1. Introduction
Since the development and use of atomic bombs at the end of World War II, the stakes of war have irreversibly changed. Even in the early stages of nuclear technology, humankind demonstrated that when forced to choose between the lives of an in-group and an out-group—especially under the influence of strong emotions and a perceived need for justice—nuclear weapons could be seen as a justified option (Walker, Reference Walker2005). Today, as tensions escalate between sovereign nations and ideologically diverse groups (Mecklin, Reference Mecklin2025), decisions about the use of nuclear weapons with vast expected military and civilian human and environmental casualties may again become a reality. These highly consequential decisions hinge on the peculiarities of human judgment, shaped by cognitive and emotional factors that can serve as both aids or obstacles—such as compassion and bounded rationality (Frith and Singer, Reference Frith and Singer2008; Jones, Reference Jones1999). How do people reason when making such decisions?
In the present research, American respondents were presented with a scenario in which they had to make a difficult decision in the context of an unavoidable war: choosing between the casualties of in-group American troops or out-group Iranian civilians. The material was adapted from the original material by Sagan and Valentino (Reference Sagan and Valentino2017) as well as later adaptations by Slovic et al. (Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020). Participants could opt for a difficult but winnable ground war, deploy nuclear weapons to kill 100,000 Iranian civilians and end the war quickly, or deploy a nuclear strike that would kill 2 million civilians. After making their decision, participants explained their thinking in their own words.
To analyze these responses, we employed natural language processing, specifically structural topic modeling, to identify common topics. These topics were then examined in relation to each participant’s preference regarding nuclear weapon use. The findings revealed significant complexity in participants’ thinking—in particular, 10 overarching patterns in reasoning around the use of nuclear weapons were derived from the analysis. Notably, the type of thinking employed was related to the selected option. Reasoning that using nuclear weapons invites retaliation with nuclear weapons—were related to a preference for conventional warfare over the use of nuclear weapons. Conversely, participants who reasoned that deployment of nuclear weapons is a type of retribution had a large proportion of participants who preferred the nuclear strike that would lead to 2 million civilian deaths. Our analysis provides valuable insights into the psychological mechanisms of public support for using nuclear weapons against out-group civilians.
2. Attitudes to nuclear weapons in America
It is ‘89 seconds to midnight’—a chilling reminder of the potential destruction of the world. As the Doomsday Clock ticks ‘closer than ever’ to catastrophe, nuclear risk remains a primary factor in setting its time (Bulletin of the Atomic Scientists, 2025). Given the current global tensions, understanding attitudes toward nuclear weapons and the thinking underlying nuclear decisions is crucial. After 1945, the American public’s attitude toward nuclear weapons shifted from widespread acceptance to growing skepticism (Tannenwald, Reference Tannenwald2007). A recent survey found that 69% of Americans believe the development of nuclear weapons has made the world less safe (Pew Research Center, 2025). Further, only 35% consider the bombings of Hiroshima and Nagasaki to have been justified—a substantial decline from the 85% who supported the bombings in 1945 (Moore, Reference Moore2005). Classical deterrence theory, as articulated by Schelling (Reference Schelling1960), assumes rational actors who weigh costs and benefits to prevent conflict—a logic of consequences. In contrast, Tannenwald (Reference Tannenwald1999) proposed the idea of a nuclear taboo—a normative belief that nuclear weapons should not be used—grounded in the moral logic of appropriateness rather than purely strategic calculation.
Although public skepticism about the US use of nuclear weapons in Hiroshima and Nagasaki grew after 1945 (Tannenwald, Reference Tannenwald2007), it is not a given that the nuclear taboo would apply in future scenarios. Instead of reduced acceptance of nuclear weapons, recent studies have found a high acceptance of their use in specific contexts (Horschig, Reference Horschig2022; Rohlfing, Reference Rohlfing2023; Sagan and Valentino, Reference Sagan and Valentino2017; Slovic et al., Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020). A study by Sagan and Valentino (Reference Sagan and Valentino2017) found that about 59% of Americans would support using nuclear weapons if framed as a quick way to end the war that would save the US lives. This suggests that, in certain contexts, the American public might deem nuclear weapons justified.
So, are we naïve in our belief in the power of the nuclear taboo? Annie Jacobsen (Reference Jacobsen2024) raises this concern in her book Nuclear War: A Scenario, where she points out that even the US military experts avoid detailed planning for nuclear weapon scenarios. Jacobsen (Reference Jacobsen2024) presents a hypothetical situation in which North Korea launches intercontinental ballistic missiles targeting the United States. When this scenario is presented to military experts, interviews reveal that the US President together with military advisors might have as little as 6 min to make critical decisions, including whether to use nuclear weapons. This clearly illustrates the immense pressure and the role of human judgment in high-stakes decisions. It may be argued that if these scenarios were to become reality, important systematic cognitive and emotional strengths and limitations of the human mind would come into play.
3. Decision making and its limitations
The challenges involved in high-stakes situations, such as an impending nuclear war, pose significant difficulties for any decision maker. The impact of a nuclear strike is difficult to grasp psychologically, due to lack of experience (fortunately), its immense complexity, and cognitive biases. One such bias is ‘scope insensitivity’—the tendency to exhibit less than linear, or even inverse, sensitivity to the magnitude of an issue (Dickert et al., Reference Dickert, Västfjäll, Kleber and Slovic2015). This could manifest in how we perceive the direct and long-term consequences of a nuclear strike. A related concept, ‘psychic numbing’, refers to the diminished response to large-scale catastrophes, which can provoke behavioral impulses to avoid confronting the situation, along with feelings of detachment (Slovic et al., Reference Slovic, Finucane, Peters and MacGregor2007). Psychic numbing can be seen as an adaptive function, which helps the individual withstand the existential threats looming large for oneself, instead handled by emotional detachment (Lifton, Reference Lifton1984). But at the same time, this is very problematic as a guide for decision making as it does not accurately represent the magnitude of consequences of nuclear decisions.
Beyond the inherent limitations in how scope is managed in decision making, basic constraints in attention play a crucial role. Our tendency to rely on what is most prominent or available in our minds—what is top of mind—could have significant implications for high-stakes decisions (Tversky et al., Reference Tversky, Sattath and Slovic1988; Tversky and Kahneman, Reference Tversky and Kahneman1973; Slovic, Reference Slovic2015). Our focus is restricted by working memory, which is curated by selective attention, a process that guides our perception (Ede and Nobre, Reference Ede and Nobre2023). A potential consequence of the process of selective attention is ‘security prominence’—the tendency for personal or national (in-group) safety to dominate our thoughts and overshadow humanitarian concerns (Delaney and Slovic, Reference Delaney and Slovic2019). Emotion is one attribute that increases the prominence of an attribute in a decision, clearly illustrated through the concept of the affect heuristic (Slovic et al., Reference Slovic, Finucane, Peters and MacGregor2007).
