviour co-evolution is measured with four effects: three influence terms and one selection term. The three influence effects assess whether aid relations are consistent with an influence process for recipients: “*Summed Incoming-Donor Rhetoric*” (`totInAlt`) sums each donor’s centred rhetoric and tests whether donor positions push the recipient’s expected rhetorical derivative in different directions; “*Logged Incoming Conflict-Aid Exposure, GDP-normalised*” (`totWInAltLog`) captures the association between total incoming aid exposure and the recipient’s rhetorical derivative irrespective of donor positions; and “*Aid-Weighted Incoming-Donor Rhetoric (asinh)*” (`totInAltWAsinh`) assesses whether a donor’s contribution to rhetorical change varies with the centred aid it provides. The final term sums centred donor rhetoric weighted by centred donor-specific aid and then applies the inverse hyperbolic sine transformation, accommodating negative weighted sums while limiting the leverage of extreme values (see Burbidge et al. 1988; Aihounton and Henningsen 2021). The selection effect is “*Selection: Donor $\times$ Recipient Rhetoric*” (`egoXaltX`). It is the product of the donor’s and recipient’s centred rhetoric scores. Large positive products arise when those scores increase together, on either side of the mean; small or negative products arise when the scores have opposite signs or when one or both remain modest. A positive coefficient therefore favours assortative matching and a negative coefficient disassortative matching. The continuous-behaviour SDE uses actors’ behaviour scores in the structural feedback term and cannot accommodate a formal similarity definition based on proximity along the dimension; the product does that job, while missing similarity near the mean. Full change statistics and implementation details appear in the Supplementary Appendix. This specification respects heterogeneity in donor positions. For instance, in contrast to donors advocating for the unequivocal condemnation of terrorism, Sweden actively provided foreign aid and rhetorical support in the 1960s and 1970s to liberation movements with recorded histories of political violence and terrorism (Ekengren 2011). Permitting each donor to exert distinct influence according to its expressed position and the volume of aid it provides is therefore essential to avoid conflating donors with divergent expressed positions. Given that the continuous dependent variable SAOM framework can only accommodate a single endogenous behaviour dependent variable per model, a separate model is run for each of the three rhetoric dimensions this paper identifies. The aid network remains consistent across these models. As a consequence, the coefficients which describe properties of network evolution remain virtually constant across the three model specifications, notwithstanding some statistical noise and slight differences resulting from the inclusion of each ego $\times$ alter effect, which differs for each unique rhetoric dimension and captures differences in tie formation and maintenance tendencies given donor and recipient positions. On the other hand, the coefficients which describe processes of changing state rhetoric are expected to vary across specifications given that they each describe a unique dimension of states’ language, i.e., a distinct dependent variable. The rhetoric and network components of each specification are visually separated for ease of interpretation, but these coefficients are jointly estimated and belong to a single model per column of the results table with two simultaneous dependent variables, network and rhetoric behaviour. ## Results Table 2. Aid-Network and Rhetoric Models
Causes Responses Perpetrators
Estimate SE Estimate SE Estimate SE
A. Aid–Rhetoric Co-Evolution
Selection: Donor × Recipient Rhetoric -0.002 (0.002) -0.002 (0.002) 0.000 (0.002)
Influence block (three parameters, tested jointly below)
Summed Incoming-Donor Rhetoric 0.839** (0.270) 0.025 (0.227) 0.440 (0.314)
Aid-Weighted Incoming-Donor Rhetoric (asinh) -0.189 (0.260) -0.034 (0.167) -0.086 (0.112)
Logged Incoming Conflict-Aid Exposure, GDP-normalised 0.084 (0.271) 0.371 (0.272) 0.320* (0.161)
Joint Wald χ2(3) for Recipient-Side Block 15.241** 23.531* * * 12.176**
B. Network Component
Donor-Recipient Distance (log km) -0.035* * * (0.005)
Donor Colonised Recipient=1 1.121* * * (0.057)
Recipient→Donor Refugees (log) 0.060* * * (0.007)
Donor GDP per capita (const. US$, log, c) -0.996* * * (0.195)
Donor Democracy (c) -4.098** (1.415)
Donor Total Incoming Refugees (log, c) 0.030* (0.014)
Recipient GDP per capita (const. US$, log) -0.155* * * (0.014)
Recipient Democracy -0.017 (0.069)
Recipient In Conflict -0.135* (0.061)
Recipient Total Departing Refugees (log) 0.022** (0.007)
Recipient # Battle Deaths 0.036** (0.012)
Recipient Popularity 0.142* * * (0.004)
Donor Activity (sqrt) 0.158* * * (0.006)
No-Donor Recipient Avoidance -0.811* * * (0.048)
Reciprocal Aid Ties (fixed) -100 —
Baseline Tie Tendency (density) -2.313* * * (0.049)
C. Behaviour Component
GDP per capita (const. 2015 US$, log) -0.173 (0.139) -0.180* (0.077) -0.031 (0.045)
Level of Democracy -0.822 (0.695) -0.055 (0.395) -2.242* * * (0.232)
# Terror Attacks (log) 0.538* * * (0.144) -0.413* * * (0.081) 0.158* * * (0.044)
Was Colonised -2.232* * * (0.444) 1.717* * * (0.248) -0.299* (0.145)
In Conflict -1.256 (0.852) -0.750 (0.487) 0.326 (0.269)
# Battle Deaths -0.031 (0.190) 0.353** (0.111) -0.178** (0.058)
Current Number of Donors -0.328** (0.124) -0.090 (0.307) -0.685 (0.494)
Feedback -0.180* * * (0.052) -0.158* * * (0.031) -0.126* * * (0.016)
Intercept -0.655 (0.854) 1.058* (0.460) 0.308* (0.122)
Rhetoric Dimension Interpretation
Higher ↑ Root Causes Legal/Judicial Non-State Actors
Lower ↓ Condemnation Force/Military State Actors
D. Model Information
Convergence: all |t-ratios| < 0.1
Convergence: max ratio < 0.25
Final Estimation Iterations 300,000
Donor Fixed-Effect Reference USA
