On 12 September 2025, I fielded a survey using CloudResearch’s Engage platform for AI interviewing and Connect platform for participant recruitment. The sample was of 520 respondents in the U.S., census-matched to the U.S. population based on gender, age, race and ethnicity.
Each respondent was asked two closed-ended questions based on the May 2025 CPOST survey (https://d3qi0qp55mx5f5.cloudfront.net/cpost/i/docs/CPOST_USPV_Survey_-_May_2025_Topline_-_PV_v3.pdf?mtime=1757700830). The responses to these two questions have been interpreted by CPOST as being a clear indicator of growing support for political violence (see e.g. https://www.nytimes.com/2025/06/16/opinion/political-violence-minnesota.html).
I will note that there were some minor differences in the wording of my questions and the wording in the CPOST survey. This was because the CPOST report from their May 2025 survey was not released publicly at the time I fielded my survey, so I based our wording on the phrasing that had been disclosed in the media earlier and the structure of questions in previous CPOST surveys (https://www.nytimes.com/2025/06/16/opinion/political-violence-minnesota.html).
The two questions were:
The use of force is justified to remove President Trump from office.”
(Strongly agree, Somewhat agree, Neither agree nor disagree, Somewhat disagree, Strongly disagree)
The use of the military is justified to stop protests against President Trump’s agenda.”
(Strongly agree, Somewhat agree, Neither agree nor disagree, Somewhat disagree, Strongly disagree)
After each of these two questions, respondents were asked a series of open-ended questions using Engage’s AI interviewing tool.
For the first question, the AI interviewer was instructed as follows:
“You are an expert on cognitive interviewing. Ask open-ended questions to understand how the respondent understood the question, why they answered the way they did, and what specific scenarios or stories they were thinking of when they answered the question.
Ask open-ended follow-up questions that encourage participants to share their emotions and detailed explanations. Probe for specific examples of scenarios or stories they thought of in their response. Please inquire more deeply about the following:
When answering this question, what did they think “use of force” meant?
Why did they choose the response they did?
What specific scenario did they envision for use of force to remove President Trump from office or are there any scenarios where they could see use of force being justified?
Listen and probe on: Language or phrasing, Emotion, Tone.
Continue with the interview until you get a FULL picture of what had occurred with this person. Ask for clarification or more details when responses are vague or general.
As no more than four to five open-ended questions. Always only ask one question at a time!”
For the second question, the AI interviewer was instructed as follows:
“You are an expert on cognitive interviewing. Ask open-ended questions to understand how the respondent understood the question, why they answered the way they did, and what specific scenarios or stories they were thinking of when they answered the question.
Ask open-ended follow-up questions that encourage participants to share their emotions and detailed explanations. Probe for specific examples of scenarios or stories they thought of in their response. Please inquire more deeply about the following:
When answering this question, what did they think “use of the military” meant?
Why did they choose the response they did?
What specific scenarios did they envision of the military being used to stop protests? What, exactly, did they think the military would be doing?
Under what, if any, scenarios would they think the use of the military would be justified?
Listen and probe on: Language or phrasing, Emotion, Tone.
Continue with the interview until you get a FULL picture of what had occurred with this person. Ask for clarification or more details when responses are vague or general.
As no more than four to five open-ended questions. Always only ask one question at a time!”
This section reports the basic attributes of the sample. As noted above, the sample was matched to U.S. census numbers based on gender, age, race and ethnicity. Because of this, there is some divergence in other areas. Nonetheless, there is still substantial diversity in other, non-matched areas like education, income and political ID. The tables below report these basic demographic characteristics of the sample.
This section reports the basline cross-tabulations between political ID and their responses to the closed-ended questions.
The table below shows the relationship between partisanship and support for “use of force” to remove President Trump from office. The results show that about 33% of Democrats support the use of force to remove President Trump. This is a little lower than the 40% reported by CPOST in May 2025, but is still quite high.
Next I look at responses to the question about military intervention into protests against President Trump’s agenda. Here, about 29% of Republicans agree that use of the military is justified. This is a little higher than in CPOST’s survey (25%), but is, again, quite close.