In the context of nuclear decisions, decision makers may struggle to fully grasp the widespread consequences of their choices (e.g., scope insensitivity). Further, the situation poses a great risk for a decrease in empathy for out-group victims as they focus their attention on the safety of their own (e.g., psychic numbing and safety prominence). These cognitive and emotional limitations are likely to play a significant role in decisions regarding the use of nuclear weapons. Some have suggested that public attitudes toward nuclear weapon use might influence policymakers’ decisions in high-stakes situations (Sagan and Valentino, Reference Sagan and Valentino2017). At the same time, policymakers themselves are not immune to cognitive biases; professionals across fields remain susceptible, despite their expertise (Berthet, Reference Berthet2022). The overwhelming nature of such a decision, coupled with the numbing effects of large-scale suffering, could push individuals to prioritize in-group safety over the value of out-group lives, even in the face of mass civilian casualties.
Surveys of the public in the United States, using the same scenario as in the current research, have found that demographic factors such as political affiliation, age, education, and gender influenced the probability of choosing nuclear weapons over conventional warfare (Sagan and Valentino, Reference Sagan and Valentino2017; Slovic et al., Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020). Specifically, Republicans, older individuals, and those who had less formal education have been found more likely to support the use of nuclear weapons. Regarding gender effects, Sagan and Valentino found that women were more willing than men to support nuclear weapons use, with the same Iran War scenario, Slovic et al. (Reference Slovic, Peterson, McDermott, Post and Västfjäll2025) found that women were consistently more supportive of the nuclear strikes when the number of US troop lives that would be saved by nuclear bombing was specified.
Going beyond demographics, Slovic et al. (Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020, Reference Slovic, Peterson, McDermott, Post and Västfjäll2025) found that respondents with a more positive attitude toward capital punishment, antiabortion, gun ownership, and stopping immigration, exhibited a greater likelihood of choosing nuclear weapons. Furthermore, those more inclined to choose nuclear weapons often viewed the civilian casualties as more distant and dehumanized. Dehumanization is closely related to the concept of virtuous violence (Fiske et al., Reference Fiske, Rai and Pinker2014)—the belief that certain acts of violence, such as nuclear strikes, may be seen as morally justified or even necessary for the greater good. Both Sagan and Valentino (Reference Sagan and Valentino2017) and Slovic et al. (Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020) found that many of their participants accepted nuclear weapons as ethically justifiable.
4. The current research
The decision to support or oppose the use of nuclear weapons in high-stakes situations is shaped by a combination of contextual factors, psychological processes, and individual characteristics. But what goes on in the mind when making such a decision? The current research builds on scenarios from Sagan and Valentino (Reference Sagan and Valentino2017) and Slovic et al. (Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020) to explore how US troop and civilian casualties influence the decision to use nuclear weapons in a hypothetical conflict with Iran. While the previous research focuses on the public’s general acceptance of nuclear weapon use—an inclination that is relatively high in this context (≈59%–33%)—this study aims to dig deeper into the thinking behind the preferences identified in Slovic et al. (Reference Slovic, Peterson, McDermott, Post and Västfjäll2025).
Participants first read a news article about a hypothetical war between the United States and Iran and then selected an initial option. Afterward, they were presented with a revised scenario that forms the core of this research: In the scenario, the United States has imposed economic sanctions on Iran for violating a nuclear agreement, which leads to an Iranian attack on the US forces, triggering a difficult ground war. While the US victory is expected, there is an estimated loss of 100 to 40,000 (depending on the condition) US troops if the war continues. In response, the United States is considering using a nuclear bomb to end the conflict and prevent any additional US casualties, but this comes at the cost of an estimated 100,000 Iranian civilians dying if the bomb is dropped on location A, or 2 million fatalities if dropped on location B.
To understand the thinking behind participants’ preferences—whether to continue the ground war, use a nuclear strike causing 100,000 civilian casualties, or use a nuclear strike causing 2 million civilian casualties—they were asked to ‘write a brief paragraph describing how you arrived at your preferred option’. These open-text responses were analyzed using Natural Language Processing (NLP) techniques, specifically Structural Topic Modeling (STM [Roberts et al., Reference Roberts, Stewart and Tingley2019]), to identify common themes in participants’ thinking about whether the use of nuclear weapons is justifiable.
Modern NLP—a machine learning method—has a rich history within psychology to access and analyze verbal data (Pennebaker et al., Reference Pennebaker, Mehl and Niederhoffer2003). Building on earlier methods like introspection, free association, and think-aloud protocols (Solomon, Reference Solomon1995), contemporary NLP tools like STM can reveal valuable insights into thoughts, emotions, values, beliefs, and decision-making processes (Mihalcea et al., Reference Mihalcea, Biester, Boyd, Jin, Perez-Rosas, Wilson and Pennebaker2024).
In this study, we use STM to explore each individual’s thinking patterns related to nuclear weapons in the given scenario. We also examine how the identified topics related to participant’s preferences to use or not to use nuclear weapons, as well as their demographic characteristics and individual traits. Our findings highlight the complexity of decision making in these high-stakes scenarios, showing that the thinking behind participants’ preferences is tied to both their individual characteristics and the scenario’s context.
5. Method
5.1. Sample
The research was approved by the Institutional Review Board of the Oregon Research Institute and was exempt from full review (IRB 00000278), and all participants provided informed consent. The sample size was based on results from pilot studies combined with available funding. There was no pre-registration, the study was exploratory. Data collection took place between September 24, 2024 and October 20, 2024. At the time of data collection, Iran–US relations were marked by heightened tensions following Iran’s April missile and drone attack on Israel and the July killing of Hamas leader Ismail Haniyeh in Tehran—an act attributed to Israel (Congressional Research Service, 2024). The study was conducted on the Prolific online survey platform, with an initial sample of 2937 participants, who were randomly assigned to 14 conditions. For the scope of the current research, 2 conditions (6 A/B; N = 484) were excluded, as their primary purpose was to evaluate the value function proposed by prospect theory (Kahneman and Tversky, Reference Kahneman and Tversky1979).
Additionally, due to the small number of participants identifying as nonbinary or preferring not to disclose their gender (N = 30), these responses were excluded from the main analysis, which applied structural topic modeling. The final sample before preprocessing (see Analyses) consisted of 2423 participants.
All survey respondents were American adults who agreed to participate in an online survey in exchange for monetary compensation, of the final sample, 49.9% identified as female. The political affiliation of participants was 48.21% Democrat and 47.79% Republican, with 4.00% identifying as independent, other, or undeclared. Participants ranged in age from 18 to 85 years (M = 41.57, SD = 13.66). Participants were highly educated, with 67.73% having attended at least some college up to completion of a graduate degree.
Regarding racial and ethnic identification, respondents were allowed to choose more than 1 category, 68.72% identified as White or Caucasian (European–American). Other frequent ethnic identifications were 19.56% as Black or African American, 8.58% as Hispanic, Latino, or Spanish, 8.25% as Asian or Asian American, and 3.64% identified with other ethnicities, or preferred not to answer.
Choice 1 characteristic for each condition
Note. Conditions 6 A and B are not included in the current research and thus are not displayed in Table 1. Conditions 7 A and B have an unspecified number of US troop losses. Instead, the political cost of the war was highlighted: ‘However, the long duration of the war has created significant political tensions and public opposition in the United States’.