Num. Actors 195
Num. Years 23
Unit Controls (Network Component Only) Donor Fixed Effects
Time Controls Year Splines (4 df)
*Notes:* $^{*}p<0.05$, $^{**}p<0.01$, $^{***}p<0.001$; c denotes a centred covariate. Responses-model estimates shown for the network component; estimates are virtually unchanged across rhetoric models. Table 2 reports one co-evolutionary model per rhetoric dimension. I begin with the network component that governs aid-tie formation, then turn to the co-evolution estimates where the joint influence-block Wald tests provide the primary inference. I next plot the expected joint contribution of the aid influence terms across the full distribution of aid volumes and aid-weighted donor rhetoric positions, holding raw incoming degree at its sample mean. This visualisation is followed by a robustness check that probes sensitivity to rhetoric measurement by re-estimating the behaviour component over posterior draws from the dynamic latent variable model and assessing how closely the joint effect of their aid-influence coefficients reproduce the primary model. Conditioning on donor fixed effects and structurally prohibiting reciprocal aid back to donors, conflict-aid ties are more likely to be formed or maintained by donors that already send aid to many recipients ($0.158^{***}$) and to recipients that already attract many donors ($0.142^{***}$). Donors are less likely to target recipients with no current donors ($-0.811^{***}$), and the baseline cost of adding ties remains strongly negative (baseline tie tendency, $-2.313^{***}$). Geographic proximity increases the probability of aid ($-0.035^{***}$ on log distance), as do historical colonial relationships between donor and recipient ($1.121^{***}$). Poorer recipients are more likely to receive or retain aid ($-0.155^{***}$ on recipient GDP per capita), and conflict severity again has mixed implications: ongoing conflict slightly reduces tie probability ($-0.135^{*}$), while greater battle deaths make aid more likely ($0.036^{**}$). The network component is substantively stable across the three rhetoric outcomes; Table 2 displays the Responses-model estimates as a representative set. Turning to co-evolution, the results reported in Panel A require a two-step reading. The three influence block parameters represent distinct components of each recipient’s exposure to aid and donor preferences: the overall sum of their donors’ centred rhetoric scores, total logged incoming aid, and the sum of each donor’s rhetoric $x$ its aid amount. Because these parameters are estimated together, the joint Wald test is the primary omnibus test and should be read first; the individual rows then provide evidence about which component is most clearly identified. Correlation among the three statistics can inflate marginal standard errors and make individual coefficients appear imprecise even when the block as a whole is informative (Ripley et al. 2026). The joint Wald $\chi^{2}(3)$ statistics reject the joint null for every rhetoric dimension: $\chi^{2}=15.241$, $p<0.01$ for Causes; $\chi^{2}=23.531$, $p<0.001$ for Responses; and $\chi^{2}=12.176$, $p<0.01$ for Perpetrators. Thus, the three-parameter recipient-side blocks jointly reject the null that all three coefficients equal zero in each dimension. The selection hypothesis (H2) is not supported. The centred donor$×$recipient rhetoric product is small and statistically indistinguishable from zero in all three models. Large positive products arise when donor and recipient scores increase together; small or negative products arise when those scores have opposite signs or remain modest. A null therefore reflects no evidence for either assortative or disassortative selection. Donors do not appear to systematically select, maintain, or break ties according to the similarity or dissimilarity of their rhetoric with recipients. In the influence block, the Causes model shows a strong positive main effect of donor rhetoric ($0.839^{**}$). Donor sets above the overall rhetoric mean contribute positively to the recipient’s fitted rhetorical derivative, while donor sets below it contribute negatively, holding the other model terms fixed. This is a directional donor-position effect rather than a formal similarity statistic. The Responses donor-position coefficient is essentially zero. No individual influence-block coefficient is significant, though the joint Wald statistic is the largest of the three models. Larger inflows are associated with movement toward legalistic and diplomatic frames, with little role for donor composition. On Perpetrators, total incoming aid is positive and significant ($0.320^{*}$). Increases in incoming conflict-aid exposure are associated with movement toward non-state actor rhetoric. That association is stronger when donors already occupy those positions. At the highest volumes, the fitted association still points toward non-state framing even when donor positions are more state-oriented. Across the remaining rhetoric terms, rhetoric exhibits self-stabilising tendencies on all three dimensions (negative feedback: $-0.180^{***}$, $-0.158^{***}$, and $-0.126^{***}$ for Causes, Responses, and Perpetrators respectively), indicating regression toward each state’s conditional equilibrium. A positive feedback parameter would predict escalatory or self-reinforcing behaviour; the uniformly negative feedback estimated here is consistent either with regression to the mean on noisy re-measurement or with substantive penalties in the international system for rhetoric that departs sharply from a state’s conditional equilibrium. The behaviour equation also adjusts for observable recipient characteristics. The logged count of terrorist attacks are associated with higher Causes, lower Responses, and higher Perpetrators positions. Battle deaths are associated with higher Responses and lower Perpetrators positions, while GDP is negatively associated with Responses and democracy with