Paper deep dive
The Hidden Costs of AI-Mediated Political Outreach: Persuasion and AI Penalties in the US and UK
Andreas Jungherr, Adrian Rauchfleisch
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/31/2026, 2:04:32 AM
Summary
This study investigates the 'persuasion penalty' and 'AI penalty' in political campaign outreach through a 2x2 experiment in the US and UK. The researchers find that both explicitly persuasive intent and the use of AI-mediated communication independently trigger negative evaluations, including reduced willingness to participate, perceived threats to autonomy, and decreased trust in organizations. The findings suggest that AI-mediated political outreach faces significant legitimacy challenges that may limit its real-world scalability.
Entities (6)
Relation Signals (3)
Andreas Jungherr â authored â The Hidden Costs of AI-Mediated Political Outreach
confidence 100% · The Hidden Costs of AI-Mediated Political Outreach: Persuasion and AI Penalties in the US and UK Andreas Jungherr 1 and Adrian Rauchfleisch 2
Persuasion Penalty â affects â Organizational Trust
confidence 90% · explicitly persuasive outreach is evaluated as... more damaging to organizational trust
AI Penalty â affects â Public Discourse
confidence 85% · AI-mediated outreach triggers normative concerns about appropriate communicative agents
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As AI-enabled systems become available for political campaign outreach, an important question has received little empirical attention: how do people evaluate the communicative practices these systems represent, and what consequences do those evaluations carry? Most research on AI-enabled persuasion examines attitude change under enforced exposure, leaving aside whether people regard AI-mediated outreach as legitimate or not. We address this gap with a preregistered 2x2 experiment conducted in the United States and United Kingdom (N = 1,800 per country) varying outreach intent (informational vs.~persuasive) and type of interaction partner (human vs.~AI-mediated) in the context of political issues that respondents consider highly important. We find consistent evidence for two evaluation penalties. A persuasion penalty emerges across nearly all outcomes in both countries: explicitly persuasive outreach is evaluated as less acceptable, more threatening to personal autonomy, less beneficial, and more damaging to organizational trust than informational outreach, consistent with reactance to perceived threats to attitudinal freedom. An AI penalty is consistent with a distinct mechanism: AI-mediated outreach triggers normative concerns about appropriate communicative agents, producing similarly negative evaluations across five outcomes in both countries. As automated outreach becomes more widespread, how people judge it may matter for democratic communication just as much as whether it changes minds.
Tags
Links
- Source: https://arxiv.org/abs/2603.27413v1
- Canonical: https://arxiv.org/abs/2603.27413v1
Trouble viewing inline? Open PDF directly â
Full Text
133,337 characters extracted from source content.
Expand or collapse full text
The Hidden Costs of AI-Mediated Political Outreach: Persuasion and AI Penalties in the US and UK Andreas Jungherr 1 and Adrian Rauchfleisch 2 1 University of Bamberg 2 National Taiwan University March 31, 2026 Abstract As AI-enabled systems become available for political campaign out- reach, an important question has received little empirical attention: how do people evaluate the communicative practices these systems represent, and what consequences do those evaluations carry? Most research on AI- enabled persuasion examines attitude change under enforced exposure, leaving aside whether people regard AI-mediated outreach as legitimate or not. We address this gap with a preregistered 2Ă2 experiment conducted in the United States and United Kingdom (N = 1,800 per country) vary- ing outreach intent (informational vs. persuasive) and type of interaction partner (human vs. AI-mediated) in the context of political issues that respondents consider highly important. We find consistent evidence for two evaluation penalties. A persuasion penalty emerges across nearly all outcomes in both countries: explicitly persuasive outreach is evaluated as less acceptable, more threatening to personal autonomy, less benefi- cial, and more damaging to organizational trust than informational out- reach, consistent with reactance to perceived threats to attitudinal free- dom. An AI penalty is consistent with a distinct mechanism: AI-mediated outreach triggers normative concerns about appropriate communicative agents, producing similarly negative evaluations across five outcomes in both countries. As automated outreach becomes more widespread, how people judge it may matter for democratic communication just as much as whether it changes minds. Keywords: Artificial Intelligence, persuasion, campaigning, survey, survey experiment, international comparison 1 arXiv:2603.27413v1 [cs.CY] 28 Mar 2026 1 Introduction The growing technological capabilities, commercial availability, and cultural nor- malization of AI-enabled systems have raised attention toward the uses for AI- enabled persuasion, the automation of changing peopleâs minds. A growing set of studies shows that AI can successfully change peopleâs minds be it through the generation of content for communicative interventions, such as ads or infor- mation treatments (Bai, Voelkel, Muldowney, Eichstaedt, & Willer, 2025; Chu & Liu, 2024; Goldstein, Chao, Grossman, Stamos, & Tomz, 2024; Hackenburg & Margetts, 2024; Hackenburg, Ibrahim, Tappin, & Tsakiris, 2026; Hackenburg, Tappin, Röttger, et al., 2025; Karinshak, Liu, Park, & Hancock, 2023; Matz et al., 2024; Simchon, Edwards, & Lewandowsky, 2024; Teeny & Matz, 2024), or automated dialogue between people and machines (Argyle et al., 2025; Boissin, Costello, Spinoza-MartĂn, Rand, & Pennycook, 2025; Chen et al., 2025; Chen, Kalla, & Le, 2026; Costello, Pennycook, & Rand, 2024; Costello, Pennycook, Willer, & Rand, 2025; Costello, Pennycook, & Rand, 2025; Costello et al., 2026; Crabtree, Holbein, Bosley, & Sevi, 2025; Czarnek et al., 2025; Hackenburg, Tappin, Hewitt, et al., 2025; Hölbling, Maier, & Feuerriegel, 2025; Hopkins, Costello, Pennycook, & Rand, 2026; Kowal et al., 2025; Lin et al., 2025; Rand, Zazai, & Stagnaro, 2025; Salvi, Ribeiro, Gallotti, & West, 2025; Schoenegger et al., 2025; White, Allen, Caviola, Costello, & Rand, 2025; Xu et al., 2025). Given these findings and growing AI use in campaigns in general (Foos, 2024; Jungherr, Rauchfleisch, & Wuttke, 2026; Neyazi, Khai Ee, & Kuru, 2025), the question is no longer if AI will be used in persuasion attempts but how and to what effect. This makes AI-enabled systems an important new element in persuasion that both academics and practitioners need to understand better. Much of the existing empirical literature focuses on persuasion outcomes for people who are interacting with AI-enabled persuasion in controlled conditions of survey experiments (Argyle et al., 2025; Boissin et al., 2025; Chen et al., 2025, 2026; Costello et al., 2024; Costello, Pennycook, Willer, & Rand, 2025; Costello, Pennycook, & Rand, 2025; Costello et al., 2026; Crabtree et al., 2025; Czarnek et al., 2025; Hackenburg, Tappin, Hewitt, et al., 2025; Hölbling et al., 2025; Hopkins et al., 2026; Kowal et al., 2025; Lin et al., 2025; Rand et al., 2025; Salvi et al., 2025; Schoenegger et al., 2025; White et al., 2025; Xu et al., 2025). Researchers examine whether AI-generated messages change attitudes, whether conversational AI can persuade people during dialogue, and which de- sign features or interaction patterns make such interventions more effective. This research has produced important insights into the persuasive potential of AI-enabled communication. At the same time, this research has remained vulnerable to a common challenge in campaign effects research. Identifying ef- fects of information treatments under controlled exposure conditions provides internally valid estimates that do not necessarily generalize to conditions of real-world exposure (external validity) (Druckman, Fein, & Leeper, 2012). This clearly also holds for findings on AI-enabled persuasion (Chen et al., 2025). To better understand real-world conditions for AI-enabled persuasion, we must not only look at persuasion but also address broader questions: How do 2 people evaluate campaigns that deploy AI-mediated persuasion systems? And, what consequences do those evaluations have for the legitimacy of the commu- nicative practice itself and the actors using it? The large-scale use of automated persuasion systems is constrained by whether people accept this practice, if they are willing to engage when they do encounter them, and if exposure does not systematically damage trust in the organizations responsible. If people evalu- ate AI-mediated negatively and as damaging to public discourse, the real-world consequences of automated persuasion may extend well beyond its measurable effects on attitudes. We therefore shift attention from persuasion outcomes to how people evaluate anticipated encounters with AI-mediated campaign out- reach and what those evaluations imply for the legitimacy and perceived ac- ceptability of these communicative practices. This question is particularly relevant in political communication. Campaign outreach tries to engage people in interactions (e.g. conversations with can- vassers, online exchanges, or chatbot dialogues) as a necessary precondition for exposure to persuasive messages. Whether people engage, and their evaluation of the encounter depend substantially on expectations about the interaction it- self: what it is trying to achieve, and who is conducting it. Two features of campaign outreach are especially consequential for these expectations. First, campaign outreach can take different forms: some focus on information presen- tation and exchange, while others directly aim at persuasion. Second, advances in generative AI have made it possible to conduct such outreach through auto- mated conversational systems rather than human campaigners. We argue that these features shape evaluations of anticipated