The code behind this analysis/simulation has been made publicly available at the Open Science Framework and can be accessed at Open Science Framework.Footnote 1
5.2. Procedure
5.2.1. Choice 1
All participants read a news story, outlining a hypothetical war scenario between the United States and Iran (see Supplementary 3, for the complete story). The news story included information about both the expected US military troop losses if the ground war continued, as well as the president’s consideration to instead use a nuclear strike, and if so, the estimated number of Iranian civilians that would be killed. The estimated number of US military troop losses and Iranian civilian casualties were altered depending on condition (see Table 1). The participants were randomized to 1 of the 12 conditions included in the current research in a between-subject design. After reading the news story, the participants were given a summary to reinforce the following key points:
The United States imposes economic sanctions on Iran in response to violations of a nuclear agreement. In retaliation, Iran attacks US forces, leading to a difficult ground war. While the United States is expected to win, continuing the conflict would result in an estimated × [100–40,000 or unspecified] US troop casualties. To reduce this number, the United States is considering using a nuclear bomb on Iran. However, this decision comes at the cost of civilian lives, with an estimated [100,000 or 2 million] Iranian civilians expected to die if a nuclear strike is carried out.
In all conditions, participants were faced with a first choice, asking if they would continue the ground war or select the nuclear strike: ‘Given the facts described in the article, if you had to choose between launching a nuclear strike against the Iranian city or continuing the ground war against Iran, which option would you prefer?’
The projected number of deaths of US troops in a continuation of the ground war that the United States was expected to win ranged from 100 to 40,000 across conditions 1 through 5, or unspecified in condition 7, as shown in Table 1. Each condition had 2 versions, A or B, depending on whether the expected civilian casualties from a nuclear strike were 100,000 or 2 million. Following their first choice, participants responded to additional questions. These questions will not be described since they were not included in any analysis here.
5.2.2. Choice 2
After the participants had completed Choice 1 with the associated questions, they were presented with a news story featuring a revised war scenario. As before, the story included information about the US military troop losses depending on the participant’s condition. However, unlike the first choice, where a single nuclear strike option had a predetermined outcome (100K or 2 million), the second choice included 3 options: continuing the ground war, launching a nuclear strike that would kill 100,000 civilians, or opting for a more devastating nuclear strike that would result in 2 million civilian deaths. The participants read the following information about the revised (second) scenario:
Suppose you learn that the military has suggested a revised bombing plan where 2 target locations in the city are under consideration. Their report to the president estimates that a nuclear strike on location A would kill an estimated 100,000 Iranian civilians and a strike on location B would be estimated to kill 2 million Iranian civilians.
[Same key points as the Choice 1 scenario] The United States has imposed economic sanctions on Iran for violating a nuclear agreement, which leads to an Iranian attack on US forces, triggering a difficult ground war. While US victory is expected, there is an estimated loss of x [100–40,000 or unspecified] US troops if the war continues .
[Revised key points for Choice 2 scenario] In response, the United States is considering using a nuclear bomb to shorten the conflict and reduce US casualties, but this comes at the cost of high civilian casualties, with an estimated 100,000 Iranian civilians dying if the bomb is dropped on location A, or 2 million if dropped on location B.
In all conditions, participants were faced with the choice between whether they would continue the ground war or select the nuclear strike at location A or location B: ‘In light of this new information, if you had to choose between launching a nuclear strike against 1 of the 2 possible targets or continuing the ground war against Iran, which option would you prefer?’
The main dependent variable of the current research is the participants’ thinking described in an open-text response following Choice 2. The participants were asked to: ‘Please write a brief paragraph describing how you arrived at your preferred option’. This verbal response, as well as the Choice 2 preceding it, is the focus of our analysis.
Following this verbal description of their thinking, the participants were asked additional questions. Later, the relevant variables included in the analysis are described. First, participants were asked: ‘Please rate how important the following consideration was to you when thinking about how much you supported or opposed using nuclear weapons…’, on a scale from 1 (not at all important) to 4 (very important). The consideration was: ‘The feeling that killing 100,000 Iranian civilians is somehow better in comparison to killing 2 million’. This item is referred to as the ‘better bomb’ comparison because many respondents used the word better in their explanatory texts. Second, participants indicated their political party affiliation: Democrat, Republican, Independent, Undeclared, or Other.
Participants also responded to 8 items assessing punitive–authoritarian–nationalist attitudes (PAN index).Footnote 2 Unless otherwise noted, items were rated on a 7-point Likert-type scale ranging from 1 (strongly disagree) to 7 (strongly agree). Sample items included: ‘The southern border should be closed to stop illegal immigration into the US; some people deserve to suffer’. Other items included support for the death penalty, belief in Hell (measured with a 5-point categorical scale), endorsement of abortion restrictions when a fetal heartbeat is present, the importance of protecting gun ownership rights over controlling gun access, confederate monuments, and the belief that: ‘our country desperately needs a mighty leader who will do what has to be done to destroy the radical new ways and sinfulness that are ruining us’. See Appendix B for exact wording of all items.
Responses were coded such that higher values reflected stronger endorsement of the PAN views. A PAN index was then calculated by assigning 1 point for each item endorsed. Scores ranged from 0 to 8 and were grouped into 3 categories: low (0–1 items endorsed), medium (2–4 items endorsed), and high (5–8 items endorsed), with approximately equal distribution of respondents within categories.Footnote 3
5.3. Analysis
The open-text response data were analyzed using Structural Topic Modeling (STM), a probabilistic machine learning technique designed to detect patterns in word usage and identify latent thematic structures within textual data. In contrast to large language models (LLMs), which are trained on massive corpora to generate or predict text based on learned linguistic patterns, STM does not produce language. Instead, it discovers patterns of word co-occurrence within a specific dataset to infer underlying topics. STM is based on a generative probabilistic framework that assumes each document (in this case, a participant’s response, e.g., see Table 2, column 3) is composed of a mixture of topics (e.g., see Table 2, column 1), and that each topic is defined as a probability distribution over words. The model estimates both the topical content (i.e., which words belong to which topics) and the topic proportions for each document by inferring the most likely latent structure that could have generated the observed text data. Each response is assigned a prevalence (referred to as a theta value, e.g., see Table 2, column 2 for average prevalence per topic) indicating the degree to which each topic is present in the text. This approach provides insight into the prevalence of topics across the dataset. Additionally, STM allows researchers to incorporate document-level metadata, such as demographic information or experimental conditions, enabling analysis of how topic usage systematically varies across different participant characteristics or groups.
Topics from the open-text responses following Choice 2
Note: Topic labels are based on term content and related survey responses, emphasizing shared themes. Topic (%) indicates the average prevalence of each topic, based on the mean theta value across all participants; this reflects how common the topic is overall. See Appendix A (Table A1) for additional response examples and Supplementary 2 for the top 5 FREX words per topic. The topics are presented in order from 1 to 10 below, the order of the topics is unrelated to the prevalence of the topic; the order is the one provided by structural topic model.
