---
title: "AI on the Battlefield? Revisiting Public Support for LAWs"
authors: ["Matthew DiGiuseppe", "Katrin Paula", "Tobias Rommel"]
year: 2025
status: "Working paper (revise & resubmit)"
topics: ["AI & Politics"]
url: "https://www.matthewdigiuseppe.com/papers/ai-battlefield-revisiting-public-support.md"
links: {"Preprint on OSF": "https://osf.io/preprints/socarxiv/s8ab5_v1"}
full_text: true
---

# AI on the Battlefield? Revisiting Public Support for LAWs

DiGiuseppe, M., Paula, K., & Rommel, T. (2025). AI on the Battlefield? Revisiting Public Support for LAWs.

- Status: Working paper (revise & resubmit)
- Topics: AI & Politics
- Preprint on OSF: https://osf.io/preprints/socarxiv/s8ab5_v1
- Listed on: https://www.matthewdigiuseppe.com/#research

<!-- END OF GENERATED HEADER: edit freely below this line; scripts/build_agent_files.py keeps it -->

## Abstract

Why do citizens support the development of lethal autonomous weapons (LAWs)? In this paper, we revisit the claim that support for taboo weapons is driven by the logic of external threats, now that individuals are more familiar with artificial intelligence and the international environment has grown more uncertain. In our study, we probe public attitudes toward both the development of LAWs and international cooperation to address the risks. Based on original survey data from 2,500 respondents in the US, we provide an observational as well as an experimental analysis. Observationally, general support for AI and favorable views of militarism and US hegemony as well as perceived competition from other countries are highly correlated with support for the use of lethal AI-weapons. Experimentally, we show that information that rivals or allies are already developing such weapons increases support for LAWs. Additionally, we find that competition from other countries does not influence support for global cooperation to limit AI on the battlefield. Overall, our analysis supports the idea that concerns for security have the power to override opposition to these taboo weapons.

## Full text

> Extracted automatically from the preprint: https://osf.io/preprints/socarxiv/s8ab5_v1. Tables, figures and equations may be garbled or missing; quote the PDF, not this text.

**AI on the Battlefield? Revisiting Public Support for LAWs**

Matthew DiGiuseppe* · Katrin Paula

Leiden University · Technical University of Munich

Tobias Rommel

Technical University of Munich

Version: May 1, 2025; Word Count: 9,832.

Abstract Why do citizens support the development of lethal autonomous weapons (LAWs)? In this paper, we revisit the claim that support for taboo weapons is driven by the logic of external threats, now that individuals are more familiar with artificial intelligence and the international environment has grown more uncertain. In our study, we probe public attitudes toward both the development of LAWs and international cooperation to address the risks. Based on original survey data from 2,500 respondents in the US, we provide an observational as well as an experimental analysis. Observationally, general support for AI and favorable views of militarism and US hegemony as well as perceived competition from other countries are highly correlated with support for the use of lethal AI-weapons. Experimentally, we show that information that rivals or allies are already developing such weapons increases support for LAWs. Additionally, we find that competition from other countries does not influence support for global cooperation to limit AI on the battlefield. Overall, our analysis supports the idea that concerns for security have the power to override opposition to these taboo weapons.

*Corresponding author: [email removed]

Smart algorithms and large language models have become an important and frequently used tool for business decisions and among the general population (Jo, 2023; von Garrel and Mayer, 2023). In many domains, the capabilities of software are matching or surpassing human intelligence opening the possibility for scientific and productivity advancements (Candrian and Scherer, 2022). Yet, artificial intelligence (AI) is by no means confined to peaceful purposes alone, but will likely define the future of warfare and deterrence. The increased use of digital technologies in conflict, such as cyber, robotics, and drones (Horowitz, 2020) already characterizes the potential of using AI for military purposes. Going one step further, recent developments in ongoing conflicts or wars, such as Israel’s war in Gaza or the Russian invasion of Ukraine, indicate that modern weapons systems are actively deployed and further developed on the ground with the help of AI-generated information (Frenkel and Odenheimer, 2025; Bergengruen, 2024).

AI-weapons systems offer tactical benefits in wars, such as reducing the direct risk to soldiers, allowing faster reaction times in combat, and more precise targeting (Horowitz, 2016). But there are also drawbacks and moral quandaries: a lack of human judgment could lead to unintended attacks and it remains unclear who should be held accountable in such cases, which results in ethical concerns whether autonomous systems should even be allowed to make life-or-death decisions (Rosendorf, Smetana, and Vranka, 2024). Governments acknowledged the military advantages of AI-weapons over conventional warfare, yet 166 countries voted for UN Resolution 79/62 in December 2024 to strengthen international regulation of autonomous weapons, illustrating the tension between normative opposition and military utility in weapons development (Press, Sagan, and Valentino, 2013). Nevertheless, countries now unilaterally commit sizable parts of their budgets to foster the development of lethal autonomous weapons (LAWs, see Humble, 2024). It thus seems to only be a matter of time, until the rapid advancement of AI makes waging war autonomous, for example in selecting who to target as well as ultimately applying force without further human intervention (McKernan and Davies, 2024).

While many governments are proceeding with plans to adopt or have already adopted LAWs, the public is far less enthusiastic. Large public opinion surveys show that the public has a strong aversion to lethal autonomous weapons, which seems to be driven by both moral and technical concerns (Zhang and Dafoe, 2020). Apart from the US, a recent Ipsos (2021) study in 28 countries indicates that majorities in all countries oppose lethal autonomous weapons.[^1] This aligns with research that has uncovered a general algorithmic aversion among the public (Mahmud et al., 2022).

Beyond observational studies, a growing literature has utilized survey experiments to understand the elasticity of opposition to AI-based weapons systems and the circumstances under which citizens will support their development (see, e.g., Young and Carpenter, 2018; Rosendorf, Smetana, and Vranka, 2022, 2024; Zwald, Kennedy, and Ozer, 2024; Kreps et al., 2023). In a pioneering study, Horowitz (2016) was the first to examine the topic by surveying a convenience sample of unrepresentative US crowd workers in 2015. He found that these respondents were more likely to support LAWs when a) they were more effective in protecting US forces and b) when they were led to believe that non-state actors and other countries are also developing these weapons. Since then, work by Young and Carpenter (2018) shows that priming citizens with iconic sci-fi movies that introduced AI in pop culture years before the current popular uptake does not move attitudes towards autonomous weapons much. Probing the ethical side of deploying AI-weapons, directly altering the perceptions of errorrates of AI-weapons or varying the autonomy of these systems increases support and trust in LAWs (Rosendorf, Smetana, and Vranka, 2022, 2024), but decreases the potential blame put on human operators (Zwald, Kennedy, and Ozer, 2024).