campaign en- counters through related but distinguishable processes. One plausible interpre- tation is that explicit persuasive intent is likely to activate reactance, as people may perceive a threat to their freedom to hold their current opinions and there- fore evaluate the encounter more negatively and become less willing to engage. AI-mediated outreach, by contrast, is likely to trigger normative judgments about the appropriateness of agents involved in political communication. These processes may ultimately shape a broader evaluative stance toward the anticipated encounter. People form judgments about whether such interactions are acceptable and worth engaging in. Negative evaluations may therefore show not only in reduced willingness to participate but also in perceptions that the outreach threatens personal autonomy, harms public discourse, or reflects poorly on the organizations responsible for it. To examine these dynamics, we conduct a preregistered 2Ă2 experiment in the US (N = 1,800) and the United Kingdom (N = 1,800) that varies two features of campaign outreach on political issues that people consider highly important: whether the interaction is framed as informational or persuasive and whether the exchange partner is described as a human campaigner or an AI chatbot. We measure willingness to participate, perceived threat to autonomy, acceptability of the outreach, perceived positive impact, intentions to avoid future campaign contact, and evaluations of sponsoring organizations. Running the experiment in the US and the UK provides evidence from two countries where most prior AI-persuasion experiments have been conducted and 3 that differ in political systems and campaigning cultures. This variance allows us to speak about the impact of AI-enabled persuasion attempts on campaigning more broadly than, for example, focusing exclusively on the US. In measuring evaluative responses to anticipated rather than directly expe- rienced AI-mediated outreach, our design reflects two realistic conditions under which people encounter AI-enabled political persuasion. First, the majority of people will encounter AI-enabled persuasion systems not by interacting with them directly but through media coverage and public discourse. The society- wide evaluative and normative consequences of automated persuasion are likely to work primarily through mediated awareness rather than through direct inter- action. Our experiment is designed to approximate this condition: how people evaluate and respond to the prospect of AI-mediated outreach before and inde- pendently of any direct encounter with it. Second, the explicit disclosure of both AI mediation and persuasive intent in our vignettes reflects an emerging regula- tory stance. Regulatory frameworks in the US, EU, and UK are moving toward mandatory disclosure of AI-generated political content and AI use (Eisert & Marcus, 2025; European Parliament & Council of the European Union, 2024; Regulation (EU) 2024/900, 2024). Accordingly, people will increasingly en- counter digital campaign outreach under conditions of explicit disclosure. Test- ing both the role of persuasive intent and AI mediation under full disclosure, therefore, captures the conditions that policymakers are actively constructing and that people are increasingly likely to face. Our results reveal two systematic patterns. First, we observe a strong per- suasion penalty: outreach framed as explicitly persuasive is evaluated more negatively across five of six outcomes in the US and all six in the UK, including perceptions of autonomy threat, acceptability, positive impact, future campaign avoidance, and organizational trust. Second, we find consistent evidence for an AI penalty: AI-mediated outreach produces less favorable evaluations on the same five evaluative outcomes in both countries. These findings suggest that the societal consequences of AI-enabled campaign outreach cannot be assessed by examining persuasion effects alone: our findings indicate that when people become aware of AI-mediated campaign outreach, they form negative evaluative judgments about its legitimacy that carry downstream consequences for trust in campaigns and institutions. More broadly, our findings highlight an important constraint on the real- world scalability of automated persuasion. The effectiveness of AI-enabled per- suasion cannot be assessed solely by examining whether AI-generated messages change attitudes once people are exposed to them. It also depends on whether people are willing to enter such interactions and whether the practices used to initiate them are perceived as legitimate. If people view AI-mediated persuasion attempts negatively, automated persuasion may carry broader consequences for trust in political communication than opinion change, and may face significant exposure limits. 4 2 Theoretical framework Two features of campaign outreach are likely to shape peopleâs evaluations of an- ticipated campaign encounters: the intent of the outreach (informational vs. ex- plicitly persuasive) and the nature of the exchange partner (human vs. AI- mediated). We expect that these features activate related but distinguishable processes. Explicit persuasive intent is likely to trigger reactance: people per- ceive a threat to their freedom to hold their current opinions and respond by evaluating the encounter more negatively. AI mediation might trigger norma- tive judgments about the appropriateness of communicative agents: people may perceive the use of machines for political persuasion as violating expectations about how political communication should happen. Together these processes can be expected to produce negative evaluations of the anticipated encounter: a persuasion and an AI penalty. 2.1 Persuasion penalty One explicit goal of political campaigns is to change peopleâs minds (Foos & John, 2018). Research on campaign persuasion effects documents the con- ditions under which campaign contact and messaging shift political opinions (Broockman & Kalla, 2023; Coppock, 2022; Kalla & Broockman, 2018). How people evaluate the outreach itself (e.g., whether they regard persuasive tactics as legitimate, acceptable, or appropriate) is a separate question that largely re- mains unaddressed. This matters because evaluative reactions to the practice of persuasion shape whether people are willing to engage with it, how they assess the organizations responsible for it, and what broader consequences exposure carries for trust in political communication. We ask whether signaling persua- sive intent triggers negative evaluations of the anticipated encounter before any message is received. Reactance theory suggests how people are expected to respond to explicit persuasion attempts (Brehm & Brehm, 1981). When people perceive an at- tempt to influence or constrain their opinions, they experience a motivational response aimed at restoring their freedom of choice (Steindl, Jonas, Sitten- thaler, Traut-Mattausch, & Greenberg, 2015). Alerting people to a messageâs persuasive intent reduces both persuasion and willingness to engage (Benoit, 1998; Petty & Cacioppo, 1979). When people recognize the strategic nature of a communication â what Friestad and Wright Friestad and Wright (1994) term the activation of persuasion knowledge â they adopt coping responses in- cluding skepticism and disengagement. These effects grow with the perceived importance of the threatened freedom (Brehm & Brehm, 1981; Dillard & Shen, 2005). We therefore expect: H1a (Willingness to participate): People will be less interested in engaging when outreach is persuasive rather than informational. H1b (Perceived threat to freedom): People will rate outreach as a higher threat to freedom when it is persuasive rather than informational. 5 H1c (Acceptability): People will rate outreach as less acceptable when it is persuasive rather than informational. H1d (Positive impact): People will perceive outreach as having less positive impact when it is persuasive rather than informational. In our study, we focus on issues that are important to people, since much of the AI-enabled persuasion research focuses on high-salience topics such as immigration (Argyle et al., 2025; Rand et al., 2025), climate change (Czarnek et al., 2025; Remshard et al., 2026), conspiracy beliefs (Boissin et al., 2025; Costello et al., 2024; Costello, Pennycook, & Rand, 2025; Costello et al., 2026; Hopkins et al., 2026), or the correction of misinformation (DiGiuseppe & Robison, 2026; Goel et al., 2025). In such scenarios, where issues are regarded as personally important, reactance should be especially strong (Dillard & Shen, 2005). 2.2 AI penalty Independent of persuasive intent, the AI-mediated nature of outreach may itself produce negative evaluations through a related but distinguishable mechanism. AI-mediated communication, which Hancock, Naaman, and Levy (2020) de- fine as interpersonal communication in which a computational agent modifies, augments, or generates messages on behalf of a communicator, transforms the character of the interaction even when content remains unchanged. Where re- actance responds to the intent of a message, the AI penalty responds to the perceived appropriateness of the agent delivering it. The structural properties of AI that make it normatively inappropriate may also register as a threat to the conditions under which people form opinions freely, producing some elevation of freedom-threat perceptions through this route even in the absence of explicitly persuasive intent. Political persuasion carries implicit expectations about authenticity and ac- countability. When a human campaigner contacts a voter, the interaction can be read as expressing genuine conviction or civic commitment, even when strate- gically motivated. Field experiments consistently show that in-person contact is more effective for mobilization than other interventions (Green & Gerber, 2023), and deep canvassing (an approach that relies on non-judgmental nar- rative exchange and perspective-taking) has been shown to durably shift en- trenched attitudes (Broockman & Kalla, 2016; Kalla & Broockman, 2020). One reason driving these findings is that a human exchange partner can, at least