Prior to model estimation, the text data were preprocessed to prepare it for analysis. This involved tokenization, where responses were split into individual words based on whitespace, and lowercasing, which standardized all text by converting it to lowercase to minimize variations caused by capitalization. Stop-word removal was applied to eliminate frequently used words with little semantic value (e.g., ‘the’ or ‘and’), preventing them from influencing the analysis. Stemming was then performed to reduce words to their root forms by removing suffixes (e.g., ‘thinking’ → ‘think’). To reduce noise, only words that appeared at least twice in the dataset were retained. One participant response was excluded during preprocessing, resulting in a final sample of 2422 responses for analysis.
6. Results
The analysis was conducted using the R package STM. The textual responses varied in length ranging from 2 to 532 words, with a median of 37 words and a mean of 42.4 words per response. The total corpus contained 102,573 words. The most frequent words were ‘nuclear’ (used 1684 times), ‘war’ (used 1464 times), ‘civilian’ (used 1369 times), ‘death’ (used 1093 times), and ‘option’ (used 830 times).
To determine the optimal number of topics for the model, we conducted a systematic comparison across a range of topic numbers from 3 to 20. Although STM can be fit with as few as 2 topics, such solutions run the risk of oversimplifying the thematic structure and limiting the insights that can be drawn. Conversely, models specifying more than 20 topics typically result in highly granular themes that are difficult to interpret and apply meaningfully. This range was therefore selected pragmatically to capture sufficient thematic detail while maintaining the practical usability and interpretability of the resulting topics.
For each topic number within this range, we conducted 30 independent runs, each starting with a different random guess of the topic structure. Each run involved up to 300 iterations, during which the model gradually adjusted its estimates step by step, similar to how iterative algorithms like Expectation–Maximization refine parameters by repeatedly improving the fit to the data. Each ‘model’ refers to the algorithm for a given topic number, including the estimated topics (word distributions), topic proportions per document, and any metadata effects. We selected the optimal model based on a balance between semantic coherence (how well the words in a topic fit together) and exclusivity (how distinct topics are from one another), along with qualitative evaluation of topic interpretability (see Figure 1). Using multiple runs per model ensured the stability of the results.
Standardized mean exclusivity and semantic coherence for models with 3–20 topics.
Note: Selected model demarcated by circle. The values presented are the average across the 30 runs for each of the separate models. These averages were than standardized for easier visualization and interpretability.
The final model was fitted with 10 topics, based on 50 independent runs with 300 iterations per run. Exclusivity values were relatively stable across runs, indicating consistent distinctiveness of topics. Due to the low variation in exclusivity, we prioritized semantic coherence when selecting the final model (see Figure 2).
Scatter plot of semantic coherence and exclusivity for the 10 best runs from fitting a 10-topic structural topic model.
Note: Each small colored dot represents an individual topic’s semantic coherence and exclusivity values. Dots with the same color belong to the same model run. For each run, the average semantic coherence and exclusivity across its 10 topics is plotted in the same color and labeled with the run number. Run 8, which showed a favorable trade-off between coherence and exclusivity, was selected as the final model.
All models included prevalence covariates (i.e., metadata: gender, political affiliation, better bomb effect, PAN index, and experimental conditionsFootnote 4). In Structural Topic Modeling, a prevalence covariate is a variable used to explain differences in how much each topic appears in different documents (in this case, a participant’s response). It works like a predictor in a regression model, in which the outcome is the proportion of each topic in a document. While these prevalence covariates directly influence which topics are more prominent in documents, they can also indirectly affect what those topics contain, since STM estimates all parts of the model together. As a result, including prevalence covariates can lead to differences in both the prevalence and the content of topics.
In addition to the main analysis, we identified each participant’s most representative topic, defined as the topic with the highest prevalence (i.e., highest theta value) in their response, as estimated by the Structural Topic Model (STM). This resulted in a grouping variable with 10 possible topic categories, each corresponding to the participant’s most representative topic. We then used this grouping variable to examine whether participants’ preferences for choice 2 differed across topic groups.
Specifically, we examined how often each of the 3 options, ground war, a nuclear strike causing 100,000 civilian casualties, and a nuclear strike causing 2 million civilian casualties, occurred within each topic group, compared to their overall frequency in the full sample. We used a chi-square test of homogeneity to assess differences across groups and analyzed the standardized corrected Pearson residuals to determine whether options appeared more or less frequently within each topic group than would be predicted based on their distribution in the full sample (see Table 3).
Chi-square test of homogeneity: Standardized residuals for preferences across topic groups
Note: Standardized residuals in the table represent the difference between the observed and expected frequencies, adjusted for the size of the data. The order of the topics is from the largest to smallest proportion of GW. A residual ±1.96 indicates deviation from the expected count at the 0.05 statistical significance level (2-tailed test).
The same grouping variable and analytical approach were also used to explore demographic variables (e.g., gender, political affiliation) and individual difference measures (e.g., PAN index, perceived scenario relevance), for an overview, see the Result section ‘Demographics and individual differences’; for statistical details, see Supplementary 1.
Table 3 shows that people whose reasoning reflected Topics 4, 5, or 10 (innocents, humanitarian ethics, nuclear escalation) were overwhelmingly more likely to choose the ground war. People whose reasoning reflected Topics 6, 7, or 9 (trade-offs, American lives first, or fewer total deaths) leaned heavily toward the 100,000-casualty strike. People in Topic 3 (payback) were most likely to choose the 2 million-casualty strike.
6.1. Topic details
6.1.1. Topic 1: Balancing strategic, ethical, and reputational concerns
Topic 1 is the least prominent topic in the model. It accounts for the smallest average share of content across responses, as well as the number of participants having the topic as their most representative (see Table 2). The topic includes answers that weigh ethical considerations together with how a nuclear strike would affect the global perception of the United States, as well as its military consequences.
Among the participants (N = 20; 0.83%) whose answer was best represented by this topic, 70.00% chose to continue the ground war, 15.00% opted for a nuclear strike killing 100,000 civilians, and 15.00% chose a nuclear strike resulting in 2 million civilian casualties (see Figure 3). Further, when studying the standardized residuals, we see that choosing a nuclear strike that will kill 2 million had a higher frequency than expected in the topic group—meaning there was higher frequency of participant selecting this option than expected frequency (see Table 3).
Stacked bar plot showing the proportion of nuclear or ground war preferences among participants within each topic group.
Note: The y-axis represents the proportion (%) of participants in each topic group who selected each of the 3 options at choice 2. For example, in Topic 1 (N = 20), 70% (N = 14) chose ground war, 15% (N = 3) chose the 100,000-casualty nuclear strike, and 15% (N = 3) chose the 2 million-casualty nuclear strike. The x-axis lists the topic groups, ordered from largest to smallest proportion of GW. The solid red line indicates the overall average proportion of participants choosing ground war; the dashed red lines show the 95% confidence interval around this average. The label for Topic Group 6 was abbreviated to improve figure readability.
6.1.2. Topic 2: Desire to avoid civilian casualties to the extent it is possible
The topic includes answers that express a fervent desire to avoid civilian casualties, acknowledging that while such outcomes may be difficult to prevent entirely in a conflict, they should be avoided whenever possible.