In this study, we revisit the role of external factors in support for overcoming the initial public hesitancy on lethal AI-weapons, by taking Horowitz’s (2016) trailblazing study as our main point of departure. We also focus on the United States, because the US is the most powerful country in terms of its military capacity, but now more than ever – at least more than 10 years ago – faces challengers (such as China and Russia) as well as challenges (such as new weapons technology) in the international arena. The United States is thus uniquely suitable to study how the public reacts to changes in the national security environment, which we regard as the most important external factor. Now that individuals are more familiar with AI in general and the international environment has, simultaneously, grown more uncertain, we find it necessary to take another close look at public support for LAWs.

Beyond emulating parts of the Horowitz (2016) study a decade later, we extend its theoretical scope and improve its research design in several ways to provide a more nuanced analysis: First, and most importantly, we add treatment arms to examine the extent by which external motivation increases support for lethal AI-weapons. Horowitz’s original survey treated respondents with the notion that various, undefined actors abroad were developing weapons. Our treatments single out US allies (Europeans) and adversaries (China and Russia) as key actors also developing AI-weapons in separate treatments. As such, our study provides more nuance on whether actions by rivals or allies are driving a reluctant embrace of LAWs. We also add a third treatment arm to empirically distinguish whether the threat to national security that stems from the fact that other countries are also developing AI-weapons is really just picking up the public’s desire to retain primacy and global leadership. The central aim of our study is, thus, to understand better what exactly it is about external threats to national security that moves the public’s position from a taboo towards support for AI-weapons development.

Second, we not only probe the public about whether their country should develop LAWs, but also ask about public support for international collaboration to manage AI-weapons to gain further insight on support for cooperation. Third, we run the analysis on a sample that reflects the US population in terms of age, gender, and partisanship. Fourth, we take steps to increase the power of the experiment by reducing measurement error, including quasi pre-treatment variables, and increasing the sample size. The increased statistical power also allows us to better examine how the emergence of consumer AI and the use of AI in other realms of government unrelated to weapons influence support for LAWs and condition information about external threats. Lastly, our survey is designed such that we are able to use a pre-treatment question on autonomous weapons to examine correlates of support for autonomous weapons on an observational basis.

Beyond the study of lethal autonomous weapons directly, our study gives greater insight into how external forces shape the development and acceptance of weapons and tactics that are deemed undesirable by majorities of the population. We show that a desire to retain leadership absent the action of other states is not a salient motivating force. Instead, we find that direct action by other states to develop weapons systems increases support for LAWs. Yet interestingly, we find an increase in public support for LAWS irrespective of whether the threat to national security emanates from friends or foes. Hence, changes in the national security environment, regardless of their origin, remove a key constraint to the adoption of ‘unwanted’ weapons systems. Our study thus also contributes to understanding how new technologies might affect the international order (see from the perspective of nuclear weapons, Gartzke and Kroenig, 2017).

### External Threats, Public Opinion, and LAWs

Some lethal weapons (chemical, biological, and nuclear) are subject to a taboo that influences policy around their use and development (Price, 1995; Press, Sagan, and Valentino, 2013; Koch and Wells, 2021; Sagan and Valentino, 2025). Recently, lethal autonomous weapons also find themselves in this class of weapons given the moral and ethical ambiguity around their use, as exemplified by the many attempts to ban them entirely or slow down their development (Maas, 2019; Rosert and Sauer, 2021; Sauer, 2016). A key policy question going forward thus centers on the elasticity of these taboos – among political elites, but critically also among citizens (see also Bloom et al., 2020). Are citizens willing to support taboo weapons under certain conditions? And if so, what conditions or motivations impact their support? These questions hold normative appeal for those advocating for the necessity of AI-weapons and those arguing against their development. However, they are also important for understanding what motivates the public’s trade-offs on the risks and benefits of lethal technology more generally.

Following the intuition of Horowitz (2016), we suggest that the fear of falling behind and heightened vulnerability are likely to increase support for the use of AI-weapons. Hence, we argue that in situations where one expects to lose ground internationally with respect to others, developments that affect the capabilities of other countries – especially in the realm of lethal weapons technology – are perceived as threats to national security. This line of reasoning is part of the rationale that also underpins Walt’s (1985) notion that countries bandwagon against threats and not balance for power. Threats thus play an important role in human behavior and, consequently, international relations (Landau-Wells, 2024).

Building on this image of international politics, we further argue that a similar dynamic may play out in public opinion where external threats should influence both elites and the public (Kertzer, 2022). While these micro-foundations have only recently been granted significant attention, the idea that the behavior of other states motivates politicians to respond in kind is the foundation of arms race behavior and the security dilemma (Jervis, 1978, 2017). Much like elites are motivated to respond to the perceived threat of other countries’ programs to develop advanced weapons systems, we expect the public to exhibit a similar reaction.

This is in line with recent research, which has suggested that priming individuals to think about threats and actual increases in threats amplifies support for both military spending (DiGiuseppe, Aspide, and Becker, 2024) and alliances (Becker et al., 2024). Feeling threatened also turns citizens into proponents of more hawkish foreign policies (Malhotra and Popp, 2012; Gadarian, 2010; Kaltenthaler, Silverman, and Dagher, 2020). The impact of threat perceptions is by no means confined to foreign and security policy alone. Leaders aim at higher levels of internal cohesion in light of threats to territorial integrity (Darden and Mylonas, 2016), while the public responds with greater support for torture (Conrad et al., 2018). Foreign threats, especially from China, also reduce domestic political polarization under some conditions (Schwartz and Tierney, 2025), even though external threat may not reduce affective polarization across the board (Myrick, 2021).

Taking into account that citizens are sensitive to national security concerns, then it should follow that information that implies heightened costs to holding steadfast to moral or practical opposition to LAWs will diminish adherence to that opposition. The development of lethal weapons systems by other countries should be both sizable as well as immediately visible enough to elevate perceptions of being threatened by other countries for the greater public. Several mechanisms for this response are plausible. On the one hand, such behavior can result from a rational calculation about the costs of inaction versus the costs of permitting taboo weapons that carry their own risks. In the case of AI, that might be a loss of control of weapons that lead to attacks on soldiers or a nation’s own citizens. On the other hand, the motivation may also be driven by fear of being exposed to a threat that causes an irrational motivation to overlook those same costs and make more utilitarian judgments (Tao et al., 2023). In brief, we suggest that fear is greater than morality among the public, as soon as external threats affect perceptions about national security. Survey methods available to us make it difficult to untangle these potential mechanisms. However, the existence of multiple potential mechanisms gives strength to the plausibility of the relationship.[^2]

Hypothesis 1: Informing respondents that other countries already started to develop AI-weapons technology increases support for autonomous lethal weapons systems compared to those not given this information.

While the prospect that other states will also develop autonomous lethal weapons may increase support for taboo weapons, we also expect variance in the strength of this effect in the response depending on which other countries are already involved. Rival states which develop threatening weapons should evoke a stronger response than allies doing the same. This is due to the fact that any reference to a geopolitical rival or enemy should lead to an even higher threat perception that citizens will want to counterbalance, in order to deter other countries from using weapons technology against their own country.