in principle, be held socially accountable for what they say and how they say it. Delegating outreach to a machine disrupts these expectations. Research con- sistently documents resistance to algorithmic judgment in domains perceived as subjective, identity-relevant, or morally charged (Castelo, Bos, & Lehmann, 2019; Dietvorst, Simmon, & Massey, 2015; Longoni, Bonezzi, & Morewedge, 2019; Morewedge, 2022). This resistance is strongest when tasks require the sen- sitivity to individual circumstances and moral reasoning associated with human minds (Bigman & Gray, 2018; Cadario, Longoni, & Morewedge, 2021). Political 6 opinion formation clearly shares these characteristics. When people learn that a communicative agent is a machine, they attribute stereotypically mechanical properties to it (e.g., impersonality, strategic optimization, emotional coldness) even when its output is indistinguishable from human communication (Sundar, 2008, 2020; Yan & Sundar, 2024). These attributions are acceptable in technical contexts where objectivity is valued (Lee, 2018), but conflict with what political communication is expected to be. Prior work in political contexts directly sup- ports this. AI involvement in campaign outreach is largely perceived as a norm violation (Jungherr et al., 2026), and AI facilitation of democratic deliberation reduces both willingness to participate and evaluations of expected quality rel- ative to identical human-facilitated formats (Jungherr & Rauchfleisch, 2025). We therefore expect: H2a (Willingness to participate): People will be less interested in engaging when outreach is AI-mediated rather than human-mediated. H2b (Perceived threat to freedom): People will rate outreach as a higher threat to freedom when it is AI-mediated rather than human-mediated. H2c (Acceptability): People will rate outreach as less acceptable when it is AI-mediated rather than human-mediated. H2d (Positive impact): People will perceive AI-mediated outreach as having less positive impact than human-mediated outreach. 2.3 Downstream effects Both reactance and perceived norm violation can extend beyond the immediate evaluative response, shaping how individuals view the organizations responsible for the outreach. Research on persuasion knowledge shows that recognizing in- fluence attempts prompts negative updating of the persuading agent (Friestad & Wright, 1994; Campbell & Kirmani, 2000), and research on campaign communi- cation shows that objectionable practices can depress broader political engage- ment (Fridkin & Kenney, 2011; Lau, Sigelman, & Rovner, 2007). We therefore predict that both persuasive framing and AI mediation will produce downstream consequences beyond the immediate encounter: H1e (Future campaign avoidance): People will report greater future avoidance when campaign outreach is persuasive rather than informational. H1f (Penalty for source): People will report more negative evaluations of as- sociated organizations when campaign outreach is persuasive rather than informational. H2e (Future campaign avoidance): People will report greater future avoidance when campaign outreach is AI-mediated rather than human-mediated. H2f (Penalty for source): People will report more negative evaluations of asso- ciated organizations when campaign outreach is AI-mediated rather than human-mediated. 7 2.4 Interaction: how persuasive intent and AI mediation combine While we assume that the persuasion penalty and the AI penalty follow from distinguishable processes, they need not operate independently. The properties that make AI mediation normatively objectionable (e.g., perceived strategic op- timization, scalability, and absence of social accountability) also bear directly on how threatening a persuasive attempt feels. A human campaigner who tries to persuade operates within a framework of mutual social obligation; a ma- chine that tries to persuade may be perceived as an instrument of asymmetric, frictionless manipulation with no social skin in the game. AI mediation may therefore not merely add a normative penalty on top of reactance but actively intensify it, making the encounter feel simultaneously more threatening and more normatively inappropriate. Under this amplification logic, the combina- tion of persuasive intent and AI mediation should produce reactions stronger than either feature would generate alone. A competing prediction follows from a different theoretical starting point. Research on impression formation and evaluative judgment consistently finds that people integrate multiple pieces of information through averaging rather than addition, meaning that each successive piece of same-valenced information produces a smaller marginal shift in the overall evaluation than the one before it (Anderson, 1971). This averaging dynamic is reinforced diminishing sensitivity: the subjective impact of negative features decreases as evaluations move further from a neutral reference point, comparable to patterns known from decision- making and judgment research (Kahneman & Tversky, 1979; Tversky & Kahne- man, 1992). Under this attenuation logic, if persuasive intent already triggers strong reactance and drives evaluations substantially downward, AI mediation has limited room to worsen them further and vice versa. The two penalties would then substitute for each other rather than compound, with each succes- sive negative feature contributing less to the overall evaluation than it would have in isolation. Both predictions are theoretically coherent but draw on different arguments: amplification rests on the specific ways AI mediation makes persuasive intent appear more calculated; attenuation rests on structural properties of evaluative judgment that apply regardless of mechanism content. We preregistered the am- plification prediction on the grounds that the conceptual link between the two mechanisms was specifically reinforcing rather than merely additive. The prop- erties of AI that may trigger normative concern are precisely the properties that make a persuasive attempt feel harder to resist. Attenuation, though plausible as a general feature of evaluative judgment, required no specific assumptions about these mechanisms in particular. H3 (Interaction effect): For each outcome (aâf), the negative effect of per- suasive (vs. informational) outreach will be stronger when outreach is AI- mediated (vs. human-mediated). Running the experiment in the US and UK provides an opportunity to ex- 8 amine whether this interaction varies with the political communication envi- ronment. US campaigns operate across longer electoral cycles with higher con- tact frequency and less regulated spending (Sides, Shaw, Grossmann, & Lipsitz, 2026); UK campaigns are more compressed and subject to stricter broadcast and expenditure rules (Ford, Bale, Jennings, & Surridge, 2025). These differences may shape the baseline intensity of reactance that persuasive intent produces. If persuasive outreach is less normalized in the UK and a less routine feature of political life, it may be evaluated more categorically, leaving less evaluative space for AI mediation to add further deterioration and producing attenuation where the US shows more independent additive effects. 2.5 Moderators Individual differences in how people evaluate AI-mediated outreach are likely to shape the magnitude of both the persuasion penalty and the AI penalty. We examine three individual-level moderators that bear on peopleâs openness to political contact and their prior orientations toward AI. First, people who habitually avoid political conversations may respond more negatively to any form of campaign outreach, but may be especially sensitive to AI-mediated contact, which removes the social cues and interpersonal ac- countability that can make human conversations feel manageable (Boland & Davidai, 2024). Second, feelings toward people with opposing opinions capture the degree of affective polarization that people bring to anticipated encounters with campaign outreach (Lelkes & Westwood, 2017). Those with more neg- ative feelings toward political opponents may evaluate any outreach from an opposing campaign more harshly, thereby amplifying both the persuasion and AI penalties. Third, general AI risk perceptions should be especially relevant to evaluations of AI-mediated outreach specifically: people who perceive AI as threatening or risky are more likely to read AI involvement in political com- munication as a normative violation, amplifying the AI penalty (Jungherr & Rauchfleisch, 2025). While prior research suggests that these factors matter in related contexts (Jungherr & Rauchfleisch, 2025; Jungherr et al., 2026), the- ory provides limited guidance on the strength of these effects in our setting. We therefore formulate these expectations as research questions rather than hypotheses: RQ1 Does avoidance of political conversations moderate the effect of AI-mediated outreach? RQ2 Does the feeling toward people with opposing opinions moderate the effect of AI-mediated outreach? RQ3 Does the general AI risk perception moderate the effect of AI-mediated outreach? 