Among the participants (N = 24; 0.99%) whose answer was best represented by this topic, 58.33% chose to continue the ground war, 37.50% opted for a nuclear strike killing 100,000 civilians, and 4.17% chose a nuclear strike resulting in 2 million civilian casualties.
6.1.3. Topic 3: It is time for payback
The topic answers describe a need for retribution as well as hate rhetoric toward out-group cultures. Further, it highlights that using nuclear weapons is a signal of US strength with a deterring effect, ‘learn to not mess with us’.
Among the participants (N = 198; 8.18%) whose answer was best represented by this topic, 18.18% chose to continue the ground war, 55.05% opted for a nuclear strike killing 100,000 civilians, and 26.77% chose a nuclear strike resulting in 2 million civilian casualties, which is the highest proportion of the 2 million option across all topic groups. In line with this, the standardized residuals show a much larger frequency than expected of the nuclear strike with 2 million civilian casualties, however, a smaller frequency than expected picking the ground war option was even more deviant relative to the frequency of all choices made in the total sample.
6.1.4. Topic 4: Wrong to kill innocent people
Topic 4 is the most prominent topic in the model. It accounts for the largest average share of content across responses (see Table 2). The topic includes responses describing that it is wrong to kill innocent civilians over military troops, troops who know the risk and still made the decision to join the war effort.
Among the participants (N = 550; 22.74%) whose answer was best represented by this topic, 94.00% chose to continue the ground war, 5.64% opted for a nuclear strike killing 100,000 civilians, and 0.36% chose a nuclear strike resulting in 2 million civilian casualties. Further, when studying the standardized residuals, we see that choosing ground war is more frequent than expected, while the other 2 options were less frequent.
6.1.5. Topic 5: Humanitarian rights and ethics
The topic includes responses describing ground war as the preferred option, as it is the option that best holds to moral and ethical standards that protect humanitarian principles. Further, the topic holds responses highlighting the negative long-term consequences of a nuclear strike.
Among the participants (N = 207; 8.56%) whose answer was best represented by this topic, 88.89% chose to continue the ground war, 8.70% opted for a nuclear strike killing 100,000 civilians, and 2.42% chose a nuclear strike resulting in 2 million civilian casualties. Further, when studying the standardized residuals, we see that choosing ground war was more frequent than expected, whereas a nuclear strike killing 100,000 civilians was less frequent.
6.1.6. Topic 6: A difficult choice: balancing the minimum loss of life for a decisive American victory with preserving American lives
The topic includes responses describing that all options felt wrong, but that they want the least American casualties while minimizing civilian death in Iran. Further, a large proportion of the responses pertaining to this topic describe the options as limited and restricting.
Among the participants (N = 244; 10.09%) whose answer was best represented by this topic, 28.69% chose to continue the ground war, 70.49% opted for a nuclear strike killing 100,000 civilians, and 0.82% chose a nuclear strike resulting in 2 million civilian casualties. Further, when studying the standardized residuals, we see that a nuclear strike killing 100,000 civilians was much more frequent than expected, while the 2 million strike were slightly less frequent, while the ground war was much less frequent.
6.1.7. Topic 7: American lives are worth more
The topic includes responses clearly outlining the cost and benefits of the 3 available options and while stating that saving American lives is prioritized over saving civilians, several mentions that killing 2 million civilians would be ‘extremely excessive’.
Among the participants (N = 332; 13.71%) whose answer was best represented by this topic, 5.72% chose to continue the ground war, 84.04% opted for a nuclear strike killing 100,000 civilians, and 10.24% chose a nuclear strike resulting in 2 million civilian casualties. Thus, participants whose answer was most represented by this topic had the highest proportion of all topics of participants choosing the nuclear strike killing 100,000 civilians. In line with this, when studying the standardized residuals, we see that a nuclear strike killing 100,000 civilians’ substantially larger frequency than expected, the 2 million nuclear strike was slightly more frequent, while ground war was substantially less frequent.
6.1.8. Topic 8: Worth it for a quick end to the war
The topic includes responses that emphasize the risks associated with a prolonged ground war, particularly the potential for significant loss of both civilian and military lives over time. Many respondents express a preference for a swift and decisive resolution to the conflict, as they believe that extended engagement could escalate the number of casualties and prolong the suffering.
Among the participants (N = 186; 7.68%) whose answer was best represented by this topic, 54.30% chose to continue the ground war, 39.79% opted for a nuclear strike killing 100,000 civilians, and 5.91% chose a nuclear strike resulting in 2 million civilian casualties. When studying the standardized residuals, we see that choosing the nuclear strike to kill 100,000 civilians was more frequent than expected, while the ground war was less frequent.
6.1.9. Topic 9: Less deaths in total
The topic includes responses that compare different options: ground war versus a 100,000 civilian nuclear strike, or a 100,000 civilian nuclear strike versus a 2 million civilian nuclear strike. While the specific comparisons differ, the underlying argument remains the same, respondents emphasize that either the ground war or the 100,000 civilian nuclear strike would result in fewer total deaths, relative to the compared option they are referring to.
Among the participants (N = 108; 4.47%) whose answer was best represented by this topic, 37.04% chose to continue the ground war, 62.04% opted for a nuclear strike killing 100,000 civilians, and 0.93% chose a nuclear strike resulting in 2 million civilian casualties. When studying standardized residuals, we see that ground war is less frequent than expected, while the 100,000 strike is more frequent.
6.1.10. Topic 10: A nuclear attack invites other nuclear attacks
Finally, Topic 10 had the largest proportion of participants having the topic as the answer that is most representative of their response. The topic includes responses that consistently highlight that if a nuclear strike is carried out, it has the risk of lowering the threshold for other countries to do the same and making the use of nuclear weapons more accessible—potentially normalizing its use in warfare.
Among the participants (N = 553; 22.84%) whose answer was best represented by this topic, 98.55% chose to continue the ground war, 1.45% opted for a nuclear strike killing 100,000 civilians, and 0.00% (N = 0) chose a nuclear strike resulting in 2 million civilian casualties. Thus, this topic has the highest proportion of participants choosing ground war, which is also clearly represented in the standardized residuals illustrating a much higher frequency than expected of participants in this group choosing ground war. Consequently, both nuclear options are clearly less frequent.
6.2. Individual differences
We again used each participant’s most representative topic, the one with the highest estimated theta value from the STM model, as a grouping variable with 10 categories, each corresponding to the participant’s most representative topic. Chi-square tests of homogeneity were then conducted to examine whether the distribution of topics varied across levels of each demographic (e.g., gender, political affiliation) and individual difference variables (e.g., PAN attitudes, ‘better bomb’ effect). Standardized corrected Pearson residuals were analyzed to identify which topics were over- or underrepresented within each group. Full results and tables can be found in Supplementary 1. Next, we highlight the main findings regarding the level on the PAN index and the degree of the ‘better bomb’ effect (see Figure 4).
Bar plot illustrates the relationship between the better bomb effect and endorsement of punitive–authoritarian–nationalist views, per topic group.