The problem with previous reporting of these results is that scholars and journalists have assumed that “use of force” matches some particular type of scenaio (i.e., illegal removal), that it takes place within current circumstances, and involves violence. “Force”, however, can take on multiple meanings in all three of these regards. Force, for example, could be taken by respondents to mean a type of extraordinary, illegal action such as a coup, assasination, or popular uprising. It could also, however, include legal action like impeachment, invocation of the 25th amendment, or the use of the police to enforce legal action (e.g., to remove a president who refuses to leave office after an election or the expiration of their term). In other words, they could think this question includes legal actions to “force” a president from office. Similarly, a respondent could be talking about actions taken in response to current situations, or they could be envisioning a scenario in the future (think of a movie plot). Finally, they could be thinking about violent action, as suggested by scholars who have used these numbers to suggest an increasing acceptance of violence, or they could be envisioning non-violent actions that still involve “force.”
To look at what people meant by their answers, I asked follow-up questions using an AI interviewing system with the prompt described above.
Once these interviews were completed, I analyzed the responses of those Democrats who said they “Strongly agree” or “Somewhat agree” that use of force is justified. I utilized GPT-4.1 on Microsoft Azure to code the interviews in terms of what the respondent understood “use of force” to mean, whether the scenario they envisioned was in response to current actions or in some hypothetical future scenario, and whether the scenario they envisioned involved violence. I also used Engage’s insight platform to verify the consistency of coding between AI systems and I read each interview to ensure the coding was credible. Two interviews were unable to be coded because of guardrails set by Azure. These were human coded. There was also one observation for which the LLM returned two responses. I checked these responses to ensure that the resulting data was identical. They were and the duplicate observation was removed.
I will start with whether the respondent viewed the “use of force” as being an illegal or legal action. In other words, were they picturing the use of something like a coup or were they thinking of actions like impeachment that would “force” the president from office? The results below suggest that 57% explicitly cited the use of legal mechanisms to remove the president from office, while another 9% gave answers that were not entirely clear. Only about 35% appear to have viewed their agreement with the statement about use of force as entailing illegal or extralegal actions to remove the president from office. As one respondent put it when asked what they thought “use of force” meant in this context, “I think using some sort of law or rule to force him out of office. Not violence…perhaps like the 25th amendment because he doesn’t seem like he’s right in the head.” Or, as another respondent put it, “I do not approve of Donald Trump in any way. But I don’t advocate violence for anybody in any way. When I think of”forcing” Trump out of office, I think of impeachment (which never seems to work for him but hey, it’s worth a shot) or finding something in the Constitution that would deem him ineligible to be in office.” Still another stated, “I was not thinking physically but more in the sense that he might need to be”fired” or forced out of office due to rules or laws.” While the authors of the question and those who marked disagreement with the question seemed to have interpreted the question as clearly involving illegal actions and violence, the large majority of those who agreed with the statement were not nearly so clear on this interpretation.
Of those who clearly thought that the “use of force” represented illegal or extralegal actions, 21% were not suggesting that this was justified in the current time. Rather they made it conditional on some future actions by the president that might or might not happen. In other words, they were suggesting that there might exist some extreme actions taken by the president in the future that would warrant extralegal action - usually this involved the president taking illegal action to hold onto power beyond his second term. As one respondent stated when describing a scenario in which physcial removal from office might become necessary, “I worry about what he will say, and what his rabid supporters will do to subvert democracy and hold on to power. I worry that this will be like an attempted coup on his part to remain in office, like he did with the 2020 election.” Indeed, the potential scenario where President Trump attempted a coup (or, more technically, an autogolpe or autocoup) to remain in power after his second term was a very consistent theme in many interviews. While these respondents still suggest the willingness of use of extralegal actions, these responses suggest an extreme hypothetical future far outside of what is defined in our legal system, not a support for such actions in the current environment.
Finally, for the remaining respondents, who see extralegal actions as justified in response to current actions by the president, about 17 respondents were coded by the AI as suggesting that violent action would be justified. Even among these respondents, however, the AI coding appears to have mostly been based on any mention of weapon involvement. Many of these responses were much more ambiguous about whether they supported violence. One of these 17 respondents, for example, pointed out the ambiguity of this question quite succinctly, “The language ‘use of force’ was a bit to broad for me. I could not justify killing Trump, for example, but less extreme uses of force were valid in my eyes.”
Taking these 17 respondents and dividing by the total number of Democrats in my sample suggests that only about 8% of the respondents are actually suggesting that illegal or extralegal violence is justified to remove President Trump for his actions to this point. This is much more consistent with the numbers from surveys that ask explicitly about violent actions (https://today.yougov.com/politics/articles/52960-charlie-kirk-americans-political-violence-poll) and is far less than the eye popping numbers found in the CPOST survey. The use of those numbers to say there is widespread support for violence in subsequent is, therefore, highly exaggerated - the numbers were more the result of the question asked and how the authors interpreted it, then actual support for violence.