Thinking in terms of allies and enemies, however, implies a certain level of perspective taking, whereby threats from rivals are perceived as more severe than potential competition from allies. However, perspective-taking may be less effective in the context of salient national security issues (Kertzer, Brutger, and Quek, 2024). The public may also be motivated to counterbalance allies, since alliances can involve internal tensions and strategic competition (Niou and Zeigler, 2019), and may become unstable over time. Irrespective of whether there is a difference in the public’s reaction to allies or rivals developing LAWs, we expect that public opposition to autonomous weapons systems is either way subject to change when national security is perceived to be at risk due to the actions of other countries, including allies.

Hypothesis 2a: Informing respondents that US rivals already started to develop AI-weapons technology increases support for autonomous lethal weapons systems compared to those not given this information.

Hypothesis 2b: Informing respondents that US allies already started to develop AI-weapons technology increases support for autonomous lethal weapons systems compared to those not given this information.

Taken together, we expect the national security dilemma to be most pronounced in the case of rival states. If we additionally see that the impact of rivals developing LAWs is substantively larger, then it would indicate that perceived threat is the primary motivator of acceptance of taboo weapons. Conversely, if there is no difference between allies and rivals, it suggests a different causal path related to loss of status or general competition.

In a final step, we also contrast tangible external threats with more general changes in the national security environment, irrespective of any direct threat. We do so because any change in public opinion towards lethal autonomous weapons might simply be due to the public’s desire to retain primacy and global leadership. Even simply suggesting that US hegemony is at risk – even without a concrete threat – may sway public opinion toward supporting government use of generally unpopular weapons. Specifically, we argue that support for military AI is more likely when respondents believe US hegemony is threatened and that developing such weapons is crucial for maintaining international power.

Hypothesis 3: Informing respondents that policymakers see AI-weapons technology as a crucial way for the US to maintain its role as a global leader increases support for autonomous lethal weapons systems compared to giving no information.

### Empirical Evidence

To test these hypotheses experimentally as well as conduct additional exploratory analyses on correlates of support for lethal autonomous weapons, we fielded a survey on attitudes towards AI on the Prolific platform in August of 2024. Prolific offers both convenience samples and quota samples, drawing from their pool of respondents. Our sample is the latter and is reflective of the US population along the dimensions of age, gender, and partisanship. To satisfy power requirements, we recruited a sample of 2,500 respondents. To avoid priming respondents, we asked pre-treatment questions about national security, then a distractor module about interest rates before returning to our experimental treatments. This setup allows us to explore the correlates of attitudes toward LAWs and allows us to include the pre-treatment questions in our experimental analysis to increase the precision of our estimates. We being our exploration of the survey findings with the former.

#### Observational Evidence of Support for LAWs

We start our analysis by looking first at the observational correlations within the data. Before we assigned respondents to treatment arms and asked the main outcome variables, we asked all respondents about their support for AI in various policy areas including health, domestic security, the promotion of local officials, and in the use as lethal weapons. Specifically we asked respondents to indicate their support for the ‘Deployment of autonomous lethal weapons in the battlefield.’ Respondents could answer on a 4-point scale for very unsupportive to very supportive. Figure 1 shows the raw distribution of the variable. Consistent with other studies, the public shows little overall support for the use of these weapons on the battlefield – a majority is ‘very’ or ‘somewhat unsupportive.’ Yet, there is a sizable part of respondents that seems to be at least ‘somewhat supportive’ to the deployment of LAWs.

Based on this dependent variable, Table 1 presents the results of four linear regression models predicting support for LAWs each adding a different set of covariates. We standardize all variables so that readers can easily compare the substantive relationship with the dependent variable. Across all models, we include basic demographics and party identification. In subsequent models, we include variables that proxy concern for threats. These include an

Figure 1: Pre-treatment support for Lethal AI-weapons on the battlefield, N=2525 additive index of militarism that relies on two questions about the necessity of using force in international relations.[^3] Consistent with our efforts to disentangle the impact of threats to national security from a desire to retain global leadership, we also asked about support for US hegemony in international politics.[^4] We also captured support for international cooperation with an additive index of three questions that centered around reespondents’ assessment of specific foreign policy goals.[^5] Finally, we include variables that capture perceptions of other states, based on the following question: How much of a challenge do you think the following countries pose for the power of the US? Respondents could reply not at all to a great deal on a 5-point scale for the following states: (1) EU members, (2) Russia, (3) China, and (4) Canada. We include these answers both as an additive index to capture the degree by which the respondent views the US in competition and individually for each country. For exploratory purposes, we also include an additive index capturing general attitudes toward AI to probe how acceptance of AI in other areas of government (employment decisions of local government officials and identification of individual risk of diseases) affect the willingness to use AI also on the battlefield.[^6]

Table 1 presents the results of the observational analysis. First of all, we see that even with an encompassing set of covariates the R[^2] is rather low, maxing out at 0.23. This suggests that quite a bit of US attitudes on LAWs are unexplained by the measures of foreign policy attitudes, demographics, and general attitudes towards the use of AI. Notably, several variables have consistent correlations with support for LAWs: First, we see that there is a gender difference, where women tend to be more skeptical towards the use of AI in the battlefield. There is mixed evidence for a partisan effect (where higher values imply stronger attachment to the Republican party), at least judged by the changes in statistical significance. However, that appears to be due to general foreign policy attitudes across the political spectrum. When we include respondents support for international cooperation, militarism, the competition index, and US hegemony, any partisan differences seem to disappear.

All measures of foreign policy views have a positive relationship with support for LAWs on the battlefield. We take this as first tentative evidence for a possible relationship between support for AI-weapons and national security concerns. However, these correlations are little help in indicating which of these concerns matters most. General militarism, support for hegemony, and viewing other states as competitors have a similar substantive relationship with LAW support. However, concerns with individual states as competitors are surprisingly statistically insignificant with the odd exception of perceptions of Canada.[^7] Even if we exclude other national security related variables that might confound the relationship, attitudes toward particular countries does not impact our findings. Notably, we see that

Table 1: Correlates of Support for Lethal AI Weapons

```text
1  2  3  4
Woman  −0.101*** −0.100***  −0.040*  −0.057**
(0.019)  (0.019)  (0.018)  (0.018)
Party ID  0.136***  0.013  0.037+  0.132***
(0.020)  (0.022)  (0.020)  (0.019)
Age  0.124***  0.051*  0.048*  0.123***
(0.020)  (0.020)  (0.019)  (0.019)
College Degree  0.031  0.028  0.006  −0.003
(0.020)  (0.019)  (0.018)  (0.019)
Support Int’l Cooperation (Index)  −0.046*  −0.062***
```

(0.020)  (0.019)

```text
Support for Militarism (Index)  0.216***  0.193***
```

(0.025)  (0.023)

```text
Support for US Hegemony  0.109***  0.090***
```

(0.024)  (0.022)

```text
Competition (Index)  0.094***  0.080***
```

(0.021)  (0.019)

EU as Competitor  0.030 (0.024) Russia as Competitor  0.039+ (0.023) China as Competitor  −0.023 (0.022) Canada as Competitor  0.106*** (0.023)

```text
AI in Gov Support  0.340***  0.351***
```

(0.019)  (0.019)

```text
Num.Obs.  2525  2525  2525  2525
R2  0.049  0.126  0.236  0.194
```

+ p <0.1, * p <0.05, ** p <0.01, *** p <0.001.

support for AI in other, non-military areas of governance are significantly correlated with support for AI-weapons.