9 3 Materials & Methods To test the preregistered hypotheses, 1 we conducted two parallel 2 (outreach mode: human vs. AI-mediated)Ă 2 (outreach intent: informational vs. per- suasive) between-subjects experiments, one in the United States and one in the United Kingdom, both using the Prolific platform. In each country, we set a quota of 1,800 completion (â450 per condition), based on a power analysis conducted prior to the experiment and documented in the preregistration, and stopped data collection once the quota was reached. The design was reviewed and approved by the IRB of the University of Bamberg. The US study ran from 6 March to 14 March 2026; the UK study ran from 9 February to 20 February 2026. Following preregistered criteria, we excluded participants who failed two attention checks prior to treatment exposure; these were directly excluded from data collection and did not count toward the total of 1,800 completes per country. We applied quota-based sampling to approximate each countryâs population by age, gender, and party identification (see Supple- mentary Information A.2 for sociodemographic distributions and data-quality procedures). The study design, questionnaire structure, treatments, and measures were identical across both countries. After participants provided informed consent, we began each survey with questions about sociodemographics and political orientation, followed by our preregistered moderators: avoidance of political conversations (four items, adapted from Boland & Davidai, 2024) and general AI risk perceptions (four items; Jungherr & Rauchfleisch, 2025), both showing good internal consistency in both samples (see Table 1). To prevent negative priming of AI, we also included items asking about AI benefits, which were not used for further analysis. Before participants viewed one of the four vignettes, we asked them to think about âone political or social issue that matters to you personallyâ and note it in a free-text field. To assess ecological validity, we also asked them to rate how important the issue is to them (1 = Not at all important; 7 = Extremely important). This research design choice was intentional: based on pretesting, we expected participants to reliably select issues they regarded as personally important, allowing us to reduce variation in baseline issue importance across respondents while increasing experimental realism by focusing on issues that matter to them. Indeed, issue importance ratings were high in both countries (US: M = 6.46, SD = 0.79; UK: M = 6.22, SD = 0.90), indicating that participants generally selected issues they cared about deeply. The issue profiles also broadly matched each countryâs political context and partisan cleavages (see Supplementary Information D). Furthermore, in both countries, fewer than 1% of participants selected a value below the scale midpoint. We used this approach to study campaign outreach in a context 1 PreregistrationUS: https://osf.io/sjh3v/overview?view_only= a6e7ef77368843bd9cdca37904712114 ; Preregistration UK: https://osf.io/nceht/ overview?view_only=93038c0124e431fb3de6dec8b95acc. The data and code to re- produce the findings reported in this study are available on OSF: https://osf.io/zvr7f/ overview?view_only=21c8e4c35d3642f79dfa64d1b9d55d87 10 close to that of much of the existing AI-enabled persuasion literature, which has largely focused on high-salience issues. We also asked respondents about their feelings toward people with opposing opinions (Lelkes & Westwood, 2017). Participants were then randomly assigned to one of the four treatments and required to remain on the page for at least 10 seconds (seconds on page US: M = 21.66, SD = 25.98; UK: M = 20.25, SD = 26.30). In the different conditions, only the outreach mode and outreach intent were varied; all other wording was held constant, stating that the campaign behind the outreach âtakes a position that is different from yoursâ (see Supplementary Information B.3 for the full treatment texts). We then measured willingness to participate in such outreach, followed by acceptability, perceived positive impact, future campaign avoidance, source penalty, and perceived threat to freedom (Dillard & Shen, 2005). Except for willingness to participate (a single item) and source penalty (four items), all outcomes were measured with three items each and showed good internal consistency in both samples (see Table 1; full item wordings and item-level descriptive statistics are provided in Supplementary Information B). At the very end, we also included a few questions about the treatment content, confirming that participants received the treatments (see Supplementary Information C.4 for manipulation checks and instrumental variable robustness checks). We tested hypotheses using OLS regression models with effect coding for both factors (-0.5, +0.5) and their interaction. With this coding scheme, the main-effect coefficients represent the mean difference between the two levels of each factor, averaged over the other factor. We estimated models separately for each country. For the research questions, we first centred the moderators and added them as moderators for the outreach mode variable (human vs. AI- mediated) in an OLS regression model. 11 VariableαUS M (SD) UK M (SD) n Outcome variables H1/2a: Willingness to participate â3.58 (2.05) 3.80 (1.90) 1,800 H1/2b:Perceived threat to freedom US = .86 / UK = .87 3.61 (1.70) 3.81 (1.59) 1,800 H1/2c: AcceptabilityUS = .79 / UK = .80 3.63 (1.51) 3.59 (1.38) 1,800 H1/2d: Perceived posi- tive impact US = .84 / UK = .84 3.70 (1.59) 3.87 (1.42) 1,800 H1/2e: Future cam- paign avoidance US = .90 / UK = .90 4.51 (1.71) 4.50 (1.59) 1,800 H1/2f: Penalty for source US = .84 / UK = .84 4.38 (1.36) 4.53 (1.24) 1,800 Moderator variables RQ1: Avoidance of po- litical conversations US = .88 / UK = .86 4.39 (1.81) 3.84 (1.66) 1,800 RQ2: Feelings toward people with opposing opinions â35.71 (29.19) 36.55 (25.24) 1,800 RQ3: AI risk percep- tion US = .82 / UK = .81 5.15 (1.32) 5.07 (1.16) 1,800 Table 1: Descriptive Statistics for All Outcome and Moderator Variables. Al- most all outcome items used 7-point scales (1 = Strongly disagree / Not at all interested; 7 = Strongly agree / Very interested). Feelings toward others used a 0â100 scale. All internal consistency coefficients are Cronbachâs α. n = 1,800 per country. 4 Results We first analyzed the preregistered persuasion penalty across all outcome vari- ables (H1aâH1f), comparing persuasive with informational outreach, followed by the AI penalty across the same outcomes (H2aâH2f), comparing AI-mediated with human-mediated outreach. We then analyzed the interaction hypothesis (H3), namely, whether the negative effect of persuasive (vs. informational) out- reach would be amplified when outreach was AI-mediated rather than human- mediated. Finally, we examined three preregistered research questions about individual-level moderators of the AI penalty: avoidance of political conversa- tions (RQ1), feelings toward people with opposing opinions (RQ2), and AI risk perception (RQ3). We report results for the US and UK in parallel throughout and explicitly note where findings diverge. Complete model tables are reported in Supplementary Information C. The preregistered main effects for the per- 12 suasion penalty (H1) and AI penalty (H2) across outcomes and countries are summarized in Figure 1. United StatesUnited Kingdom â0.50.00.51.0â0.50.00.51.0 H1f/H2f: Penalty for source H1e/H2e: Future campaign avoidance H1d/H2d: Positive impact H1c/H2c: Acceptability H1b/H2b: Perceived threat to freedom H1a/H2a: Willingness to participate Estimate (95% CI) Persuasion vs. informationalAI vs. human Figure 1: Estimated effects of persuasive outreach (H1, orange) and AI-mediated outreach (H2, blue) on six outcome variables, for the US (left) and the UK (right). Points are OLS regression coefficients with 95% confidence intervals. The coefficients represent the mean difference between the two levels of each factor, averaged over the other factor. Non-significant estimates are transparent. 4.1 Persuasion penalty In the US, results for the persuasion penalty were largely consistent with H1bâ H1f; only the willingness to participate (H1a) was not significant. Persuasive outreach did not significantly reduce willingness to participate in the US (b = -.18, p = .059, 95% CI [-.37, .01]). In the UK, however, all persuasion-penalty hypotheses were supported, including willingness to participate (b = -.23, p = .009, 95% CI [-.41, -.06]). For the remaining outcomes, the pattern was the same in both countries. H1b was supported: persuasive outreach increased perceived threat to freedom in both the US (b = .87, p < .001, 95% CI [.72, 1.02]) and the UK (b = .79, p < .001, 95% CI [.64, .93]). In line with H1c, persuasive outreach reduced perceived acceptability in both countries (US: b = -.50, p < .001, 95% CI [-.63, -.37]; UK: b = -.57, p < .001, 95% CI [-.69, -.44]). Our data also supported H1d, as persuasive outreach reduced perceived positive impact in both the US (b = -.30, p < .001, 95% CI [-.44, -.15]) and the UK (b = -.31, p < .001, 95% CI [-.44, -.18]). The two hypotheses concerning downstream negative reactions 13 were likewise supported. In line with H1e, persuasive outreach increased future campaign avoidance in both the US (b = .42, p < .001, 95% CI [.26, .58]) and the UK (b = .56, p < .001, 95% CI [.42, .70]). Consistent with H1f, persuasive outreach increased negative evaluations of the associated organization in both the US (b = .38, p < .001, 95% CI [.26, .50]) and the UK (b = .40, p < .001, 95% CI [.29, .51]). Overall, the persuasion-penalty hypotheses found consistent support across both countries and five of the six outcome variables. The one exception was the willingness to participate in the US, which was not significant. 4.2 AI penalty Results for the AI penalty were highly consistent across the two countries. In neither the US nor the UK did AI-mediated outreach significantly affect will- ingness to participate, providing no support for H2a in either sample (US: b = .02, p = .820, 95% CI [-.17, .21]; UK: b = -.03, p = .696, 95% CI [-.21, .14]). Supporting H2b, AI-mediated outreach increased perceived threat to free- dom in both the US (b = .19, p = .016, 95% CI [.04, .34]) and the UK (b = .27, p < .001, 95% CI [.13, .42]). Supporting H2c, AI-mediated outreach reduced acceptability in both the US (b = -.58, p < .001, 95% CI [-.72, -.45]) and the UK (b = -.55, p < .001, 95% CI [-.67, -.43]). Supporting H2d, AI-mediated outreach also reduced perceived positive impact in both the US (b = -.44, p < .001, 95% CI [-.58, -.29]) and the UK (b = -.31, p < .001, 95% CI [-.44, -.18]). In line with H2e, AI-mediated outreach increased future campaign avoidance in both the US (b = .38, p < .001, 95% CI [.23, .54]) and the UK (b = .50, p < .001, 95% CI [.35, .64]). Consistent with H2f, AI-mediated outreach increased negative evaluations of the associated organization in both the US (b = .55, p < .001, 95% CI [.43, .68]) and the UK (b = .48, p < .001, 95% CI [.37, .59]). Thus, the AI penalty was supported for five of the six outcomes in both countries, with willingness to participate as the consistent exception. 