## [1] 0.08292683There is a similar trend when looking at Republican support for use of the military in response to protests against President Trump’s agenda. Reporting on this seems to take positive responses to mean violent action against peaceful protests. But that is not really implied by the question wording. Both Republican and Democratic presidents and governors have deployed the National Guard into situations where protests either became or threatened to become violent, most notably during the Civil Rights Era.
Here again, I utilized AI interviews to ask respondents what scenario they were envisioning when they reponded to this question, and I utilized AI labeling to classify the responses. And, again, the respondents who expressed support for military intervention in protests were, on the whole, not arguing for the use of the military to suppress peaceful protests. Almost all of the respondents who gave positive answers to this question were suggesting that the military could intervene, now or in some future scenario, to respond to protests that turned violent.
I start by looking at the scenarios under which respondents thought use of the military would be justified. Only 4 respondents (about 12% of Republicans who agreed on this question) gave an indication that they might support the use of the military in response to non-violent protests. Almost all of these respondents believed they were suggesting the use of the military as a response to violence on the part of protesters. To quote one of the respondents, “I pictured protestors becoming rioters and vandalizing businesses, police cars, and police stations. I also pictured the protestors throwing items at the police.” Another was even more explicit, “I see the military coming and acting as a police force to stop or prevent the demonstrations that become violent. Peaceful protesters must be allowed to exercise their right to free speech.” Still another, “When protests become more about looting stores, burning cars and buildings, and hurting people who disagree with them, then people are not acting like civil Americans and thus should not be treated as such… Protesting is one thing but intruding upon other peoples/bystanders lives doesn’t help anyones [SIC] cause.”
Dividing this by the total number of Republicans in the sample suggests that the number of respondents who think deployment of the military was a response to non-violent protests against President Trump’s agenda suggests that these make up only around 2.5% of Republicans in the sample.
## [1] 0.02597403Additionally, most respondents who said they supported the use of the military, almost 52%, saw the role of the military as one of maintaining order, not breaking up protests. Only about 9% of the Republicans in this sample suggested the military should actively break up protests. When asked what type of actions one respondent thought the military would take, the respondent stated, “juste [SIC] them ensuring no harm come to the citizens of the US and making arrests where required.” Another said the military should, “JUST STAND AROUND KEEPING EVERYONE IN LINE [SIC].” And yet another explicitly rejected the violence interpretation of this question, “When I read ‘Use of the military,’ I pictured soldiers in uniform, maybe with trucks or gear, standing near protest to keep order. Not like attacking people, more like a strong presence that makes people think twice before things get out of hand. To me it was more about control and keeping peace, not violence.” Still another characterized the military’s role as: “They would make sure things dont [SIC] get out of hand . An example would be to give verbal warnings or arrests people.”
## [1] 0.0974026Finally, in terms of the means the military would be using in these situations, a minority saw the military’s role as potentially using violence, about 10% of all Republicans in the sample. A significant portion of respondents explicitly saw the military’s role in dealing with protests as non-violent - either as a symbolic action to maintain order or as playing a support role for police and law enforcement. As one respondent put it, “The military is there to support the local authorities so their police officers have support to stop a large number of people who are destroying infrastructure with burning and looting… They can assist the police force to keep up the resistance and enable them to arrest more people.” Another characterised the role of the military as, “Mainly crowd control and to stop groups from destroying property.”
## [1] 0.1038961Here again, while those who oppose these actions usually interpreted military action as being authoritarian action against peaceful political protests, this does not seem to be the way that people who expressed support for this question interpreted it (indeed, 13% of Democrats also agreed on this question). Instead, they largely viewed military action as a response in extreme circumstances where protests had become violent actions against people and property (more akin to riots).
I would also note that some additional caution is warranted when we are talking about numbers in the single digits with online surveys. Pew Research has suggested that most online survey panels, even those considered “high quality” in terms of respondents passing attention checks, have anywhere from 4% to 7% of respondents they characterize as “bogus respondents” (https://www.pewresearch.org/methods/2020/02/18/assessing-the-risks-to-online-polls-from-bogus-respondents/). These bogus respondents are likely to simply agree with any statement, even statements that contradict each other. For both of these questions, “agree” responses were indicative of greater use of force. In other words, we cannot discount that a significant portion respondents who did respond in a way consistent with the reporting around these numbers are roleplaying what they think they should have said, given that they had already agreed with the prior question. At a minimum, the valence of such important questions should be reversed to see how much of these percentages might stem from such “bogus” responses.