In all, these observational results suggest that support for LAWs correlated highly with respondents’ more general views on foreign policy. Most notably, increases in support for AI-weapons follows from concerns about military primacy, support for US hegemony, and the perception that other countries step up as competitors. Unsurprisingly, attitudes towards the military use of AI are also related to broader support for government use of AI outside of military contexts. Hence, we now know that the public generally forms preferences towards AI-weapons development in line with their general views on foreign policy. We now proceed to our experimental research design and analysis, in which we put the impact of external threats to a rigorous empirical test.

#### Experimental Evidence and Results of External Threats

Following the pre-treatment questions that allow the observational analysis, we present each respondent with the following scenario and briefly describe the potential benefits and drawbacks of lethal autonomous weapons (while randomizing the order of both):

The Department of Defense is thinking about using computer algorithms and artificial intelligence (AI) to increase national security, focusing especially on lethal autonomous weapons systems. Autonomous weapons are weapons that can select and attack targets without direct human control. They can take the form of drones or robots that can identify and engage enemy soldiers or targets on their own.

Military and foreign policy experts have noted a few drawbacks of autonomous lethal weapons systems. They include: Lack of human judgment could lead to unintended attacks; unclear accountability if autonomous weapons make mistakes; ethical concerns about machines making life-or-death decisions. Autonomous lethal weapons systems also have benefits according to military and foreign policy experts. They include: Less risk to US soldiers; faster reaction times in combat; more precise targeting to reduce civilian casualties.

The general introduction serves to anchor respondents on the same costs and benefits. After this, we randomly assigned respondents to one of four treatment conditions. We made sure to keep the wording of these conditions the same and only change the language on the national security environment:

Table 2: Treatment Conditions and Associated Text Condition  Text Shown to Participants Control  As you can see, foreign policy and military experts are, thus, divided on the use of AI on the battlefield.

Global Leader  As you can see, foreign policy and military experts are, thus, divided on the use of AI on the battlefield. Yet, many policymakers see the technology as a crucial way for the US to maintain its role as a global leader.

Ally Competition As you can see, foreign policy and military experts are, thus, divided on the use of AI on the battlefield. Yet, many policymakers see the technology as a crucial way for the US to maintain its role as a global leader, because European allies have already started to develop such weapons.

Rival Competition As you can see, foreign policy and military experts are, thus, divided on the use of AI on the battlefield. Yet, many policymakers see the technology as a crucial way for the US to maintain its role as a global leader, because China and Russia have already started to develop such weapons.

In addition to invoking the threat perception that specific countries have already started to develop LAWs (‘Rival Competition’ and ‘Ally Competition’), we also include a condition on general ‘Global Leader’ that aims to test if threat and competition are the primary drivers of support for AI-weapons or if a desire to be a global leader is already sufficient to increase support for these weapons.

Following the treatments, we ask respondents to answer several questions that serve as our main outcome variables, which we will also combine into an additive index. The first two variables posit unilateral action by the US government:

Outcome 1: Do you agree or disagree that the United States should develop autonomous (AI) weapons systems? [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

Outcome 2: Do you agree or disagree with the following statement: It is important for the US to unilaterally stop the development of its own autonomous (AI) weapons systems. [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

We next ask respondents two more outcome questions that capture the same sentiment, but require respondents to consider cooperation with other countries. The aim here is to examine the elasticity between threat perceptions and international cooperation. If respondents are provided an opportunity to resolve threats with cooperation, it would suggest that the relationship between threat and unilateral development of weapons could be addressed through collaborative measures.

Outcome 3: The United Nations (UN) is currently discussing a potential treaty to ban or restrict on the use of autonomous weapons. This would be similar to a ban or restriction on the use land mines and the prohibition on the use of nuclear weapons and would also apply to other countries. Do you agree or disagree that the United States should sign a treaty to restrict the use of autonomous weapons? [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

Outcome 4: Do you agree or disagree with the following statement: The U.S. should work hard to cooperate with the international community to develop safe autonomous lethal weapons, even if doing so requires giving up some of the U.S.’s advantages. Cooperation could include collaborations research labs and creating and committing to common safety standards for AI. [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

Figure 2 presents the distribution of each outcome variable in the control group (N=651). We see that around 37%-40% of respondents are somewhat or strongly supportive of developing AI-weapons or not stopping their development. The rest of the public is either neutral or opposed to such weapons. This is consistent with our pre-treatment measure. Comparatively, there is strong support for international cooperation to either restrict AI-weapons or cooperate to develop safe AI-weapons.

Figure 2: Outcome Distribution: This figure shows the distribution of each outcome variable as answered by those in the control group (N=651).

We begin to examine the results of our experiment by looking at support for developing or unilaterally stopping the development LAWs (outcomes 1 and 2) and the additive index of the two (after rescaling). Figure 3 presents the ATE for each combination of the outcomes and treatment conditions. The outcome variables are standardized. As such, the ATEs can be interpreted substantively as a percentage of a standard deviation.

Figure 3: Treatment Effects on Support for Developing and Banning LAWs: The figure shows the coefficients and 95% confidence intervals for each of the treatments for three different outcome variables: Support for developing AI-weapons, support banning AI- Weapons, and an additive index for both these outcomes once they are rescaled.

Our experiment reveals that the information that either rivals or allies have already started to develop AI-weapons has a significant impact on all outcome variables related to the unilateral development of AI-weapons. Substantively, researchers usually expect small effects from information treatments of about 10% of a standard deviation (Coppock, 2023). Here, we find that the effect of the rival treatment is about 25% and the effect of allies treatment is about 20% of a SD. We find that, despite strong statistical power, the ‘Global Leader’ treatment does not have a significant impact on support for developing LAWs. These results support Hypotheses 1, 2a, and 2b.

Attention to the information in treatments is always a concern for any survey and survey experiment. While we tried our best to engage the respondents, other incentives may have led some to complete the survey as quick as possible. To address concerns that inattentiveness may diminish the substantive size of the effect, we asked respondents to recall information from the treatments at the very end. Specifically, we asked: Earlier in the survey, we also mentioned that some countries have already started to develop lethal autonomous weapons. Respondents could answer with ‘Brazil and South Africa’, ‘China and Russia’, ‘European allies’, ‘Japan and South Korea’ or ‘We didn’t name any country’. Table 3 shows that 80% could recall the adversaires treatment. Far less could recall the allies treatment – which is surprising given the similarities of the substantive effects. In the control condition as well as the global leader condition, we did not name any country; and between 75-78% of respondents were able to recall that correcly.