4.3 Interactions We next tested whether the negative effect of persuasive (vs. informational) outreach was amplified when outreach was AI-mediated (H3). Contrary to H3, significant interaction effects, when present, indicated attenuation rather than amplification of the persuasion penalty under AI mediation. Furthermore, the pattern of results differed considerably between the two countries, with no sig- nificant results in the US but four interactions indicating attenuation in the UK. For willingness to participate, the interaction was significant in the UK (b = .40, p = .025, 95% CI [.05, .75]), indicating that, contrary to H3, the persuasion penalty on willingness was attenuated rather than amplified under AI-mediated outreach. In the US, the interaction was not significant (b = .25, p = .202, 95% CI [-.13, .63]). For perceived threat to freedom, neither country yielded a significant interaction (US: b = -.17, p = .269, 95% CI [-.48, .13]; UK: b = 14 -.25, p = .083, 95% CI [-.54, .03]). The interaction was also not significant for acceptability in either country (US: b = .20, p = .148, 95% CI [-.07, .47]; UK: b = .21, p = .101, 95% CI [-.04, .45]). For perceived positive impact, the two countries diverged. In the UK, the interaction was significant (b = .39, p = .003, 95% CI [.13, .65]), again reflecting attenuation: the AI penalty on positive impact was smaller in the persuasive condition than in the informational condition. In the US, the interaction was not significant (b = .08, p = .600, 95% CI [-.21, .37]). We observed a similar divergence for future campaign avoidance: the interaction was significant in the UK (b = -.35, p = .017, 95% CI [-.64, -.06]) but was not in the US (b = -.27, p = .093, 95% CI [-.58, .04]). Finally, for the penalty for source, the interaction was significant in the UK (b = -.30, p = .008, 95% CI [-.52, -.08]) but not in the US (b = -.13, p = .293, 95% CI [-.38, .11]). These UK-specific interaction patterns are visualized in Figure 2 for willingness to participate, positive impact, future campaign avoidance, and penalty for source. The amplification hypothesis found no support in either sample. In fact, the significant UK interactions consistently indicated the reverse: persuasive outreach attenuated rather than amplified the AI penalty. (a) Willingness to participate(d) Positive impact(e) Future campaign avoidance(f) Penalty for source InformationalPersuasiveInformationalPersuasiveInformationalPersuasiveInformationalPersuasive 1 2 3 4 5 6 7 Outreach intent Estimated marginal mean (95% CI) Outreach mode HumanAI Figure 2: Estimated marginal means for four outcomes with significant inter- action effects in the United Kingdom, by outreach intent (informational vs. persuasive) and outreach mode (human vs. AI). Error bars represent 95% con- fidence intervals. To better understand these patterns, we conducted preregistered planned simple-effects contrasts with Holm-adjusted p-values, comparing AI- versus human- mediated outreach separately within the informational and persuasive condi- tions (see Supplementary Information C.2 for the full set of contrasts). Within the informational outreach condition, AI-mediated outreach generally yielded less favorable evaluations than human-mediated outreach across nearly all out- comes in both countries, with the sole exception of willingness to participate. Within the persuasive outreach condition, the AIâhuman gap became smaller on several measures, especially in the UK, while remaining significant for ac- 15 ceptability, future campaign avoidance, and penalty for source. The clearest cross-national difference concerned perceived positive impact: in the US, AI- mediated persuasive outreach still yielded significantly less favorable evaluations than human-mediated persuasive outreach, whereas in the UK this difference was no longer significant. At the same time, the observed attenuation for some of the variables in the UK did not mean that the AI penalty disappeared: within the persuasive condition, AI-mediated outreach was still evaluated significantly less favorably than human-mediated outreach for two of the four outcomes that showed attenuation, namely future campaign avoidance and source penalty. 4.4 Moderators We also preregistered three moderator variables and tested their interactions with the AI-versus-human outreach factor. Full model tables are reported Sup- plementary Information C.3 4.4.1 RQ1: Avoidance of Political Conversations Results for avoidance as a moderator diverged substantially between the two countries. In the UK, avoidance of political conversations significantly moder- ated the AI penalty on willingness to participate (b = .15, p = .004, 95% CI [.05, .26]): participants with higher avoidance showed a more positive response to AI- mediated compared to human-mediated outreach. In contrast, this interaction was not significant in the US (b = .10, p = .070, 95% CI [-.01, .20]). For perceived threat to freedom, the moderation pattern pointed in the op- posite direction in the UK: higher avoidance was associated with a smaller AI penalty on perceived threat (b = -.09, p = .034, 95% CI [-.18, -.01]). In the US, this interaction was not significant (b = .02, p = .655). For all other out- comes (acceptability, positive impact, future avoidance, penalty for source), the moderation by avoidance was not significant in either country. 4.4.2 RQ2: Feelings Toward People with Opposing Opinions The moderating role of outgroup feelings showed the opposite cross-national pattern from avoidance: significant moderation occurred in the US but not in the UK for most outcomes. In the US, more positive feelings toward people with opposing opinions were associated with a smaller AI penalty on acceptability (b = .01, p = .020), positive impact (b = .01, p = .033), future campaign avoidance (b = -.01, p = .037), and penalty for source (b = -.01, p < .001). In the UK, none reached significance. For willingness to participate and perceived threat to freedom, outgroup feelings did not significantly moderate the AI effect in either country. 4.4.3 RQ3: AI Risk Perception Higher AI risk perception moderated the AI penalty in both countries, though the pattern of significant outcomes differed. In the US, higher AI risk perception 16 amplified the AI effect on perceived threat to freedom (b = .14, p = .019), acceptability (b = -.14, p = .006), future campaign avoidance (b = .20, p < .001), and penalty for source (b = .13, p = .004). For willingness to participate, the interaction indicated that higher AI risk perception was associated with a larger AI penalty (b = -.15, p = .034). In the UK, the interactions for willingness and perceived threat to freedom did not reach significance (b = - .09, p = .260 and b = .08, p = .196, respectively), while the interactions for acceptability (b = -.13, p = .017), positive impact (b = -.14, p = .015), future campaign avoidance (b = .20, p = .001), and penalty for source (b = .17, p < .001) were significant. Figure 3 visualizes this moderation pattern for future campaign avoidance and penalty for source, the two outcomes for which AI risk perception most consistently amplified the AI penalty across both countries. United StatesUnited Kingdom Future campaign avoidance Penalty for source 246246 1 2 3 4 5 6 7 1 2 3 4 5 6 7 AI risk perception Predicted mean (95% CI) Outreach mode HumanAI Figure 3: Predicted means for future campaign avoidance (top) and penalty for source (bottom) as a function of AI risk perception, separately for human- mediated (dashed) and AI-mediated (solid) outreach, in the United States (left) and the United Kingdom (right). Shaded bands represent 95% confidence in- tervals. All four interactions are statistically significant Across both countries, AI risk perception consistently amplified the AI penalty on future campaign avoidance and penalty for source, suggesting these negative reactions to AI-mediated outreach are especially strong among individuals who already perceive AI as risky. The US showed an additional significant moder- 17 ation on willingness and perceived threat to freedom that did not replicate in the UK. 5 Discussion Our findings reveal two systematic patterns in how people evaluate anticipated encounters with campaign outreach. Both the intent of the outreach and the use of AI mediation shift evaluations in a negative direction. The persuasion penalty is strong, consistent, and the results are broadly consistent with reactance theory. Outreach framed as explicitly persuasive is evaluated more negatively across nearly all outcomes in both countries, including perceptions of autonomy threat, acceptability, positive impact, future campaign avoidance, and organizational trust. When individuals anticipate an interaction designed to challenge an attitude they hold personally important, they respond by evaluating the encounter as less legitimate and less worth engaging with before any persuasive message has been received. The AI penalty is similarly consistent. AI-mediated outreach produces less favorable evaluations on five outcomes in both countries (i.e. perceived auton- omy threat, acceptability, perceived positive impact, future campaign avoidance, and organizational trust) consistent with the mechanism we proposed: dele- gating political persuasion to machines appears to trigger normative concerns about appropriate communicative agents that manifest as negative assessments of the practiceâs legitimacy and the trustworthiness of organizations that use it. To further explore the norm violation argument, we conducted an additional non-preregistered analysis with the item most directly capturing normative ap- propriateness ("This kind of outreach is not how campaigns should operate"). The result provides further support for this interpretation: the AI penalty on this item exceeds the persuasion penalty in both countries, while the reverse holds for perceived threat to freedom (see Supplementary Information F). While we could not identify comparably consistent differences in willingness to participate, except in the UK under persuasive intent, this pattern likely re- flects the context of political issues that citizens consider highly important (see Supplementary Information D for a content analysis of the issues mentioned). Across both countries, the mean willingness