Perpetrators. These coefficients are controls within the joint model rather than tests of the paper’s aid hypotheses. ![Figure 1. Conditional Aid-Related Rhetorical Drift](https://www.charlieacarter.com/research/figures/paper2/aid-rhetoric-surface.png) Fitted aid-related rhetorical drift across observed combinations of incoming aid and aid-weighted donor rhetoric. The vertical axis ranks recipients by the aid-share-weighted mean of their donors’ rhetoric so that similar donor mixes appear together. That ranking is for reading the figure; the influence terms still enter as totals over each recipient’s donors. The plotted quantity sums the three recipient-side aid terms and the separate incoming-degree control, with raw incoming degree held at its surface-sample mean. It is a partial drift contribution rather than the model’s full expected annual change. Warmer colours indicate positive drift, cooler colours negative drift, and retained black lines mark zero boundaries of sign components occupying at least four 31 x 31 cells. The influence block’s terms are not independent, as changes in donor aid provided and donor positions affect multiple terms simultaneously. Accordingly, their coefficients should not be interpreted as isolated effects, but as components of a conditional joint contribution whose substantive implications depend on the values of the other terms (Brambor et al. 2006). Figure 1 visualises the influence terms’ contribution to drift, the instantaneous expected rate of change in rhetoric, across observed percentiles of aid volumes and donor positions, with incoming degree held at its mean to retain the structural contribution of incoming ties. Variation across each panel comes from donor composition, total aid exposure, and how aid is distributed among donors; each cell summarises nearby observed actor-period configurations rather than an unsupported combination of inputs. The Causes panel varies mainly with donor positions. Recipients move toward the locations their donors occupy. The Perpetrators panel rises with incoming aid and with more non-state-oriented donor sets. At the largest inflows, it remains toward non-state frames even among more state-oriented donor sets. The Responses panel varies mainly with incoming aid, toward legalistic and diplomatic frames, and shows little independent contribution from donor positions. The rhetoric trajectories behind the results are themselves estimated with uncertainty: the Bayesian latent variable model returns a full posterior distribution over trajectories, of which the primary model uses only each state’s posterior median. The appendix tests whether this single summary drives the findings. It draws 30 complete trajectories from the posterior distribution and re-estimates the behaviour equation on each. Every draw’s fitted aid coefficients are then evaluated on the same input grid as Figure 1, with rhetoric inputs and percentile ranks held fixed, so that only the coefficients change. This procedure injects the model’s own measurement uncertainty into the fitted surface. The surface is preserved strongly for Perpetrators, whose posterior-draw median surface reproduces the primary surface closely ($\rho = 0.996$), and substantially for Causes ($\rho = 0.843$). Responses is the exception: its posterior-draw median surface aligns only weakly with the primary ($\rho = 0.379$). I therefore identify strong coefficient robustness for Perpetrators and substantial evidence for Causes. Responses, in contrast, is more sensitive to measurement uncertainty. In summary, established conflict-aid relationships predict subsequent terrorism rhetoric through an influence block that is jointly informative across all three dimensions. The same three recipient-side terms recur, with different weights. H1 receives direct support on Causes, where donor positions are identified. On Perpetrators, incoming aid is identified and donor positions modify that volume association. On Responses, the primary fitted association is with incoming aid rather than donor positions, and the detailed surface is the one most sensitive to rhetoric-measurement draws. H2 is not supported. This null is an absence of evidence for assortative or disassortative selection under the centred-product statistic. Similarity near the mean is not separately identified, and the null does not rule out all strategic allocation. These estimates concern changes within established donor–recipient relationships. They do not identify aggregate effects outside those relationships. ## Discussion A growing body of research treats the rhetoric states use in international institutions in relation to conflict and political violence as substantively meaningful public political positioning (Knapp 2026; Brandt and Kotajoki 2026). This paper’s evidence that conflict-aid ties are associated with shifts in recipients’ terrorism rhetoric accords with and extends this literature on both substantive and methodological grounds. Across the three dimensions considered here, the joint influence blocks are associated with subsequent recipient public rhetoric through established aid ties. The donor–recipient rhetoric-product terms provide little evidence of assortative or disassortative selection. Donor positions, incoming aid, and their combination appear throughout, with different weights by issue. Perpetrators shows why those weights have to be read jointly. High inflows can still move recipients toward non-state frames when donors themselves are more state-oriented. Those differences are substantively informative. They do not imply that each dimension follows a separate process. The qualification is clearest for Responses. It returns the largest joint test but no individually significant coefficient. In the primary fit, donor positions contribute little and larger inflows are associated with more legalistic and diplomatic public rhetoric even where donors remain more repressive. That pattern is consistent with a