Note: The X-axes represent the deviation between expected and observed frequencies presented in percent (%) of participants in each topic group. Expected values are based on the overall distribution in the full sample. For example, in Topic 1, Balancing strategic, ethical, and reputational concerns (N = 20), about 6 participants were expected to have low punitiveness while the observed frequency was 8 participants = approximately 10% deviation. The y-axes list the topic groups in order from smallest to largest better bomb effect. The dotted black line demarcates which topics were most associated with preferences of GW or the 100K nuclear option. The label for Topic Group 6 (a difficult choice) was abbreviated to improve figure readability.
The support of PAN attitudes (low, medium, high) influenced the distribution of all but one topic. Individuals with low support were more consistently associated with thinking patterns favoring ground war. In contrast, individuals with medium to high support exhibited more variable thinking patterns, selecting both ground war and nuclear weapon preferences.
After selecting their preferred option in Choice 2 and explaining their reasoning, participants rated how much the presence of the 2 million-casualty strike made the 100,000-casualty strike seem more acceptable to them. This rating, referred to as the ‘better bomb’ effect, indicated whether participants found the comparison ‘very important’ (suggesting a strong better bomb effect) or ‘not at all important’ (suggesting no effect).
Among participants who reported a strong better bomb effect (i.e., very important), thinking patterns associated with a higher likelihood of choosing the 100,000-casualty nuclear option were more frequent. Conversely, among participants reporting no better bomb effect, thinking patterns favoring ground war were more prominent.
Taken together, Figure 4 illustrates a pattern where 9 of 10en topics (all except Topic 5; Humanitarian rights and ethics) show an inverse relationship: in topic groups where individuals frequently display a strong better bomb effect, few individuals display low support for PAN views, and vice versa.
6.3. Topic co-occurrence
In addition, we conducted a correlation analysis on the prevalence of each topic to see patterns of co-occurrence (see Table 4 for correlations and Table 5 for topic labels). Spearman’s correlations were conducted due to violations of assumptions for a Pearson’s correlation. All correlations were significant except for the one between Topics 4 and 7, thus the focus below is on the particularly strong relationships appearing between specific topics (rho ≥0.6).
Spearman’s correlation in topic prevalence
Note. All participants received a prevalence value from each topic ranging from 0 to 1 (i.e., 0–100%). These prevalent values were correlated. p < .001 = ***, p < .01 = ***, p < .05 = *. Strong correlations (rho ≥0.6) highlighted in bold.
Topic labels for readability
Topic 10 had 3 strong negative correlations, with Topics 6, 7, and 9. Interestingly, these 3 topics all had strong positive correlations with each other. This indicates that thinking the choice to be particularly difficult, describing American lives as worth more, and reflecting on the total amounts of deaths of relative options, seem often to co-occur in the same participant’s response. Further, the common theme across these appears to be the act of comparison. In contrast, those who reasoned along the lines of Topics 6, 7, and 9 did not consider the risk of a nuclear attack inviting retaliation attacks. As an example of a positive correlation, the thinking response below had a relatively high prevalence of both Topic 7 (43.23%) and Topic 9 (27.51%): ‘Being an American citizen, I prefer no American lives be lost [Topic 7]’. ‘100,000 versus 2 million lives of Iranians lost is much less. Still not the best outcome for Iranians but better than 2 million [Topic 9]’.
Further, Topic 1 had strong significant correlations with both Topics 2 and 3, which both also had significant positive correlations with each other. This suggests that participants who view their decision as a trade-off between strategic, ethical, and reputational concerns often express a desire to avoid civilian casualties, which can surprisingly co-occur with thoughts of retribution or the need to signal US military strength. To illustrate a positive correlation, the thinking response below had a relatively high prevalence of both Topics 1 (20.27%) and 3 (45.02%):
‘…. Prior to this I would be against the war, quite honestly and would have sent the missile the minute they hit national waters. There would be no discussion, it is within my authority I believe [Topic 1]’. ‘I say this because we are at a point where these weapons are at hand. I do believe we have the upper hand in general but unfortunately you can’t allow another country to be stronger in this aspect. Because they will certainly do it to us [Topic 3]…’.
7. Discussion
In the current global political climate, the risk of nuclear conflict remains a pressing concern (Bulletin of the Atomic Scientists, 2025; Wolfsthal et al., Reference Wolfsthal, Kristensen and Korda2025). Understanding how individuals’ reason about such existential is crucial. Our study examined the American public’s thinking behind their preference to use or not to use nuclear weapons in a hypothetical war with Iran. Using structural topic modeling on open-text responses, we identified 10 distinct thinking patterns that were strongly associated with participants’ decisions to either continue the ground war or opt for a nuclear strike—either with 100,000 or 2 million civilian casualties.
Across the 10 identified topics, the most representative topic for the largest proportion of participants also had the highest proportion of participants choosing not to deploy nuclear weapons. Thinking that ‘nuclear attacks invite other nuclear attacks’ (Topic 10), essentially arguing that a nuclear strike would lower the threshold for the use of nuclear weapons in warfare, had the largest proportion of participants choosing the ground war over either type of nuclear strike. Additionally, 2 other topics were associated with a high proportion of choosing the ground war. Notably, participants who reasoned that it is ‘wrong to kill innocent people’ (Topic 4) or those who focused on ‘humanitarian rights and ethics’ (Topic 5) had a lower frequency of participants opting for a nuclear strike. Put simply, individuals who engaged more in these 3 patterns of thinking chose not to use nuclear weapons.
These topics associated with low acceptance indicate that the reasons for choosing not to deploy nuclear weapons can be qualitatively diverse—reasoning that nuclear attacks may lower the threshold for future nuclear retaliation (‘nuclear attacks invite other nuclear attacks’), empathy for innocent victims (‘wrong to kill innocent people’), and moral and humanitarian standards (‘humanitarian rights and ethics’). In essence, ‘nuclear attacks invite other nuclear attacks’ can be viewed as a type of fear appeal, suggesting that using nuclear weapons now will lead to a higher risk of receiving the same treatment in future conflicts. Research has found that increased fear can lead to more risk-averse behavior (Nguyen and Noussair, Reference Nguyen and Noussair2014).
Thinking that it is ‘wrong to kill innocent people’, instead, seems to be driven by the focus on the potential victims, calling on a more instinctually empathetic response for avoiding nuclear weapons—with individuals as such appearing less susceptible to psychic numbing (Slovic, Reference Slovic2007). There is evidence that empathy for out-group members can increase the chances of conflict resolution. Empathy promotes prosocial behavior and more favorable attitudes toward out-group members, which can reduce aggression (Klimecki, Reference Klimecki2019).
Those concerned with ‘humanitarian rights and ethics’ emphasize the importance of following moral principles that hold human societies together. According to Erlandsson and Dickert (Reference Erlandsson and Dickert2024), moral principles such as responsibility, fairness, and not causing harm are fundamental in guiding prosocial behavior (Erlandsson and Dickert, Reference Erlandsson and Dickert2024). These patterns of reasoning may reflect different decision-making logics. Some participants appear to follow a logic of appropriateness, guided by moral norms and principles, while others rely on consequentialist reasoning, focusing on outcomes and trade-offs (Müller, Reference Müller2004).