We should also be careful about interpreting statements online completely literally, since there are other reasons someone might respond with support for violence in an anonymous online setting. We have long known that people are willing to say things in anonymous online settings and suggest they would support actions that they would never advocate, support or do in real life - the “online disinhibition effect” (https://www.liebertpub.com/doi/abs/10.1089/1094931041291295?casa_token=4FVXj89kV8MAAAAA%3ANCAUlrxCcQ1FI-XEdGEljLZlUZHmK6H7rSZa2TxJxMWAf2OCt6kxl63GULi8zxEHh4f8AE5Vm2I). So we must be careful not to assume that these are statements of support for violence, rather than emotional expressions of general frustration. Put another way, these could be “speech acts,” where the respondent is using this online format to express frustration rather than to state their support for a real world action if they were given the choice to actually do it. Take, for example, one of the respondents who “Strongly agreed” with use of force to remove President Trump and whose interview suggested they meant violent removal for current actions, “It makes you feel hopeless, because if you can’t trust your leader to act rationally, then who can you rely on?”
Finally, we have little to no idea how firm these attitudes are. As John Zaller, among many others, have noted, survey respondents can respond very differently over time, depending on the salience of particular examples. Even responses that seem firmly held during one interview can reverse quickly, based on what examples, scenarios or stories are salient to the respondent when they are taking the survey. Among those who responded positively to use of the military to stop protests against President Trump’s agenda, there was clear indication among almost all respondents that the way they interpreted recent protests, as portrayed in the media sources to which they were exposed, as quite violent affairs. Their responses seem likely to be highly contingent on this media exposure, rather than general statements about support for such actions.
These, however, enter into the realm of speculation and should be read with caution by the reader. For an issue like support for political violence, where inaccurate poll results hold the potential to throw gasoline into an already polarized political fire, such caution seems warranted.
The bottom line is that the CPOST numbers, at least as it has been reported in the media, wildly exaggerate support for violent among both Democrats and Republicans. While there is plenty of legitimate evidence that political violence is an increasing problem in the United States, support for such actions is not widespread.
I will end by noting a few caveats to these findings. First, while I have reviewed the interviews and generally found the AI-generated labels to be credible (indeed, my evaluation made me think that the AI was still over-estimating support for violence), these results should still be considered highly preliminary. Language is ambiguous and I will be taking additional steps to verify the accuracy of the AI-generated (or human-generated for that matter) labels, including using multiple human coders, multiple runs with different LLMs (though I should note that I did run the labeling procedure five times and the overall percentages remained consistent), and inviting other scholars to review the results. Normally, I would take all of these steps prior to release of the data, but, in this case, the urgency of pointing out that these widely quoted statistics were not nearly as solid as they are being portrayed seemed to take precedence.
Second, I am sure that a critical reader will argue that people are less likely to express their true feelings directly rather than indirectly. This is likely part of the reason why CPOST likely used “use of force” in place of “use of violence,” thinking that people were more likely to provide their true attitudes in response to the more euphemistic phrase. This is certainly possible - social desirability bias is a well-noted phenomenon. Nonetheless, the inconsistency I note here between how CPOST’s results have been interpreted in public and what respondents said they meant when responding still stands as prima facie evidence that the public interpretation has been incorrect. Absent any evidence to the contrary, beyond pure conjecture by those doing the interpretation of vague multiple-choice answers, this is the best evidence we have for how people are interpreting these questions.
Finally, I should discuss the expected error of the percentages displayed above. While I could use statistical procedures to place estimates of sampling error on these estimates, there are multiple reasons to take them as guidelines rather than formal estimates of population parameters. The sample used here was not as carefully drawn as that done by NORC for the CPOST surveys, a luxury that both urgency and cost precluded for this report. Also, as with any qualitative research technique, generalizing from semi-structured interviews poses issues. Nonetheless, the percentages are highly informative and, at a minimum, suggest large portions of respondents are not interpreting the questions in the way that they are publicly being presented. There is no reason to assume that these differences in interpretation are any less prominent in NORC’s panel than in our CloudResearch sample.