disciplining or conditionality reading attached to larger aid relationships. It is not well described as alignment with donors’ own legalistic positions. Professionalisation of recipient security institutions remains observationally possible. The detailed surface is also the most sensitive to rhetoric-measurement draws. This substantive contribution builds on a literature documenting aid’s role in multilateral alignment. Previous work has shown that the United States provides aid to states in order to secure their votes in UNGA if it can do so cheaply enough to constitute a near-minimal winning coalition (Alexander and Rooney 2019). This paper extends that literature with evidence consistent with conflict-aid relationships and donor composition predicting subsequent public rhetorical positioning beyond discrete vote buying. The composition of a recipient’s aid relationships matters: donor positions, donor-specific contributions, and the relative leverage each donor possesses over a recipient relative to other donors all shape the relevant association. The evidence cannot distinguish aid-for-policy bargaining from social influence; both mechanisms remain observationally consistent with the observed associations. These results also speak to the debate, set out at the outset of the paper, over why terrorism remains contested in the UNGA. If contestation were primarily an artefact of the term’s semantic ambiguity, there would be little reason to expect material transfers to influence how states frame terrorism. The association between conflict aid and systematic movement in recipients’ rhetoric instead suggests that public positions are responsive, at least in part, to the material and strategic incentives created by aid relationships. The evidence is therefore more consistent with an account in which contestation reflects interest-based differences in expressed positions than with one in which it is a purely definitional impasse. Methodologically, the paper pairs transformer-based embeddings of local terrorism contexts with an adapted version of the model presented by Reuning et al. (2019). This approach addresses challenges of missing data, high observation variability and measurement error, and dependence across time that frequently arise in studies of state rhetoric. The resulting estimates represent ideal-point-style measures of states’ public rhetorical positions over time. They remain sensitive to rapid changes in rhetorical positioning, incorporate uncertainty while remaining robust to textual noise and semantic variation, and carry direct substantive interpretations. Modelling these textual position scores as endogenous outcomes within a joint network–behaviour framework builds directly on other International Relations applications of this model family (see Warren 2016; Chyzh 2016; Kinne et al. 2026), extending them to the continuous behavioural variable case. Formal crisis-bargaining models often treat rhetoric as cheap talk and therefore uninformative (see Fearon 1995; Spaniel 2023). In this setting, however, a rhetorical shift may carry downstream costs through an aid relationship, including possible changes in assistance. Repeated public claims can acquire audience and reputational costs, while rhetorical change may affect whether donors maintain support. Observed rhetoric can therefore have substantive relevance even though the evidence cannot identify which influence mechanism produced it. The cross-lagged coevolutionary design helps adjudicate between selection and influence, but it provides neither causal identification from randomisation nor an as-if-random treatment through quasi-experimental techniques. Either aid-for-policy bargaining or social influence could generate the directional associations between donor positions, aid exposure, and recipients’ rhetorical derivatives recovered by the model. Future experimental or quasi-experimental designs could address these limitations. A second limitation concerns measurement. The embedding-based ideal points are measures of expressed rhetoric rather than direct measures of unexpressed views. Missing data, high observation variability, semantic variation, and measurement error remain relevant, and the instability of the Responses surface across measurement draws underscores the need for caution in treating that detailed pattern as robust. Finally, the analysis covers 1990–2012 given data limitations; AidData Core Research Release 3.1 ends in 2013, while the present analysis ends in 2012 because activity-code coverage declines sharply in the final year (see AidData 2017). The political salience of terrorism in this period and American efforts to build an international coalition around the GWoT render it particularly relevant to the research question, but its distinctiveness complicates the assumption of continuity to contemporary international politics and limits generalisability beyond UNGA terrorism rhetoric. As alternative data sources become available, tests of the findings’ generalisability to other periods and issue areas might be conducted. Contemporary relevance therefore rests on a bounded implication. If the historical associations persist, reduced conflict aid would thin the observed link between assistance and recipients’ public rhetoric. ## Conclusion Commentary on the shuttering of USAID and the broader contraction of foreign-aid expenditure among other major donors has noted a crisis of perception surrounding aid: persistent beliefs that it is wasteful or ineffective, even as evidence mounts of its effectiveness at attaining humanitarian, social, public health, and economic goals (Miliband 2026). Tackling misperceptions of aid’s effectiveness on these issues among donor-state publics remains important. In an era of heightened competition and renewed challenges to multilateralism, however, it is also important to counter the perception that aid is merely a philanthropic luxury without geopolitical benefits, or