By contrast, several reasoning patterns justified the use of nuclear weapons: Approximately 8% of participants reasoned that it was ‘time for payback’ (Topic 3). This type of thinking was associated with the largest proportion choosing the nuclear strike that would result in the death of 2 million civilians. However, the largest proportion choosing nuclear weapons overall (i.e., across both 100,000 and 2 million), with 94.3% choosing a nuclear strike, were individuals thinking that ‘American lives are worth more’ (Topic 7).
Reasoning that it is ‘time for payback’ or that ‘American lives are worth more’ seem to be linked to a devaluing of out-group lives of the Iranian civilians. Devaluing out-group human life has been shown to reduce prosociality and increase antisocial behaviors (for review, see Haslam and Loughnan, Reference Haslam and Loughnan2014). However, the mechanisms underlying devaluing of the out-group are complex, but findings suggest that emotional states (i.e., disgust), sociality needs (i.e., feeling socially connected), cognitive factors (i.e., self-centered thinking), perceived threats (i.e., fear of death), and power dynamics (i.e., feeling powerful), all are relevant factors (Haslam and Loughnan, Reference Haslam and Loughnan2014; Markowitz and Slovic, Reference Markowitz and Slovic2021). The theme of disgust—in the form of repugnance—a strong aversion expressed in statements such as ‘…get rid of their awful culture…’—found in responses clustered under the ‘time-for-payback topic’. In addition, the theme of self-centered in-group thinking was found in ‘American lives are worth more’ responses. Both topics focus on national or in-group security, aligning with the importance of ‘security prominence’ that is associated with genocide and other mass atrocities (Delaney and Slovic, Reference Delaney and Slovic2019).
In addition, the perception of the decision as ‘a difficult choice’ (Topic 6), the view that their preferred option would cause ‘fewer deaths in total’ (Topic 9), and the topic stating that ‘American lives are worth more’ (Topic 7)—are examples of participants weighing the costs and benefits of each option—albeit with different factors considered. This suggests that they may be approaching the decision more as a pragmatic trade-off, rather than being driven by moral condemnation of nuclear weapons, akin to what has been called a nuclear taboo (Tannenwald, Reference Tannenwald2007).
The tendency to choose a nuclear option, such as killing 100,000 civilians, could also be explained through the lens of prospect theory, decoy, and attraction effects (Huber et al., Reference Huber, Payne and Puto1982; Kahneman and Tversky, Reference Kahneman and Tversky1979; Wedell and Pettibone, Reference Wedell and Pettibone1996). The decoy effect makes one option appear more attractive when compared to a clearly inferior alternative, and the attraction effect explains how such comparisons shift trade-offs, nudging preference toward the better option. Prospect theory explains why extreme risks may seem justified in scenarios where losses are highly salient to people under threat. Thus, when presented with the options of 2 nuclear strikes and a ground war, the more extreme nuclear strike establishes a reference point that can makes the less extreme option, killing 100,000 civilians, appear better by comparison.
The fact that the goodness or badness of an option or outcome is not absolute, but rather a comparative process influenced by the options available to decision makers, is evident in the thinking of our participants. The 3 topics associated with pragmatic trade-offs and comparative thinking—‘a difficult choice; American lives are worth more; fewer deaths in total’—all show strong topic co-occurrences. These 3 topics also show the highest proportion of participants opting for the ‘better bomb’ (~62–84%)—killing 100,000 civilians. These findings could be seen as a type of compromise effect—taking the middle road—psychologically appearing that it can partially satisfy multiple factors relevant to the choice (Shaddy et al., Reference Shaddy, Fishbach and Simonson2021). Slovic et al. (Reference Slovic, Peterson, McDermott, Post and Västfjäll2025) argue that drawing meaning from comparisons is fundamental to human thinking. They note that the better bomb effect is closely related to the lure of the ‘lesser evil’ that goes back to Aristotle (Reference Aristotle, Crisp and Crisp2014), who observed that it appears in the mind as something good. Centuries later, in the aftermath of World War II, the moral philosopher Arendt (Reference Arendt1963) cautioned that the lesser evil was still evil. It is clear that lesser evil thinking is a common deception that confuses ‘betterness’ with goodness and acceptability.
Finally, one additional topic group was associated with an overrepresentation of participants preferring the 100,000 nuclear strike: the belief that it would be ‘worth it for a quick end to the war’ (Topic 8). Most participants with a ‘worth it for a quick end to the war’ thinking, justified their preference through a utilitarian (i.e., consequentialist) lens, arguing that while a nuclear strike would cause immense immediate harm, an extended ground war would ultimately lead to civilian casualties either way—‘The number of civilian deaths by continuing the ground war was not reported, but it is never zero’. From this perspective, the decision is framed as a trade-off between 2 undesirable outcomes—prioritizing reducing the duration of suffering rather than focusing on the immediate moral implications of using nuclear weapons. Such thinking aligns with consequentialist perspectives on military ethics, where the emphasis is placed on the outcome rather than the means of achieving it (Olsthoorn, Reference Olsthoorn2011). This type of reasoning points to an ethical divide in nuclear decision making. As Hardin and Mearsheimer (1985) describe, consequentialist arguments often support nuclear deterrence on the grounds that it prevents greater harm, even though it involves threatening massive civilian casualties. In contrast, deontological reasoning emphasizes that threatening or killing innocents is inherently immoral, regardless of potential benefits.
One can also argue that ending the war quickly is an expression of the peak-end rule of emotional memory (Kahneman et al., Reference Kahneman, Fredrickson, Schreiber and Redelmeier1993). In their seminal study, Kahneman et al. (Reference Kahneman, Fredrickson, Schreiber and Redelmeier1993) noted a tendency for individuals to place less weight on the total experience (i.e., affective sum) and instead focus on the most intense moment (i.e., peak effect) and the final part of the experience (i.e., end effect). In this context, the immediate saving of American troops can be seen as a dramatic peak that overwhelms the statistical numbing associated with saving 100,000 distant out-group lives.
Understanding the psychological mechanisms behind nuclear decision making is crucial for policy and diplomacy. How can this knowledge help reduce the threat of nuclear weapons? Perhaps we can start by recognizing certain patterns of thinking that would lead to the decision whether or not to choose nuclear weapons? Previous research has shown that this decision varies to a high degree depending on demographic factors and individual traits and characteristics (e.g., virtuous violence and PAN attitudes; Sagan and Valentino, Reference Sagan and Valentino2017; Slovic et al., Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020, Reference Slovic, Peterson, McDermott, Post and Västfjäll2025). The current research has extended these findings by identifying that the thinking patterns behind preferences for ground war or nuclear strikes are subject to individual differences.
Across the 10 topic groups, 2 traits stood out most strongly—endorsement of punitive–authoritarian–nationalist (PAN) views and the ‘better bomb’ effect. Acceptance of nuclear weapons use was associated with a profile of high PAN endorsement—and a strong better bomb effect. The reverse was true for those leaning toward ground war—low PAN endorsement and a weaker better bomb effect. As shown in Figure 4, topics with higher nuclear-use probability also showed stronger better bomb effects, while low PAN endorsement was underrepresented. This inverse relation suggests that high punitiveness may facilitate framing the 100,000-casualty strike as an acceptable ‘lesser evil’, whereas low punitiveness may be a psychological barrier to accepting any nuclear use (Arendt, Reference Arendt1963).