Table 3: Recall Percentage by Treatment Condition

```text
Brazil and  China and European  Japan and  We didn’t
South Africa  Russia  allies  South Korea name any country
Control  0.00  21.30  2.50  0.80  75.30
Enemy  0.00  79.70  1.30  0.30  18.70
Ally  0.20  16.60  45.40  0.80  37.10
Leader  0.30  19.20  2.20  0.60  77.60
```

The recall is useful beyond assessing inattentiveness. It can also help correct for it. We use the recall responses as indication that the respondents ‘received’ the treatments and estimated a two-stage least-squares model to estimate the complier average causal effect (CACE). The treatment assignment serves as an ‘instrument’ to predict if respondents ‘received’ the treatment. In the second stage, we estimate the effect of receiving the treatment on our dependent variable. This estimand can then be interpreted as the effect of the treatment on those who received the treatment. In our analysis, we restrict our statistical models to a single treatment and the control group. Figure 4 presents the final estimates for the adversaries and allies treatment. As we can observe, the substantive effect increases (though so does statistical uncertainty) to an effect size of around 40% of a standard deviation. Further, we see that the CACEs do not diverge when taking into account the different recall rates between the allies and enemy treatment.

Figure 4: CACE of Rival and Ally Treatments on Support for LAWs: This figure shows the complier average causal effect (CACE) of the Rival and Ally treatments on the support for AI-weapons index. Each estimate is the product of a separate model that includes only one treatment arm and the control arm. The points indicate the coefficients and bars indicate the 95% confidence intervals.

Taken together, these results suggest that support for AI-Weapons is directly responsive to concerns about competition with other states.[^8] This provides some support for a broader security dilemma or arms race type mental calculus rather than a desire to ‘be the best.’ It corresponds with other research that suggests threats play an important role not just in the judgment of leaders but in the public’s willingness to commit resources to security (DiGiuseppe, Aspide, and Becker, 2024).

Notably, however, the effects of allies or rivals is statistically indistinguishable from each other. There are several interpretations of this finding. First, respondents may not see allied states as allies for the foreseeable future. As such, the threat motivation still applies. Second, despite the null results of the ‘global leader’ condition, respondents are not motivated by fear of rivals, but of general concern that they are indeed at a disadvantage absent these weapons. To gain further insight on these competing interpretations, we examine if the treatment effect of the ‘allies’ treatment is conditional on the views of allies. If those who view allies as a competitor are moved by the treatment, it would provide support for our first speculative explanation. In Figure 5, we plot the marginal effect of the ‘allies’ treatment (against the control condition) across respondents’ views of the EU as a competitor to US power. We see that the positive relationship is strongest among those that think the EU poses no threat at all. As such, there is little evidence that the treatment is due to viewing allies as threats and suggest that respondents are more concerned with falling behind but not necessarily about supremacy. Future research should follow up on this dynamic. Do citizens fear allies with an advantage? Or do they just have a strong preference to maintain dominance, but only in the face of competition?

Figure 5: This plot shows the marginal effect of the Allies Treatment across respondents belief that Europe is a Challenger to the United States. The figure shows the mean prediction and 95% confidence intervals generated from a Kernel estimation as described in (Hainmueller, Mummolo, and Xu, 2019). The barplot plots the number of observations in each condition by control (grey) and treatment (red).

Lastly, we turn our attention to the cooperative outcomes (3 and 4 as well as an additive index of the two). Do threats also increase support for international cooperation to restrict or safely develop LAWs? The answer from our study is no. While according to Figure 2 support for international cooperation to regulate AI-weapons is generally high, our treatments do not move support for these outcomes. Further, adjusting for inattentiveness, as we did above, does not change our inferences.

If ceiling effects are not to blame for the lack of movement, it suggests that threats themselves may increase support for unilateral action on developing weapons, but not a call for more cooperation to address the threat of AI-weapons. Lastly, we see that the effect of the treatment on support for cooperation are similarly not heterogeneous. Across a variety of dimensions and indicators, the treatment effects remain zero.

Figure 6: Treatment Effects on Support for Cooperating on LAWs: The figure shows the coefficients and 95% confidence intervals for each of treatments for three different outcome variables: Support for a treaty to ban AI-weapons, support for international cooperation to develop safe AI-weapons systems.

### Conclusion

Our study builds on the growing body of research concerning public attitudes toward the development of lethal autonomous weapons. Understanding public opinion is crucial, not only because it reflects societal values, but because it can actively shape elite decision-making and foreign policy orientations (Tomz, Weeks, and Yarhi-Milo, 2020; Lin-Greenberg, 2021; Chu and Recchia, 2022; Peez and Bethke, 2024).

Consistent with prior work (Horowitz, 2016), we find that public aversion to lethal autonomous weapons remains strong, suggesting that the moral taboo surrounding these systems has not significantly eroded over the past decade. Our observational evidence further indicates that general support for artificial intelligence in government is positively associated with support for AI-integrated weapons systems. This relationship implies that as AI becomes more normalized in civilian contexts, resistance to its deployment in military applications may gradually decline. In line with broader trends in public opinion on military issues, we also find that general hawkishness and militarism are significant predictors of support for LAWs.

However, our findings also underscore that this normative stance is neither fixed nor immune to contextual influences. The international environment plays a crucial role. Our experimental results show that when respondents are reminded that other states are already developing LAWs, public support for these weapons increases in the U.S. – strikingly, regardless of whether those states are considered rivals or allies. This suggests that the public is responsive to competitive framing, even when it involves traditional allies such as the European Union. Moreover, this sense of competition does not appear to promote interest in collaborative development. Rather, it strengthens support for unilateral advancement of autonomous weapons.

There are several possible reasons for why we do not observe a significant difference in public reaction to rival and allied states. First, elite cues – shaped by shifting geopolitical dynamics – can alter how the public perceives alliances (Alley, 2023). Yet, while the second Trump administration adopted a more skeptical stance toward European allies, our experiment was still conducted during the Biden administration, when relations with the EU were still largely positive. Indeed, public support for NATO remained strong throughout this period, and a majority of U.S. citizens expressed favorable views of the alliance (Gallup, 2024). More plausibly, the perception of allies and rivals is neither uniform nor permanent. As Niou and Zeigler (2019) put it, “in any alliance, uncertainty reigns supreme; there exists at least a minimal expectation that allied groups may fight one another in addition to their common foe. Members within them appreciate this contingency and respond accordingly.” Alliances frequently involve internal tensions and strategic competition, which may diminish the expected distinction between allies and adversaries and influence how the public interprets salient issues, particularly if it comes to critical domains such as foreign military armament and military competition. Future research should dig much deeper into disentangling these conjectures.