that it may carry geopolitical costs. This study speaks directly to that perception. Conflict-aid relationships are associated with subsequent recipient public rhetoric through jointly informative recipient-side terms whose weights differ by dimension. Recipients move toward donors’ positions most clearly on Causes. Larger inflows are associated with non-state actor framing on Perpetrators, more strongly when donors already occupy those positions and still at high volumes when donors are more state-oriented. On Responses, incoming aid is associated with more legalistic and diplomatic rhetoric, with little independent role for donor positions, though that detailed pattern is sensitive to measurement. These associations concern politically consequential dimensions: contestation over terrorism’s Causes, Responses, and Perpetrators has historically impeded multilateral counterterrorism efforts. For International Relations theory, the primary implications of this study are twofold. First, the evidence is consistent with conflict aid operating as a channel of ongoing political influence over public rhetorical positioning within the multilateral politics of political violence. The design identifies longitudinal associations rather than a causal effect. The empirical result is consistent with more diffuse, general influence than that identified in studies examining transfers used to purchase discrete votes, fund recipient counterterrorism capacity, or reward already aligned partners. Second, aid-related associations depend on the composition of a recipient’s donor relationships. Assessing a purported effect of aid requires consideration of donors’ individual positions, their donor-specific aid contributions, and the relative leverage each donor possesses over a recipient relative to other donors. Research designs should therefore preserve heterogeneity among connected donors rather than treating recipient behaviour as a function of aggregate aid volume alone. Incoming aid can be the identified term on some issues. That does not make donor composition irrelevant on others. Separating the aid-for-policy and social-influence mechanisms that the present design cannot distinguish is one priority for future research. A second is extending the series beyond 2012, whether through a direct extension of the AidData series or through alternative ODA series, would establish whether the relationship holds across the markedly changed donor system, the contemporary turn towards greater multipolarity, declines in the American provision of global public goods, and the distinctive terrorism agenda of the subsequent decade. None of this settles whether such influence ought to be sought. The study documents an association, not an endorsement of using aid to shape another state’s public rhetoric. Leverage exercised through assistance can distort the expressed will of recipient publics, generate resentment, and erode trust; any gains must be understood in light of the power and resource asymmetries of international politics. Contemporary decisions to curtail aid also carry humanitarian costs, so claims about aid’s geopolitical utility should be assessed alongside those costs. Returning to the premise with which this paper opened, the claim that conflict aid buys nothing of geopolitical value is an empirical one, and the evidence assembled here does not support it. Conflict aid appears to be associated with how recipients speak about terrorism on politically contested dimensions. A contraction of conflict aid may weaken one observed channel linking those relationships to how recipients speak about terrorism. ## References
Abrahms, Max. 2007. “Why Democracies Make Superior Counterterrorists.” *Security Studies* 16 (2): 223–53. .
Acharya, Upendra D. 2009. “War on Terror or Terror Wars: The Problem in Defining Terrorism.” *Denver Journal of International Law and Policy* 37 (4): 653–80.
Adhikari, Bimal. 2019. “United Nations General Assembly Voting and Foreign Aid Bypass.” *International Politics* 56 (4): 514–35. .
AidData. 2017. “AidData Core Research Release, Version 3.1.” AidData. .
Aihounton, Ghislain B D, and Arne Henningsen. 2021. “Units of Measurement and the Inverse Hyperbolic Sine Transformation.” *The Econometrics Journal* 24 (2): 334–51. .
Aldrich, Daniel P. 2014. “First Steps Towards Hearts and Minds? USAID’s Countering Violent Extremism Policies in Africa.” *Terrorism and Political Violence* 26 (3): 523–46. .
Alesina, Alberto, and David Dollar. 1998. *Who Gives Foreign Aid to Whom and Why?* No. w6612. National Bureau of Economic Research. .
Alexander, Dan, and Bryan Rooney. 2019. “Vote-Buying by the United States in the United Nations.” *International Studies Quarterly* 63 (1): 168–76. .
Ball, Deborah Yarsike, and Theodore P. Gerber. 2005. “Russian Scientists and Rogue States: Does Western Assistance Reduce the Proliferation Threat?” *International Security* 29 (4): 50–77. .
Banks, Nicola, David Hulme, and Michael Edwards. 2015. “NGOs, States, and Donors Revisited: Still Too Close for Comfort?” *World Development* 66 (February): 707–18. .
Becker, Tal. 2006. *Terrorism and the State: Rethinking the Rules of State Responsibility*. Hart Monographs in Transnational and International Law. Hart.
Boyle, Michael J. 2018. “The Military Approach to Counterterrorism.” In *Routledge Handbook Of Terrorism And Counterterrorism*, 1st ed., edited by Andrew Silke. Routledge. .
Brambor, Thomas, William Roberts Clark, and Matt Golder. 2006. “Understanding Interaction Models: Improving Empirical Analyses.” *Political Analysis* 14 (1): 63–82. .
Brandt, Caroline M., and Jenniina Kotajoki. 2026. “Naming and Shaming Nonstate Armed Groups at the United Nations Security Council.” *International Organization* 80 (2): 431–55. .
Burbidge, John B., Lonnie Magee, and A. Leslie Robb. 1988. “Alternative Transformations to Handle Extreme Values of the Dependent Variable.” *Journal of the American Statistical Association* 83 (401): 123–27. .