Rathbun and Stein (Reference Rathbun and Stein2020) found that a belief in retribution, understood as a strong endorsement of punishment, was closely linked to support for nuclear weapons use and a reduced sensitivity to civilian casualties. They also observed that individuals who strongly endorsed binding moral values—such as in-group loyalty, moral purity, and respect for authority—were more likely to support nuclear strikes, though these effects were less pronounced than those associated with retributive beliefs. These findings are consistent with earlier research showing that retributive and authoritarian traits predict support for punitive military actions, including nuclear strikes (Liberman, Reference Liberman2006; McFarland, Reference McFarland2005). Together, these studies highlight how moral rigidity, and blind patriotism can diminish empathy for civilian victims and increase acceptance of extreme violence.
Future research should assess the generality of these topics (i.e., thinking patterns) across cultures. Additionally, experimental research should examine whether acceptance of nuclear weapons can be influenced by leveraging these thinking patterns by introducing or removing options, for example, facilitating comparative thinking or highlighting short-term or long-term security risks, among others. One potential psychological intervention for reducing the likelihood of nuclear war could be to heighten fear associated with risks posed by choosing nuclear conflict, by vividly presenting the long-term consequences. While fear-based messaging could potentially trigger defensive reactions, such as heightened in-group security concerns (Delaney and Slovic, Reference Delaney and Slovic2019), meta-analytic evidence suggests that fear appeals are generally effective for promoting one-time behaviors, with minimal support for the occurrence of psychological reactance (Tannenbaum et al., Reference Tannenbaum, Hepler, Zimmerman, Saul, Jacobs, Wilson and Albarracín2015).
Beyond developing an understanding of nuclear decision-making psychology, past, present, and future findings should be integrated into protocols governing the use of nuclear weapons. Ignoring both public pressure on the President and military advisors—as well as their own susceptibility to cognitive biases—leaves decision makers vulnerable to unplanned influences when the 6 min start ticking (Jacobsen, Reference Jacobsen2024). In a companion paper, Slovic et al. (Reference Slovic2025) argue that decision-making procedures should be improved—by incorporating bias mitigation strategies, increasing resistance to comparative thinking, and as such the lure of the so-called ‘better bomb’. We propose that assessing the thinking patterns identified in the current study and applying them to hypothetical future scenarios in which nuclear warfare is on the table—while rigorously assessing both short- and long-term consequences—may help decision makers become better prepared and less likely to fall prey to unwise decisions.
8. Conclusion
Using structural topic modeling of open-ended responses to a hypothetical Iran war scenario, we identified a wide range of reasoning patterns that reflect the diverse moral, strategic, and emotional facets, Americans draw on when justifying or rejecting the use of nuclear weapons. The diverse reasoning patterns incorporate beliefs about the value of civilian life, retributive justice, protecting American troops, preventing nuclear escalation, and minimizing total casualties, among others. Comparative thinking triggered by the choice set—balancing the number of lives lost, and whether those lives are American or Iranian—was especially evident in topics favoring a ‘better bomb’, highlighting how moral trade-offs can make killing 100,000 civilians acceptable as a ‘lesser evil’.
Importantly, these reasoning patterns are unlikely to be limited to the public alone. Similar ways of thinking (e.g., psychic numbing; lesser evil comparisons, prioritizing near-term security over long-term humanitarian consequences) are fundamental to human cognition, thus likely to influence how presidents, military advisors, and policy elites respond in real nuclear crises (Berthet, Reference Berthet2022; Slovic et al. Reference Slovic2025).
The results of this study, along with those from a recent paper on crisis decision making (Keeney et al., Reference Keeney, Gregory and Slovic2025), highlight the importance of preplanned frameworks for thinking about nuclear decision making. Advisory frameworks, structured decision protocols, and tools like decision matrices that make moral and strategic trade-offs explicit and more comprehensive could result in more balanced and ethically considered outcomes. Similarly, training for military and civilian leaders on cognitive bias, framing effects, and the emotional pitfalls of comparative reasoning—such as the ‘better bomb’ effect—could strengthen decision making in critical moments. While some of these protocols and tools are labor intensive and may thus be resisted, simpler procedures such as checklists designed to guide reflection and warn of known biases may be quite effective in reducing impulsive emotion driven responses. For example, at the state level, mechanisms to slow the current arms race and restore arms control treaties that have been cancelled or allowed to lapse are essential. The alternative to not incorporating behavioral insights into procedures and policy making and continuing unabated on the current path is what the Bulletin of the Atomic Scientists warned is ‘a form of madness’ (Mecklin, Reference Mecklin2025).
The Iran war scenario used in this study is not just about choosing between 2 military options. It offers a glimpse into the full psychological, moral, and strategic complexity that real-world nuclear decisions involve. Preparing for that complexity—through education, planning, and reflection—is not just ideal—but essential. The future of nuclear restraint may depend as much on the integrity of human judgment as it does on strategy or technology.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/jdm.2025.10017.
Funding statement
Funding for this research came from the Emerging Challenges Fund and Nuclear Weapons Policy Fund at Longview Philanthropy.
Competing interest
We have no conflicts of interest to disclose.
Appendix A
Additional examples of the responses pertaining to the topics from the open-text responses following Choice 2
Note: Five participants response for each topic with the highest prevalence of the topic.
Appendix B
Punitive–authoritarian–nationalist index items
All respondents were asked the following 8 questions measuring their attitudes and opinions pertaining to punitive tendencies, authoritarianism, and sadism. Questions such as these were found by Slovic et al. (Reference Slovic, Mertz, Markowitz, Quist and Västfjäll2020) to be highly predictive of preference and support for using nuclear weapons against enemy civilians. Unless otherwise noted, the response options were as follows:
-
• Do you strongly favor, favor, oppose, or strongly oppose the death penalty for persons convicted of murder?
-
• To what extent do you believe in hell? (absolutely certain hell exists; fairly certain; not too certain; do not know; do not believe in hell)
-
• Our country desperately needs a mighty leader who will do what has to be done to destroy the radical new ways and sinfulness that are ruining us.
-
• Confederate monuments should not be removed from public spaces.
-
• Some people deserve to suffer.
-
• If a fetal heartbeat is present, a doctor should be prohibited from performing an abortion, unless it is necessary to save the mother’s life or ‘to prevent a serious risk of the substantial and irreversible impairment of a major bodily function.’
-
• What do you think is more important: To protect the right of Americans to own guns OR To control gun ownership?
-
• The southern border should be closed to stop illegal immigration into the United States.
We created a punishment index for each respondent by giving them one point for each of the 8 punitive attitudes or antisocial beliefs they endorsed.Footnote 5 Scores were categorized as low (0 or 1 items endorsed), medium (2, 3, or 4 items endorsed), and high (5–8 items endorsed) to achieve roughly equal numbers of respondents in each category.
Open access