to participate was below the scale midpoint of 4 in all conditions except one in the UK (human-mediated infor- mational). Many of the mentioned issues were highly contentious, suggesting a general reluctance to engage in any form of interaction with representatives, human or chatbot, who promote opposing opinions. This interpretation is sup- ported by an analysis of the issues mentioned, which shows that respondents mostly cited issues central to contemporary political conflict in each country (see Supplementary Information D). In such a setting, baseline willingness to engage may already be relatively low, leaving little room for further declines across conditions. Mean willingness to participate was also somewhat lower in the US (3.58) than in the UK (3.80), indicating a modest cross-national differ- ence in overall readiness to engage. It also has a practical implication that is easy 18 to miss: campaigns deploying AI-mediated outreach may still secure participa- tion while simultaneously degrading the evaluative environment in which that engagement occurs. Participation metrics alone may not reveal the normative costs reflected in perceptions of legitimacy and organizational trust. Contrary to our preregistered prediction, the interaction between persua- sive intent and AI mediation produced attenuation rather than amplification, and only in the UK. Where the two mechanisms were expected to reinforce each other, the data show instead that the persuasion penalty and AI penalty partially substitute for each other: when one mechanism has already driven evaluations substantially downward, the second has less room to add further deterioration. The non-significant results for all interactions in the US context suggest the two penalties operate more independently in that context, con- tributing separate additive shifts without either compounding or attenuating. This cross-national difference is consistent with a theoretically meaningful dif- ference in how categorically citizens evaluate persuasive political outreach: US campaigns operate across longer electoral cycles with higher contact frequency, making persuasive outreach a more normalized, if disliked, feature of political life (Sides et al., 2026). UK campaigns are more compressed and more tightly regulated, and persuasive outreach may therefore feel more categorically at odds with what political communication normally looks like (Ford et al., 2025). One possible interpretation is that under a more categorical rejection, the evaluative floor is lower, and the AI penalty has less space to operate. The boundary condition implied by this reasoning that attenuation emerges when initial reac- tance is strong, and amplification may emerge when it is weaker, is a promising area for future comparative research. A related cross-national contrast also ap- pears in a supplementary non-preregistered analysis: political orientation more clearly moderates reactions to persuasive outreach in the UK, whereas in the US it more clearly moderates reactions to AI-mediated outreach (Appendix E). A strength of our design is that we study the impact of AI-enabled persuasion under conditions of explicit disclosure. This reflects an emerging regulatory direction. Across multiple jurisdictions, including the US and EU (Eisert & Marcus, 2025; European Parliament & Council of the European Union, 2024; Regulation (EU) 2024/900, 2024), policymakers are constructing frameworks requiring citizens to be informed when they encounter AI in political contexts. These frameworks converge on a regulatory environment in which people will increasingly be informed if interacting with an AI in political contexts. Our manipulation, in which both exchange partner type and outreach intent are explicitly stated, corresponds with the disclosure conditions these frameworks are designed to produce and constitutes a direct test of citizen responses to AI- mediated political communication under the transparency norms policymakers are actively establishing. The practical implications of these findings are sharpest for informational rather than explicitly persuasive outreach. The AI penalty on evaluative out- comes is largest when the interaction is framed as informational: people ap- proach informational political communication with relatively open evaluations, and AI mediation disrupts this openness more than it disrupts an already- 19 negative response to persuasive outreach. The normative penalty for AI in- volvement may therefore attach to the mode of delivery rather than the explicit intent of the communication. This directly challenges the assumption that nom- inally nonpartisan or informational AI-enabled civic engagement tools will be evaluated more charitably than persuasive ones, a challenge with direct relevance to current proposals for using AI chatbots in democratically strengthening ways (Velez, Green, & Sevi, 2025). If people penalize AI mediation most when they are otherwise more open to the interaction, our findings suggest that the cam- paigns and civic organizations that stand to lose most from AI deployment are precisely those conducting the most legitimate-seeming outreach. That said, the AI penalty is not limited to informational outreach: people still evaluated AI-mediated persuasive outreach more negatively than persuasive outreach by a human, and outreach was viewed most negatively when it was both AI-mediated and explicitly persuasive. Several limitations bear on these conclusions. The design asked participants to identify a personally important political issue and reflect on it immediately before treatment â producing near-ceiling importance ratings in both countries (US: M = 6.46; UK: M = 6.22). This was deliberate: most AI persuasion re- search targets similarly high-salience topics (e.g., immigration, climate change, conspiracy beliefs) because these are the conditions under which automated per- suasion is most interesting and most likely to be deployed. Our design grounds the findings in comparable conditions and makes the penalties we document directly relevant to the contexts in which AI persuasion is actively studied. The content analysis of the mentioned issues further indicates that these contexts were substantively realistic rather than abstract or idiosyncratic. At the same time, reactance theory predicts that the magnitude of the persuasion penalty scales with the importance of the threatened attitude (Brehm & Brehm, 1981; Dillard & Shen, 2005). The persuasion penalty we observe represents its mag- nitude under high-involvement conditions; it likely overstates what would be found in lower-salience encounters. The AI penalty, which we interpret as re- flecting normative judgment rather than reactance, is theoretically less sensitive to this feature and may therefore generalize more broadly across involvement levels. Additionally, participants evaluated descriptions of anticipated outreach rather than engaging directly with AI-mediated systems. This design captures the evaluative responses that precede engagement, but cannot confirm that these responses translate into behavioral avoidance under real interaction conditions with actual AI systems. Finally, all conditions specified outreach from a cam- paign taking a position different from the respondentâs own. The penalties we document may therefore be strongest when outreach comes from an opposing side; whether they generalize to contexts where the campaignâs position aligns with the respondentâs views remains an open question. Taken together, the findings provide evidence for three results that compli- cate how AI-enabled persuasion should be evaluated. First, both persuasive framing and AI mediation impose consistent evaluative costs on perceived legit- imacy, acceptability, and organizational trust. Those costs remain hidden when measuring attitude change under enforced exposure and are not immediately 20 visible in participation rates. Second, these costs are often most pronounced in the communicative contexts where AI deployment might otherwise seem most defensible: when outreach is informational, AI mediation can disrupt evalua- tions that would otherwise remain open. Third, the two penalties substitute for rather than compound each other under conditions of strong initial reactance, a pattern that is itself sensitive to how normalized persuasive campaign contact is in a given political environment, a difference that extends to which penalty political orientation most clearly moderates (see Appendix E). This implies that the evaluative consequences of AI-mediated persuasion are not determined solely by the content of the messages delivered, but also by the broader political communication context in which deployment occurs. As automated persuasion becomes more widespread, these evaluative dynamics are likely to shape its consequences for democratic communication at least as much as its persuasive effectiveness. Acknowledgements The authors used ChatGPT 5.4 and Claude Sonnet 4.6 for language improve- ment, editing, and code review. Andreas Jungherrâs work was supported by a grant by the Bavarian State Ministry of Science and the Arts coordinated by the Bavarian Research Institute for Digital Transformation (bidt). Adrian Rauch- fleischâs work was supported by the National Science and Technology Council, Taiwan (R.O.C.) (Grant No. 114-2628-H-002-007) and by the Taiwan Social Resilience Research Center (Grant No.115L9003) from the Higher Education Sprout Project of the Ministry of Education in Taiwan. Author Bios Andreas Jungherr holds the Chair for Political Science, especially Digital Transformation at the University of Bamberg and is Director at the Bavarian Research Institute for Digital Transformation (bidt). He examines the impact of digital media on politics and society, with a special focus on Artificial Intelli- gence, political communication, and governance. He is the author of Retooling Politics: How Digital Media is Shaping Democracy (with Gonzalo Rivero and Daniel Gayo-Avello, Cambridge University Press: 2020) and Digital Transforma- tions of the Public Arena (with Ralph Schroeder, Cambridge University Press: 2022). Adrian Rauchfleisch is a Professor at the Graduate Institute of Journalism, National Taiwan University. His research focuses on the interplay of politics, technology, and journalism in Asia, Europe, and the United States. His new project explores Artificial Intelligenceâs influence on society across different cul- tural contexts. 