Byman, Daniel. 2019. “Counterterrorism Strategies.” In *The Oxford Handbook of Terrorism*, edited by Erica Chenoweth, Richard English, Andreas Gofas, and Stathis N. Kalyvas. Oxford University Press. .
Byman, Daniel. 2023. “How to Think About State Sponsorship of Terrorism.” *Survival* 65 (4): 101–22. .
Capie, David. 2004. “Between a Hegemon and a Hard Place: The ‘War on Terror’ and Southeast Asian-US Relations.” *The Pacific Review* 17 (2): 223–48. .
Carlitz, Ruth D., and Sebastian Ziaja. 2024. “Harmony or Cacophony? A Disaggregated Analysis of Aid Fragmentation.” *Studies in Comparative International Development* 59 (1): 176–98. .
Chambers, Paul. 2004. “U.S.- Thai Relations After 9/ 11: A New Era in Cooperation?” *Contemporary Southeast Asia* 26 (3): 460–79. .
Chyzh, Olga. 2016. “Dangerous Liaisons: An Endogenous Model of International Trade and Human Rights.” *Journal of Peace Research* 53 (3): 409–23. .
Coppedge, Michael, John Gerring, Carl Henrik Knutsen, et al. 2021. “V-Dem Country-Year/Country-Date Dataset v11.” Varieties of Democracy (V-Dem) Project. .
Crelinsten, Ronald. 2018. “Conceptualising Counterterrorism.” In *Routledge Handbook Of Terrorism And Counterterrorism*, 1st ed., edited by Andrew Silke. Routledge. .
Cronin, Patrick M. 2004. “Foreign Aid.” In *Attacking Terrorism: Elements of a Grand Strategy*, 1st ed., edited by Audrey Kurth Cronin and James M. Ludes. Georgetown University Press.
Davies, Shawn, Garoun Engström, Therése Pettersson, and Magnus Öberg. 2024. “Organized Violence 1989–2023, and the Prevalence of Organized Crime Groups.” *Journal of Peace Research* 61 (4): 673–93. .
De Mesquita, Bruce Bueno, and Alastair Smith. 2009. “A Political Economy of Aid.” *International Organization* 63 (2): 309–40. .
Dugard, John. 1974. “International Terrorism: Problems of Definition.” *International Affairs* 50 (1): 67–81. .
Dunning, Thad. 2004. “Conditioning the Effects of Aid: Cold War Politics, Donor Credibility, and Democracy in Africa.” *International Organization* 58 (2): 409–23. .
Duxbury, Scott. 2023. “Network Influence Models.” In *Longitudinal Network Models*. Quantitative Applications in the Social Sciences 192. SAGE Publications. .
Ekengren, Ann-Marie. 2011. “How Ideas Influence Decision-Making: Olof Palme and Swedish Foreign Policy, 1965–1975.” *Scandinavian Journal of History* 36 (2): 117–34. .
Fábián, Sándor, and Andrew Boutton. 2024. “Foreign Military Training and Socialization: An Examination of Human Capital and Norm Transmission Between Allies.” *International Studies Quarterly* 68 (2): sqae040. .
Fearon, James D. 1995. “Rationalist Explanations for War.” *International Organization* 49 (3): 379–414. .
Febrica, Senia. 2010. “Securitizing Terrorism in Southeast Asia: Accounting for the Varying Responses of Singapore and Indonesia.” *Asian Survey* 50 (3): 569–90. .
Findley, Michael G. 2018. “Does Foreign Aid Build Peace?” *Annual Review of Political Science* 21 (1): 359–84. .
Friedrichs, Jörg. 2006. “Defining the International Public Enemy: The Political Struggle Behind the Legal Debate on International Terrorism.” *Leiden Journal of International Law* 19 (1): 69–91. .
Ganor, Boaz. 2002. “Defining Terrorism: Is One Man’s Terrorist Another Man’s Freedom Fighter?” *Police Practice and Research* 3 (4): 287–304. .
Ganor, Boaz. 2018. “The Use of Force to Combat Terrorism.” In *Contemporary Debates on Terrorism*, 2nd ed., edited by Richard Jackson and Daniela Pisoiu. Routledge. .
Gelman, Andrew, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B. Rubin. 2014. *Bayesian Data Analysis*. Third edition. Texts in Statistical Science Series. CRC Press, Taylor & Francis Group. .
Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg, and Håvard Strand. 2002. “Armed Conflict 1946-2001: A New Dataset.” *Journal of Peace Research* 39 (5): 615–37. .
Gregory, Frank. 2005. “The Eu’s Response to 9/11: A Case Study of Institutional Roles and Policy Processes with Special Reference to Issues of Accountability and Human Rights.” *Terrorism and Political Violence* 17 (1-2): 105–23. .
Grewal, Sharan. 2022. “Norm Diffusion Through US Military Training in Tunisia.” *Security Studies* 31 (2): 291–317. .
Halberstam, Malvina. 2003. “The Evolution of the United Nations Position on Terrorism: From Exempting National Liberation Movements to Criminalizing Terrorism Wherever and by Whomever Committed.” *Columbia Journal of Transnational Law* 41 (3): 573–84. .