Our findings showcase that the international environment can significantly influence, even reshape, the trade-off between normative opposition and military utility that lies at the heart of weapons development. This underscores the volatility of moral taboos and shows that public resistance to new technologies such as lethal autonomous weapons is not rigid, but contingent upon external cues and contextual settings. As such, our results have important implications not only for policymakers and stakeholders involved in the AI-weapons debate, but also for broader considerations of democratic responsiveness in security policy. The public emerges as an important political actor, with preferences that are both malleable and sensitive to perceptions of external threats and international competition.

Our insights likely extend beyond the U.S. context. As illustrated by shifts in European public opinion, e.g., following the Russian invasion of Ukraine, increasing support for defense initiatives and hawkish attitudes toward advanced weapons systems are not exclusively a American phenomena (Dill, Sagan, and Valentino, 2022; Onderco, Smetana, and Etienne, 2023). Public hawkishness, particularly when driven by fears of strategic vulnerability, appears to be a feature of democratic societies under pressure. This reinforces the need to account for public opinion as a dynamic force in shaping national and international security policies. Taken together, our study contributes to both the empirical understanding of public attitudes toward LAWs and the general literature on public opinion in international security. As the technological and geopolitical landscape continues to evolve, future research should further explore how moral thresholds shift, how elite-public feedback loops operate, and how these dynamics shape the trajectory of new technology development across democratic societies.

### Appendix

#### Attention Check

We evaluate respondents’ attention to the question wording before we present the treatments using the following survey item:

People are very busy these days and many do not have time to look up specific pieces of information. There are many websites offering the same content at different levels of detail. Some have the time to search for information all day, but some do not even have the time to read questions carefully. To show that you’ve read this much, please ignore the question below and just click the answer that includes seven.

About how many web sites do you visit daily to look up information on current issues? [0; 1-2; 3-5; 6-9; 10 and more]

#### Pre-treatment Questions

We measure the following pre-treatment variables (in that order):

1. Importance of US hegemony is measured by a survey question that asks respondents: How important do you think it is that the US stays the most powerful countries in the world? [5-point scale, ranging form ‘not at all important’ to ‘extremely important’] 2. Military assertiveness is measured as an index comprised of two survey items (randomized order): How much do you agree or disagree with the following statements? (1) The best way to ensure peace is through American military strength. (2) The use of military force only makes problems worse. [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

3. Importance of international cooperation is measured as an index consisting of three survey items (randomized order): How important are the following foreign policy goals for you personally? (1) Sharing the costs of maintaining world order with other countries. (2) Preventing the spread of weapons of mass destruction (WMDs). (3) Improving relationships with our allies. [5-point scale, ranging from ‘strongly disagree’ to ‘strongly agree’]

4. Familiarity with the use of AI is measured by a survey questions that asks respondents: How often do you use artificial intelligence tools like Chat GPT? [5-point scale, ranging from ‘have never used it’ to ‘daily’]

5. General willingness to use AI is measured as an index consisting of the following four items (randomized order): How do you feel about using AI more frequently in the following areas? (1) Identification of who is at risk of various diseases. (2) Deployment of autonomous lethal weapons in the battlefield. (3) Job selection and promotion of local officials. (4) Surveillance of national security threats. [4-point scale, ranging from ‘very unsupportive’ to ‘very supportive’]

Additional pre-treatment variables include socio-demographics (age, gender, education, income, military service, state of residency) and political variables (partisanship and perception of countries challenging US power).

#### Treatment Manipulation Check

At the very end of the survey, we ask respondents a simple recall question that allows us to assess whether respondents remember receiving the treatment. This survey item also allows us to estimate a LATE/CACE effect (see below) to see if there is an effect among those who actually received the treatment:

Earlier in the survey, we also mentioned that some countries have already started to develop lethal autonomous weapons. Can you remember which countries we mentioned? [China and Russia; European allies; Japan and South Korea; Brazil and South Africa; We didn’t mention any country]

#### Exploratory Conditional Hypotheses

We also assume that the treatment effect will be stronger among certain segments of the population.

- The treatment effects will be larger among respondents who are more likely to support the use of military force compared than among those who are less likely to support the use of military force.

- The treatment effects will be larger among respondents who place less importance on international cooperation.

- The treatment effects will be larger among respondents who place more importance on US hegemony in the international system.

- The treatment effects will be larger among respondents who are more familiar with artificial intelligence.

- The treatment effects will be larger among respondents who are more supportive of the use of artificial intelligence in general.

#### Statistical Models

To test our hypotheses, we estimate a linear model that estimates the ATE for each of our treatments US hegemony (USHeg ), competition from allies (Comp), and competition from rivals (Rival ) relative to the control condition. From this equation, we can back out comparisons among the treatment conditions.

Y = β0 + β1U SHeg + β2Comp + β3Rival + λD + ϵ We also include a matrix of pre-treatment covariates and their coefficients on the right hand side λD. We select these variables agnostically with a LASSO selection model, following the recommendation of Bloniarz et al. (2016). We include each continuous variable individually and each categorical variable as dummies in a model predicting the outcome. The LASSO model returns only variables with non-zero coefficients. We then include these predicted variables in our model estimating treatment effects.

We also examine the complier average causal effect (CACE) for each treatment. In this case, we estimate a two-stage least squares model that includes an individual treatment category and the control group while excluding non-relevant treatment categories. We first estimate the effect of the assigned treatment on the likelihood of recalling (receiving) the treatment. We then use this to estimate the impact of receiving the treatment on support for autonomous weapons in the second stage of the model. Essentially, the assignment of the treatment is used as an instrumental variable to predict receiving (recalling) the treatment.

Precieved = π0 + πT reat + λD + v Y = β0 + β1Pˆrecieved + λD + ϵ

In each of our models, we estimate the equations on two dependent variables: The two-question additive index of support for unilateral autonomous weapons and the two-question additive index of support for cooperative measures to address autonomous weapons.

#### Consent Form

We start the survey by asking whether respondents agree to take part in our study. We screen out respondents who do not agree after the following initial text:

Thank you for agreeing to take part in this research study. The data we collect will be used in academic research to help us understand your perspectives on government policies. If you agree to participate in this study, you will be asked to complete an on-line survey that will take about 7 minutes.

There are no foreseeable risks associated with this project. However, your participation in this study is completely voluntary and you are free to withdraw at any time. Your survey responses will be strictly confidential and data from this research will be reported only in anonymized form. The data will be stored on a secure server and will be opened only by the researchers when conducting analysis on aggregate data. None of your personal information will be collected. We will preserve your data in perpetuity and protect any confidential data. Anonymized data will be shared with others upon publication of any academic papers resulting from the project. We will use the data to conduct statistical analysis from which we will draw general conclusions. The project will be published in open access format so that individuals that are interested can see the final project.

By clicking ‘I agree’ below you are indicating that you are at least 18 years old, have read and understood this consent form, and agree to participate in the research study. [I agree; I do not agree]

#### Ethics

We received ethical approval from the German Association for Experimental Economic Research e.V. (No. 39cUYan7) on August 23, 2024.