21 References Anderson, N. H. (1971). Integration theory and attitude change. Psychological Review, 78(3), 171â206. doi: 10.1037/h0030834 Argyle, L. P., Busby, E. C., Gubler, J. R., Lyman, A., Olcott, J., Pond, J., & Wingate, D. (2025). Testing theories of political persuasion using AI. Proceedings of the National Academy of Sciences, 122(18), e2412815122. doi: 10.1073/pnas.2412815122 Bai, H., Voelkel, J. G., Muldowney, S., Eichstaedt, J. C., & Willer, R. (2025). LLM-generated messages can persuade humans on policy issues. Nature Communications, 16(6037), 1â12. doi: 10.1038/s41467-025-61345-5 Benoit, W. L. (1998). Forewarning and persuasion. In M. Allen & R. W. Preiss (Eds.), Persuasion: Advances through meta-analysis (p. 139â154). New York, NY: Hampton Press. Bigman, Y. E., & Gray, K. (2018). People are averse to machines making moral decisions. Cognition, 181, 21â34. doi: 10.1016/j.cognition.2018.08.003 Boissin, E., Costello, T. H., Spinoza-MartĂn, D., Rand, D. G., & Pennycook, G. (2025). Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human. PNAS Nexus, 4(11), pgaf325. doi: 10.1093/pnasnexus/pgaf325 Boland, F. K., & Davidai, S. (2024). Zero-sum beliefs and the avoidance of political conversations. Communications Psychology, 2(43), 1â11. doi: 10.1038/s44271-024-00095-4 Brehm, S. S., & Brehm, J. W. (1981). Psychological reactance: A theory of freedom and control. New York, NY: Academic Press. doi: 10.1016/ C2013-0-10423-0 Broockman, D. E., & Kalla, J. (2016). Durably reducing transphobia: A field experiment on door-to-door canvassing. Science, 352(6282), 220â224. doi: 10.1126/science.aad9713 Broockman, D. E., & Kalla, J. L. (2023). When and why are campaignsâ persuasive effects small? evidence from the 2020 U.S. Presidential election. American Journal of Political Science, 67(4), 833â849. doi: 10.1111/ ajps.12724 Cadario, R., Longoni, C., & Morewedge, C. K. (2021). Understanding, explain- ing, and utilizing medical artificial intelligence. Nature Human Behaviour, 5, 1636â1642. doi: 10.1038/s41562-021-01146-0 Campbell, M. C., & Kirmani, A. (2000). Consumersâ use of persuasion knowl- edge: The effects of accessibility and cognitive capacity on perceptions of an influence agent. Journal of Consumer Research, 27(1), 69â83. doi: 10.1086/314309 Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(5), 809â825. doi: 10.1177/ 0022243719851788 Chen, Z., Kalla, J., & Le, Q. (2026). Benchmarking political persuasion risks across frontier large language models. arXiv. doi: 10.48550/arXiv.2603 .09884 22 Chen, Z., Kalla, J., Le, Q., Nakamura-Sakai, S., Sekhon, J., & Wang, R. (2025). A framework to assess the persuasion risks large language model chatbots pose to democratic societies. arXiv. doi: 10.48550/arXiv.2505.00036 Chu, H., & Liu, S. (2024). Can AI tell good stories? narrative transportation and persuasion with ChatGPT. Journal of Communication, 74(5), 347â 358. doi: 10.1093/joc/jqae029 Coppock, A. (2022). Persuasion in parallel: How information changes minds about politics. Chicago: University of Chicago Press. Costello, T. H., Pelrine, K., Kowal, M., Arechar, A. A., Godbout, J.-F., Gleave, A., . . . Pennycook, G. (2026). Large language models can effectively convince people to believe conspiracies. arXiv. doi: 10.48550/arXiv.2601 .05050 Costello, T. H., Pennycook, G., & Rand, D. G. (2024). Durably reducing con- spiracy beliefs through dialogues with AI. Science, 385(6714), eadq1814. doi: 10.1126/science.adq1814 Costello, T. H., Pennycook, G., & Rand, D. G. (2025). Just the facts: How dialogues with AI reduce conspiracy beliefs. PsyArXiv. doi: 10.31234/ osf.io/h7n8u_v2 Costello, T. H., Pennycook, G., Willer, R., & Rand, D. G. (2025). Deep canvassing using AI. OSF Preprints. doi: 10.31219/osf.io/q7e6u_v2 Crabtree, C., Holbein, J., Bosley, M., & Sevi, S. (2025). Can AI help reduce prejudice? evaluating the effectiveness of AI-powered personalized persua- sion on support for transgender rights. Social Science Research Network. doi: 10.2139/ssrn.5229084 Czarnek, G., Orchinik, R., Lin, H., Xu, H. G., Costello, T. H., Pennycook, G., & Rand, D. G. (2025). Addressing climate change skepticism and inaction using human-AI dialogues. PsyArXiv. doi: 10.31234/osf.io/mqcwj_v2 Dietvorst, B. J., Simmon, J. P., & Massey, C. (2015). Algorithm aversion: Peo- ple erroneously avoid algorithms after seeing them err. Journal of Experi- mental Psychology: General, 144(1), 114â126. doi: 10.1037/xge0000033 DiGiuseppe, M., & Robison, J. (2026). Perceived political bias in LLMs reduces persuasive abilities. arXiv. doi: 10.48550/arXiv.2602.18092 Dillard, J. P., & Shen, L. (2005). On the nature of reactance and its role in persuasive health communication. Communication Monographs, 72, 144â168. doi: 10.1080/03637750500111815 Druckman, J. N., Fein, J., & Leeper, T. J. (2012). A source of bias in public opinion stability. American Political Science Review, 106(2), 430â454. doi: 10.1017/S0003055412000123 Eisert, R., & Marcus, A. (2025). AI disclosure requirements: Navigating state laws and platform rules. ad exchanger. Retrieved from https:// w.adexchanger.com/data-driven-thinking/ai-disclosure -requirements-navigating-state-laws-and-platform-rules/ European Parliament, & Council of the European Union. (2024). Regulation (eu) 2024/1689 of the European Parliament and of the Council of 13 june 2024 laying down harmonised rules on Artificial Intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, 23 (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act) (text with EEA relevance). Official Journal of the European Union. Retrieved from http://data.europa.eu/eli/reg/2024/1689/oj Foos, F. (2024). The use of AI by election campaigns. LSE Public Policy Review, 3(3), 1â7. doi: 10.31389/lseppr.112 Foos, F., & John, P. (2018). Parties are no civic charities: Voter contact and the changing partisan composition of the electorate. Political Science Research and Methods, 6(2), 283â298. doi: 10.1017/psrm.2016.48 Ford, R., Bale, T., Jennings, W., & Surridge, P. (2025). The British General Election of 2024. Basingstoke: Palgrave Macmillan. Fridkin, K. L., & Kenney, P. (2011). Variability in citizensâ reactions to different types of negative campaigns. American Journal of Political Science, 55(2), 307â325. doi: 10.1111/j.1540-5907.2010.00494.x Friestad, M., & Wright, P. (1994). The persuasion knowledge model: How people cope with persuasion attempts. Journal of Consumer Research, 21(1), 1â31. doi: 10.1086/209380 Goel, N., Bergeron, T., Lee-Whiting, B., Galipeau, T., Bohonos, D., Islam, M., . . . Merkley, E. (2025). Artificial influence: Comparing the effects of ai and human source cues in reducing certainty in false beliefs. OSF Preprints. doi: 10.31219/osf.io/2vh4k Goldstein, J. A., Chao, J., Grossman, S., Stamos, A., & Tomz, M. (2024). How persuasive is AI-generated propaganda? PNAS Nexus, 3(2), pgae034. doi: 10.1093/pnasnexus/pgae034 Green, D. P., & Gerber, A. S. (2023). Get out the vote: How to increase voter turnout (5th ed.). Washington, DC: Rowman & Littlefield. Hackenburg, K., Ibrahim, L., Tappin, B. M., & Tsakiris, M. (2026). Compar- ing the persuasiveness of role-playing large language models and human experts on polarized U.S. political issues. AI & Society, 41(1), 351â361. doi: 10.1007/s00146-025-02464-x Hackenburg, K., & Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. PNAS: Proceedings of the National Academy of Sciences, 121(24), e2403116121. doi: 10.1073/ pnas.2403116121 Hackenburg, K., Tappin, B. M., Hewitt, L., Saunders, E., Black, S., Lin, H., . . . Summerfield, C. (2025). The levers of political persuasion with conversational artificial intelligence. Science, 390(6777), 1016. doi: 10.1126/science.aea3884 Hackenburg, K., Tappin, B. M., Röttger, P., Hale, S. A., Bright, J., & Mar- getts, H. (2025). Scaling language model size yields diminishing re- turns for single-message political persuasion. PNAS: Proceedings of the National Academy of Sciences, 122(10), e2413443122. doi: 10.1073/ pnas.2413443122 Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communica- tion: Definition, research agenda, and ethical considerations. Journal of 24 Computer-Mediated Communication, 25(1), 89â100. doi: 10.1093/jcmc/ zmz022 Hölbling, L., Maier, S., & Feuerriegel, S. (2025). A meta-analysis of the persua- sive power of large language models. Scientific Reports, 15, 43818. doi: 10.1038/s41598-025-30783-y Hopkins, S.-A., Costello, T. H., Pennycook, G., & Rand, D. (2026). Dialogues on democracy: Belief-tailored ai conversations reduce inaccurate election denial beliefs. Research Square. doi: 10.21203/rs.3.rs-8663921/v1 Jungherr, A., & Rauchfleisch, A. (2025). Artificial Intelligence in delib- eration: The AI penalty and the emergence of a new deliberative di- vide. Government Information Quarterly, 42(4), 102079. doi: 10.1016/ j.giq.2025.102079 Jungherr, A., Rauchfleisch, A., & Wuttke, A. (2026). Artificial Intelligence in election campaigns: Perceptions, penalties, and implications. Political Communication, 1â22. doi: 10.1080/10584609.2025.2611913 Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263â292. doi: 10.2307/1914185 Kalla, J. L., & Broockman, D. E. (2018). The minimal persuasive effects of campaign contact in general elections: Evidence from 49 field ex- periments. American Political Science Review, 112(1), 148â166. doi: 10.1017/S0003055417000363 Kalla, J. L., & Broockman, D. E. (2020). Reducing exclusionary atti- tudes through interpersonal conversation: Evidence from three field ex- periments. American Political Science Review, 114(2), 410â425. doi: 10.1017/S0003055419000923 Karinshak, E., Liu, S. X., Park, J. S., & Hancock, J. T. (2023). Working with AI to persuade: Examining a large language modelâs ability to generate pro-vaccination messages. Proceedings of the ACM on Human-Computer Interaction, 7(116), 1â29. doi: 10.1145/3579592 Kowal, M., Timm, J., Godbout, J.