Hensel, Paul R., and Sara M. Mitchell. 2007. “The Issue Correlates of War (ICOW) Project Supplementary Data Set: Colonial History Data Set.” Version V2. Harvard Dataverse. .
Hoeffler, Anke, and Verity Outram. 2011. “Need, Merit, or Self‐Interest—What Determines the Allocation of Aid?” *Review of Development Economics* 15 (2): 237–50. .
Jankin, Slava, Alexander Baturo, and Niheer Dasandi. 2017. “United Nations General Debate Corpus 1946-2022.” Version 11.0. Harvard Dataverse. .
Jenkins, Brian Michael, and J. J. Johnson. 1975. *International Terrorism: A Chronology, 1968-1974*. R-1597-DOS/ARPA. RAND Corporation. .
Jost, Tyler, and Joshua D. Kertzer. 2024. “Armies and Influence: Elite Experience and Public Opinion on Foreign Policy.” *Journal of Conflict Resolution* 68 (9): 1769–97. .
Jost, Tyler, Joshua D. Kertzer, Eric Min, and Robert Schub. 2025. “The Role of Advisors: How Ideas Shape International Crises.” In *Inside the Situation Room: The Theory and Practice of Crisis Decision-Making*, edited by Hillary Rodham Clinton and Keren Yarhi-Milo. Oxford University Press. .
Joyce, Renanah Miles. 2022. “Soldiers’ Dilemma: Foreign Military Training and Liberal Norm Conflict.” *International Security* 46 (4): 48–90. .
Kanbur, Ravi. 2006. “Chapter 26 The Economics of International Aid.” In *Handbook of the Economics of Giving, Altruism and Reciprocity*, edited by Serge-Christophe Kolm and Jean Mercier Ythier, vol. 2. Elsevier. .
Kinne, Brandon J, Juan F. Tellez, Anya Stewart, et al. 2026. “Transnational Networks and Interstate Competition: How Support for Nonstate Actors Increases Conflict Between States.” *International Organization* 80 (2): 281–323. .
Knapp, Andrea. 2026. “Speech Is Silver, Resolutions Are Gold: State Influence at the United Nations Security Council (1990–2019).” *Conflict Management and Peace Science*, February 4, 07388942261417494. .
Kozlowski, Austin C., Matt Taddy, and James A. Evans. 2019. “The Geometry of Culture: Analyzing the Meanings of Class through Word Embeddings.” *American Sociological Review* 84 (5): 905–49. .
Kruglanski, Arie W., Martha Crenshaw, Jerrold M. Post, and Jeff Victoroff. 2007. “What Should This Fight Be Called?: Metaphors of Counterterrorism and Their Implications.” *Psychological Science in the Public Interest* 8 (3): 97–133. .
Kydd, Andrew H., and Barbara F. Walter. 2006. “The Strategies of Terrorism.” *International Security* 31 (1): 49–80. .
Laksmana, Evan A., and Michael Newell. 2020. “Political Violence and Counterterrorism: Disputed Boundaries of a Postcolonial State.” In *Non-Western Responses to Terrorism*, edited by Michael J. Boyle. Manchester University Press. .
Lee, James. 2022. “Foreign Aid, Development, and US Strategic Interests in the Cold War.” *International Studies Quarterly* 66 (1): sqab090. .
Leifeld, Philip. 2013. “Reconceptualizing Major Policy Change in the Advocacy Coalition Framework: A Discourse Network Analysis of German Pension Politics.” *Policy Studies Journal* 41 (1): 169–98. .
Levitt, Geoffrey. 1986. “Is "Terrorism" Worth Defining?” *Ohio Northern University Law Review* 13 (1): 97–116.
Mani, Rama. 2004. “The Root Causes of Terrorism and Conflict Prevention.” In *Terrorism and the UN: Before and After September 11*, edited by Jane Boulden and Thomas G. Weiss. Indiana University Press. .
Maoz, Zeev, and Belgin San-Akca. 2012. “Rivalry and State Support of Non-State Armed Groups (NAGs), 1946–2001.” *International Studies Quarterly* 56 (4): 720–34. .
Miliband, David. 2026. “One Year After USAID: How to Actually Fix Foreign Aid.” *Foreign Policy*, February 19. .
Molenaers, Nadia, Sebastian Dellepiane, and Jorg Faust. 2015. “Political Conditionality and Foreign Aid.” *World Development* 75 (November): 2–12. .
Morgenthau, Hans. 1962. “A Political Theory of Foreign Aid.” *American Political Science Review* 56 (2): 301–9. .
Musk, Elon. 2025. *Post on x*. .
National Consortium for the Study of Terrorism and Responses to Terrorism. 2021. “Global Terrorism Database.” National Consortium for the Study of Terrorism and Responses to Terrorism. .
Niezink, Nynke M. D., and Tom Snijders. 2017. “Co-Evolution of Social Networks and Continuous Actor Attributes.” *The Annals of Applied Statistics* 11 (4). .