#### Sampling

Our analysis relies on a sample drawn from a pool of of adult respondents on the Prolific survey platform. Prolific offers a sample that reflects the US population on the dimensions of age, gender, and political affiliation. We use this quota sampling approach to build our sample. We will screen out respondents under the age of 18. Based on our power analysis, we will recruit 2,500 respondents.

#### Power Analysis

We conducted simulations using Declare Design (Blair, Coppock, and Humphreys, 2023) to inform our sample size. Given the constrained budget to collect 2,500 responses, we plot the minimal detectable effect assuming a correlation of R=0.2 between the set of covariates and the outcome variables. Our analysis relies on 1,000 simulation of each assumed effect. Our analysis reveals that, for each treatment, we can retrieve an effect of d=0.125 with 80% power as shown in Figure 7. In our opinion, this is sufficient to capture a meaningful effect.

Figure 7: Our estimates of the minimum detectable effect are conduced with Declare Design Package in R. We assume a sample of N=2,500 and a set of covariates that correlate with the outcome at 0.20. Our estimates of the power rely on 1000 simulations of each value of the average treatment effect (ATE).

### References

Alley, Joshua. 2023. “Elite Cues and Public Attitudes Towards Military Alliances.” Journal of Conflict Resolution 67 (7-8): 1537–1563.

Becker, Jordan, Matthew DiGiuseppe, Haemin Jee, and TongFi Kim. 2024. “External Threats and Support for International Security Cooperation.” SSRN 4949045 .

Bergengruen, Vera. 2024. “How Tech Giants Turned Ukraine Into an AI War Lab.” TIME February 8: https://time.com/6691662/ai–ukraine–war–palantir/.

Blair, Graeme, Alexander Coppock, and Macartan Humphreys. 2023. Research design in the social sciences: declaration, diagnosis, and redesign. Princeton University Press.

Bloniarz, Adam, Hanzhong Liu, Cun-Hui Zhang, Jasjeet S Sekhon, and Bin Yu. 2016. “Lasso adjustments of treatment effect estimates in randomized experiments.” Proceedings of the National Academy of Sciences 113 (27): 7383–7390.

Bloom, Pazit Ben-Nun, Shaul Kimhi, Shani Fachter, Michal Shamai, and Daphna Canetti. 2020. “Coping with Moral Threat: Moral Judgment amid War on Terror.” Journal of Conflict Resolution 64 (2-3): 231–260.

Candrian, Cindy, and Anne Scherer. 2022. “Rise of the Machines: Delegating Decisions to Autonomous AI.” Computers in Human Behavior 134: 1–16.

Chu, Jonathan A., and Stefano Recchia. 2022. “Does Public Opinion Affect the Preferences of Foreign Policy Leaders? Experimental Evidence from the UK Parliament.” The Journal of Politics 84 (3): 1874–1877.

Conrad, Courtenay R., Sarah E. Croco, Brad T. Gomez, and Will H. Moore. 2018. “Threat Perception and American Support for Torture.” Political Behavior 40 (4): pp. 989–1009.

Coppock, Alexander. 2023. Persuasion in parallel: How information changes minds about politics. University of Chicago Press.

Darden, Keith, and Harris Mylonas. 2016. “Threats to Territorial Integrity, National Mass Schooling, and Linguistic Commonality.” Comparative Political Studies 49 (11): 1446– 1479.

DiGiuseppe, Matthew, Alessia Aspide, and Jordan Becker. 2024. “Threats and the public constraint on military spending.” British Journal of Political Science 54 (3): 649–666.

Dill, Janina, Scott D. Sagan, and Benjamin A. Valentino. 2022. “Kettles of Hawks: Public Opinion on the Nuclear Taboo and Noncombatant Immunity in the United States, United Kingdom, France, and Israel.” Security Studies 31 (1): 1–31.

Frenkel, Sheera, and Natan Odenheimer. 2025. “How Israel Sent A.I. Into Combat.” New York Times April 30 (International Edition): 4.

Gadarian, Shana Kushner. 2010. “The Politics of Threat: How Terrorism News Shapes Foreign Policy Attitudes.” The Journal of Politics 72 (2): 469–483.

Gallup. 2024.  “Americans Remain Committed to NATO, Critical of UN.” https://news.gallup.com/poll/611261/americans-remain-committed-nato-critical.aspx(Accessed April, 14 2024).

Gartzke, Erik, and Matthew Kroenig. 2017. “Social Scientific Analysis of Nuclear Weapons: Past Scholarly Successes, Contemporary Challenges, and Future Research Opportunities.” Journal of Conflict Resolution 61 (9): 1853–1874.

Hainmueller, Jens, Jonathan Mummolo, and Yiqing Xu. 2019. “How much should we trust estimates from multiplicative interaction models? Simple tools to improve empirical practice.” Political Analysis 27 (2): 163–192.

Horowitz, Michael. 2016. “Public Opinion and the Politics of the Killer Robots Debate.” Research and Politics 3 (1): 1–8.

Horowitz, Michael. 2020. “Do Emerging Military Technologies Matter for International Politics?” Annual Review of Political Science 23: 385–400.

```text
Humble,  Kristian. 2024.  “War,  Artificial Intelligence,  and the Future
of  Conflict.”  Georgetown  Journal  of  International  Affairs  (July  12):
```

https://gjia.georgetown.edu/2024/07/12/war–artificial–intelligence–and–the–future–of–conflict/.

Ipsos. 2021. “Global Survey Highlights Continued Opposition to Fully Autonomous Weapons.” (February). Accessed: 2025-01-21.

Jervis, Robert. 1978. “Cooperation under the security dilemma.” World Politics 30 (2): 167–214.

Jervis, Robert. 2017. Perception and misperception in international politics: New edition. Princeton University Press.

Jo, Hyeon. 2023. “Understanding AI tool engagement: A study of ChatGPT usage and wordof-mouth among university students and office workers.” Telematics and Informatics 85: 102067.

Kaltenthaler, Karl C., Daniel M. Silverman, and Munqith M. Dagher. 2020. “Nationalism, Threat, and Support for External Intervention: Evidence from Iraq.” Security Studies 29 (3): 549–573.

Kertzer, Joshua D. 2022. “Re-assessing elite-public gaps in political behavior.” American Journal of Political Science 66 (3): 539–553.

Kertzer, Joshua, Ryan Brutger, and Kai Quek. 2024. “Perspective Taking and Security Dilemma Thinking: Experimental Evidence from China and the United States.” World Politics 76 (2): 334–378.

Koch, Lisa Langdon, and Matthew Wells. 2021. “Still Taboo? Citizens’ Attitudes toward the Use of Nuclear Weapons.” Journal of Global Security Studies 6 (3).

Kreps, Sarah, Julie George, Paul Lushenko, and Adi Rao. 2023. “Exploring the artificial intelligence “Trust paradox”: Evidence from a survey experiment in the United States.” Plos one 18 (7): e0288109.