Niezink, Nynke M. D., Tom A. B. Snijders, and Marijtje A. J. van Duijn. 2019. “No Longer Discrete: Modeling the Dynamics of Social Networks and Continuous Behavior.” *Sociological Methodology* 49 (1): 295–340. .
Nincic, Miroslav. 2006. “The Logic of Positive Engagement: Dealing with Renegade Regimes.” *International Studies Perspectives* 7 (4): 321–41. .
Nincic, Miroslav. 2010. “Getting What You Want: Positive Inducements in International Relations.” *International Security* 35 (1): 138–83. .
OECD. 2023. *Converged Statistical Reporting Directives for the Creditor Reporting System (CRS) and the Annual DAC Questionnaire*. Unclassified Document DCD/DAC/STAT(2023)9/FINAL. Development Co-operation Directorate, Development Assistance Committee. .
Pemstein, Dan, Kyle Marquardt, Eitan Tzelgov, et al. 2020. “The v-Dem Measurement Model: Latent Variable Analysis for Cross-National and Cross-Temporal Expert-Coded Data.” *SSRN Electronic Journal*, ahead of print. .
Peterson, M. J. 2004. “Using the General Assembly.” Chap. 8 in *Terrorism and the UN: Before and After September 11*, edited by Jane Boulden and Thomas G. Weiss. Indiana University Press.
Reuning, Kevin, Michael R. Kenwick, and Christopher J. Fariss. 2019. “Exploring the Dynamics of Latent Variable Models.” *Political Analysis* 27 (4): 503–17. .
Ripley, Ruth M., Tom A. B. Snijders, Zsófia Boda, and András Vörös. 2026. *Manual for RSiena*. University of Oxford, Department of Statistics; Nuffield College, released March 21. .
Risse, Thomas, and Kathryn Sikkink. 1999. “The Socialization of International Human Rights Norms into Domestic Practices: Introduction.” In *The Power of Human Rights: International Norms and Domestic Change*, edited by Thomas Risse, Stephen C. Ropp, and Kathryn Sikkink. Cambridge Studies in International Relations. Cambridge University Press. .
Rosendorff, B. Peter, and Todd Sandler. 2004. “Too Much of a Good Thing? The Proactive Response Dilemma.” *Journal of Conflict Resolution* 48 (5): 657–71. .
Saul, Ben. 2006. *Defining Terrorism in International Law*. Oxford Monographs in International Law. Oxford University Press. .
Schvitz, Guy, Luc Girardin, Seraina Rüegger, Nils B. Weidmann, Lars-Erik Cederman, and Kristian Skrede Gleditsch. 2022. “Mapping the International System, 1886-2019: The CShapes 2.0 Dataset.” *Journal of Conflict Resolution* 66 (1): 144–61. .
Spaniel, William. 2023. *Formal Models of Crisis Bargaining: Applications in the Politics of Conflict*. 1st ed. Cambridge University Press. .
Stoltz, Dustin S., Marshall A. Taylor, and Jennifer S. K. Dudley. 2025. “A Tool Kit for Relation Induction in Text Analysis.” *Sociological Methods & Research* 54 (2): 565–604. .
Terman, Rochelle L. 2023. *The Geopolitics of Shaming: When Human Rights Pressure Works—and When It Backfires*. Princeton Studies in International History and Politics. Princeton University Press. .
Tierney, Michael J., Daniel L. Nielson, Darren G. Hawkins, et al. 2011. “More Dollars Than Sense: Refining Our Knowledge of Development Finance Using AidData.” *World Development* 39 (11): 1891–906. .
United Nations General Assembly. 2003. *Revitalization of the Work of the General Assembly*. A/RES/58/126. United Nations. .
United Nations High Commissioner for Refugees. 2023. “UNHCR Refugee Population Statistics Database.” .
United States of America. 1993. *Measures to Eliminate International Terrorism: Report of the Secretary-General*. General Assembly Document A/48/267/Add.1. United Nations. .
U.S. House of Representatives Committee on Oversight and Government Reform, Subcommittee on Delivering on Government Efficiency. 2025. *America Last: How Foreign Aid Undermined u.s. Interests Around the World*. Hearing before the Subcommittee on Delivering on Government Efficiency of the Committee on Oversight and Government Reform, U.S. House of Representatives. .
Warren, T Camber. 2016. “Modeling the Coevolution of International and Domestic Institutions: Alliances, Democracy, and the Complex Path to Peace.” *Journal of Peace Research* 53 (3): 424–41. .
Whitaker, Beth Elise. 2007. “Exporting the Patriot Act? Democracy and the ‘War on Terror’ in the Third World.” *Third World Quarterly* 28 (5): 1017–32. .
World Bank. 2025. “GDP per capita (constant 2015 US\$).” World Development Indicators. World Bank. .
Yarhi-Milo, Keren. 2013. “Tying Hands Behind Closed Doors: The Logic and Practice of Secret Reassurance.” *Security Studies* 22 (3): 405–35. .
Zhang, Yanzhao, Mingxin Li, Dingkun Long, et al. 2025. “Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.” Version 3. Pre-published. .
Ziaja, Sebastian. 2020. “More Donors, More Democracy.” *The Journal of Politics* 82 (2): 433–47. .