Landau-Wells, Marika. 2024. “Building from the brain: Advancing the study of threat perception in International Relations.” International Organization 78 (4): 627–667.

Lin-Greenberg, Erik. 2021. “Soldiers, Pollsters, and International Crises: Public Opinion and the Military’s Advice on the Use of Force.” Foreign Policy Analysis 17 (3): orab009.

Maas, Matthijs. 2019. “How viable is international arms control for military artificial intelligence? Three lessons from nuclear weapons.” Contemporary Security Policy 40 (3): 285–311.

Mahmud, Hasan, A.K.M. Najmul Islam, Syed Ishtiaque Ahmed, and Kari Smolander. 2022. “What influences algorithmic decision-making? A systematic literature review on algorithm aversion.” Technological Forecasting and Social Change 175: 121390.

Malhotra, Neil, and Elizabeth Popp. 2012. “Bridging Partisan Divisions over Antiterrorism Policies: The Role of Threat Perceptions.” Political Research Quarterly 65 (1): 34–47.

```text
McKernan,  Bethan,  and Harry Davies. 2024.  “‘The machine did it
```

coldly’:  Israel used AI to identify 37,000 Hamas targets.” The Guardian https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes.

Myrick, Rachel. 2021. “Do External Threats Unite or Divide? Security Crises, Rivalries, and Polarization in American Foreign Policy.” International Organization 75 (4): 921–958.

Niou, Emerson M. S., and Sean M. Zeigler. 2019. “External Threat, Internal Rivalry, and Alliance Formation.” The Journal of Politics 81 (2): 571–584.

Onderco, Michal, Michal Smetana, and Tom W. Etienne. 2023. “Hawks in the making? European public views on nuclear weapons post-Ukraine.” Global Policy 14 (2): 305–317.

Peez, Anton, and Felix S Bethke. 2024. “Does Public Opinion on Foreign Policy Affect Elite Preferences? Evidence from the 2022 US Sanctions against Russia.” International Studies Quarterly 69 (1): sqae145.

Press, Daryl G., Scott D. Sagan, and Benjamin A. Valentino. 2013. “Atomic Aversion: Experimental Evidence on Taboos, Traditions, and the Non-Use of Nuclear Weapons.” American Political Science Review 107 (1): 188–206.

Price, Richard. 1995. “The Genealogy of the Chemical Weapons Taboo.” International Organization 49 (1): 73–103.

Rosendorf, Ondrej, Michal Smetana, and Marek Vranka. 2022. “Autonomous weapons and ethical judgments: Experimental evidence on attitudes toward the military use of “killer robots”.” Peace and Conflict: Journal of Peace Psychology 28 (2): 177.

Rosendorf, Ondrej, Michal Smetana, and Marek Vranka. 2024. “Algorithmic Aversion? Experimental Evidence on the Elasticity of Public Attitudes to “Killer Robots”.” Security Studies 33 (1): 115–145.

Rosert, Elvira, and Frank Sauer. 2021. “How (not) to stop the killer robots: A comparative analysis of humanitarian disarmament campaign strategies.” Contemporary Security Policy 42 (1): 4–29.

Sagan, Scott D, and Benjamin A Valentino. 2025. “Atomic Arguments and Counter- Arguments: How Exposure to Conflicting Information Influences American Public Support for the Use of Nuclear Weapons.” International Studies Quarterly 69 (1): sqae154.

Sauer, Frank. 2016. “Stopping ‘killer robots’: Why now is the time to ban autonomous weapons systems.” Arms Control Today 46 (8): 8–13.

Schwartz, Joshua A., and Dominic Tierney. 2025. “Us and Them: Foreign Threat and Domestic Polarization.” Journal of Conflict Resolution 69 (2-3): 352–380.

Tao, Y, J Dong, H Niu, Y Lv, X He, S Zhang, and X Liu. 2023. “Fear facilitates utilitarian moral judgments: Evidence from a moral judgment task.” Psych Journal 12 (5): 680–689.

Tomz, Michael, Jessica L.P. Weeks, and Keren Yarhi-Milo. 2020. “Public Opinion and Decisions About Military Force in Democracies.” International Organization 74 (1): 119– 143.

von Garrel, J¨org, and Jana Mayer. 2023. “Artificial Intelligence in studies—use of Chat- GPT and AI-based tools among students in Germany.” Humanities and Social Sciences Communications 10: 799.

Walt, Stephen. 1985. “Alliance Formation and the Balance of World Power.” International Security 9 (4): 3–43.

Young, Kevin L, and Charli Carpenter. 2018. “Does science fiction affect political fact? Yes and no: A survey experiment on “Killer Robots”.” International Studies Quarterly 62 (3): 562–576.

Zhang, Baobao, and Allan Dafoe. 2020. US public opinion on the governance of artificial intelligence. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. pp. 187–193.

Zwald, Zachary, Ryan Kennedy, and Adam Ozer. 2024. “The Political Viability of AI on the Battlefield: Examining US Public Support, Trust, and Blame Dynamics.” Journal of Peace Research .

### Notes

[^1]: The Ipsos survey asked the following question: “The United Nations is reviewing the strategic, legal and moral implications of lethal autonomous weapons systems. These weapons systems would be capable of independently selecting targets and attacking those targets without human intervention. They are thus different than current day ‘drones’ where humans select and attack targets. How do you feel about the use of such lethal autonomous weapons systems in war?”
[^2]: All hypotheses have been pre-registered. Our pre-analysis plan also includes more detailed hypotheses to allow for a ranking of individual treatments relative to each other. This experiment was conducted as part of a broader survey that included additional experimental work on AI governance in monetary policy-making.
[^3]: The questions asked for agreement with the following statements: (1) the best way to ensure peace is through American military strength and (2) the use of military force only makes problems worse.
[^4]: The question read: How important do you think it is that the US stays the most powerful country in the world?
[^5]: We asked how important each of the following foreign policy goals was for the respondent: (1) sharing the costs of maintaining the world order, (2) preventing the spread of weapons of mass destruction, and (3) improving relationships with our allies.
[^6]: We asked: Artificial Intelligence (AI) refers to computer systems that perform tasks or make decisions that usually require human intelligence. AI can perform these tasks or make these decisions without explicit human instructions. How do you feel about using AI more frequently in the following areas? (1) Identification of who is at risk of various diseases and (2) job selection and promotion of local officials. Again, we create an additive index of these items.
[^7]: We can likely attribute this finding to small sample bias. Only 4% view Canada as a challenger.
[^8]: In addition to examining the ATE, we also attempted to find conditional effects (CATE) in line with our pre-registered exploratory hypotheses. We find, consistent with Coppock (2023), that respondents moved in parallel across a number of attributes. We find no significant interaction effects across militarism, partisanship (7-point), views on China, views on Russia, desire for international cooperation, the use of AI, and support of AI in general. The treatment effect, is surprisingly consistent across different segments of society.