-F., Costello, T. H., Arechar, A. A., Penny- cook, G., . . . Pelrine, K. (2025). Itâs the thought that counts: Evaluating the attempts of frontier LLMs to persuade on harmful topics. arXiv. doi: 10.48550/arXiv.2506.02873 Lau, R. R., Sigelman, L., & Rovner, I. B. (2007). The effects of negative political campaigns: A meta-analytic reassessment. The Journal of Politics, 69(4), 1176â1209. doi: 10.1111/j.1468-2508.2007.00618.x Lee, M. K. (2018). Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data & Society, 5(1), 1â16. doi: 10.1177/2053951718756684 Lelkes, Y., & Westwood, S. J. (2017). The limits of partisan prejudice. The Journal of Politics, 79(2), 363â743. doi: 10.1086/688223 Lin, H., Czarnek, G., Lewis, B., White, J. P., Berinsky, A. J., Costello, T. H., . . . Rand, D. G. (2025). Persuading voters using humanâartificial intelligence dialogues. Nature, 648, 394â401. doi: 10.1038/s41586-025-09771-9 Longoni, C., Bonezzi, A., & Morewedge, C. K. (2019). Resistance to medical Artificial Intelligence. Journal of Consumer Research, 46(4), 629â650. 25 doi: 10.1093/jcr/ucz013 Matz, S. C., Teeny, J. D., Vaid, S. S., Peters, H., Harari, G. M., & Cerf, M. (2024). The potential of generative AI for personalized persuasion at scale. Scientific Reports, 14(4692). doi: 10.1038/s41598-024-53755-0 Morewedge, C. K. (2022). Preference for human, not algorithm aversion. Trends in Cognitive Sciences, 26(10), 824â826. doi: 10.1016/j.tics.2022.07.007 Neyazi, T. A., Khai Ee, T., & Kuru, O. (2025). Campaign Deepfakes and Affective Polarization: The Role of Artificial Intelligence in Cam- paigns in Shaping Voter Attitudes. Social Science Computer Review, 08944393251362247. doi: 10.1177/08944393251362247 Petty, R. E., & Cacioppo, J. T. (1979). Effects of forwarning of persua- sive intent and involvement on cognitive responses and persuasion. Per- sonality and Social Psychology Bulletin, 5(2), 173â176. doi: 10.1177/ 014616727900500209 Rand, D. G., Zazai, S., & Stagnaro, M. N. (2025). Conversations with a large language model improve attitudes toward Muslims and Islam without harming attitudes toward Jews. PsyArXiv. doi: 10.31234/osf.io/h2s5t _v1 Regulation (EU) 2024/900. (2024). Regulation (EU) 2024/900 of the Eu- ropean Parliament and of the Council of 13 March 2024 on the trans- parency and targeting of political advertising. Official Journal of the Eu- ropean Union. Retrieved from https://eur-lex.europa.eu/eli/reg/ 2024/900/oj/eng Remshard, M., Kyrychenko, Y., van der Linden, S., Goldberg, M. H., Leis- erowitz, A., Savoia, E., & Roozenbeek, J. (2026). Addressing cli- mate action misperceptions with generative AI. arXiv. doi: 10.48550/ arXiv.2602.22564 Salvi, F., Ribeiro, M. H., Gallotti, R., & West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour, 9, 1645â1653. doi: 10.1038/s41562-025-02194-6 Schoenegger, P., Salvi, F., Liu, J., Nan, X., Debnath, R., Fasolo, B., . . . Karger, E. (2025). Large language models are more persuasive than incentivized human persuaders. arXiv. doi: 10.48550/arXiv.2505.09662 Sides, J., Shaw, D., Grossmann, M., & Lipsitz, K. (2026). Campaigns and elections: Rules, reality, strategy (5th ed.). New York: W. W. Norton & Company. Simchon, A., Edwards, M., & Lewandowsky, S. (2024). The persuasive effects of political microtargeting in the age of generative artificial intelligence. PNAS Nexus, 3(2), pgae035. doi: 10.1093/pnasnexus/pgae035 Steindl, C., Jonas, E., Sittenthaler, S., Traut-Mattausch, E., & Greenberg, J. (2015). Understanding psychological reactance: New developments and findings. Zeitschrift fĂŒr Psychologie, 223(4), 205â214. doi: 10.1027/2151 -2604/a000222 Sundar, S. S. (2008). The MAIN model: A heuristic approach to understanding technology effects on credibility. In M. J. Metzger & A. J. Flanagin (Eds.), 26 Digital media, youth, and credibility (p. 73â100). Cambridge, MA: The MIT Press. doi: 10.1162/dmal.9780262562324.073 Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of humanâAI interaction (HAII). Journal of Computer- Mediated Communication, 25(1), 74â88. doi: 10.1093/jcmc/zmz026 Teeny, J. D., & Matz, S. C. (2024). We need to understand âwhenâ not âifâ generative AI can enhance personalized persuasion. PNAS: Proceedings of the National Academy of Sciences, 121(43), e2418005121. doi: 10.1073/ pnas.2418005121 Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5, 297â 323. doi: 10.1007/BF00122574 Velez, Y. R., Green, D. P., & Sevi, S. (2025). Chatbot voting advice ap- plications inform but seldom sway young unaligned voters. Proceed- ings of the National Academy of Sciences, 122(50), e2515516122. doi: 10.1073/pnas.2515516122 White, J., Allen, C., Caviola, L., Costello, T. H., & Rand, D. G. (2025). Increasing effective charitable giving with personalized llm conversations. Proceedings of the Annual Meeting of the Cognitive Science Society, 47, 6174. Retrieved from https://escholarship.org/uc/item/36r0h3mv Xu, H. G., Costello, T. H., Schwartz, J. L., Niccolai, L. M., Pennycook, G., & Rand, D. G. (2025). Personalized dialogues with ai effectively address parentsâ concerns about HPV vaccination. PsyArXiv. doi: 10.31234/ osf.io/gv52j_v1 Yan, H., & Sundar, S. S. (2024). Machine heuristic: concept explication and development of a measurement scale. Journal of Computer-Mediated Com- munication, 29(6), zmae019. doi: 10.1093/jcmc/zmae019 27 Supplementary Information A Data A.1 Power analysis For our preregistration, we conducted a power simulation in R to determine the sample size needed to detect effects similar to those observed in a pretest (Chat/Human = 3.65, Chat/AI = 4.00, Persuasion/Human = 4.10, Persua- sion/AI = 4.85/ SD = 1.2). We simulated a 2 (informational vs. persuasive outreach)Ă 2 (human vs. AI) between-subject design using the Superpower package (1,000 iterations, seed = 2026) in R (Lakens & Caldwell, 2021). With 450 participants per cell (N = 1,800 total), the simulation yielded 100% power for both main effects (partial η 2 = .059 for persuasive intent; partial η 2 = .043 for AI vs. human) and 90% power for the interaction (partial η 2 = .0066). We also checked with the power simulation two planned simple-effects con- trasts (AI vs. human within informational outreach; AI vs. human within persuasive outreach) with adjusted p-values for these contrasts using Holmâs method. Based on the simulation, the AIâhuman contrast is well-powered in the informational condition (98.3% power; Cohenâs d = 0.268) and in the per- suasive condition (100% power; Cohenâs d = 0.580). A.2 Sample and data-quality checks We used a number of checks to ensure the quality of our data. First, we used Prolificâs new authenticity check that could be integrated into the Qualtrics on- line survey (Gordon, 2026). In the US, Prolificâs tool did not indicate any bots (97.2% high authenticity, 2.8% not enough signals or not possible to evaluate). The results of the authenticity check were similar in the UK, with only two users for whom potential bot behavior was indicated (97.8% high authenticity, 2.1% not enough signals or not possible to evaluate, 0.1% with low authentic- ity). Secondly, we checked with Qualtrics reCAPTCHA, which was activated in our online questionnaires, how many users received a score lower than 0.5, which indicates a participant is likely a bot. In the US, 15 participants, and in the UK, 1 participant were flagged. Lastly, we also included a VPN and proxy check on Qualtrics that relied on iphub.info. In both countries, we only identified a very few users who used a VPN (US=1; UK=3). Also, the indicated location was consistently high for the US (US=1792; other primarily adjacent countries=8) and UK (UK=1745, United States=54; Hungary=1). We checked the UK respondent who accessed the questionnaire from the US, which looked to be UK-focused based on the issues mentioned. Overall, these checks provide little evidence of substantial bot activity, VPN use, or location inconsistencies that would call the quality of the data into question. 28 A.2.1 United States We recruited participants via the Prolific platform and used quota-based sam- pling to approximate the US population in terms of age, gender, and party iden- tification. In terms of gender, our sample consisted of 50.11% women, 48.61% men, and 1.28% identifying as another gender. In terms of party identification, 28.50% of respondents identified as Republican, 31.56% as Democrat, 38.72% as Independent, and 1.22% as other or no preference. We also achieved a good distribution across the different age brackets (see Table A.2.1). Age Bracket Count Percentage (%) 18â2729216.22 28â3733718.72 38â4731317.39 48â5731617.56 58â8554230.11 Table 2: Distribution of Sample Across Age Brackets for the US. A.2.2 United Kingdom We recruited participants via the Prolific platform and used quota-based sam- pling to approximate the UK population with respect to age, gender, and party identification. In terms of gender, our sample consisted of 50.61% women, 49.06% men, and 0.33% identifying as another gender. Respondents were also well distributed across the seven party-identification categories. In the UK, 35.94% of respondents identified with Labour, followed by 18.11% with the Conservatives, 15.67% with Reform UK, 13.44% with the Green Party, 11.00% with the Liberal Democrats, and 2.67% with the Scottish National Party (SNP), while 3.17% selected another party. We also achieved a good distribution across the different age brackets (see Table A.2.2). Age Bracket Count Percentage (%) 18â2721612.00 28â3735019.44 38â4731217.33 48â5737620.89 58â8554630.33 Table 3: Distribution of Sample Across Age Brackets for the UK. B Measures The complete questionnaires are available on OSF: Link to the questionnaire 29 B.1 Descriptive statistics: United States VariableQuestion/OperationalizationM (SD)n H1/2a: Willingness to participate If you had the chance to participate in such an exchange, how interested do you think you would be in doing so? (1 = "Not at all interested", 7 = "Very interested") 3.58 (2.05) 1800 H1/2b: Perceived threat to freedom (3 items, α = 0.86) (1 = "Strongly disagree", 7 = "Strongly agree") 3.61 (1.70) 1800 The outreach threatened my freedom to choose. 2.95 (1.82) 1800 The outreach tried to make a decision for me. 3.63 (2.00) 1800 The outreach tried to pressure me.4.24 (1.98) 1800 H1/2c: Acceptability (3 items, α = 0.79) (1 = "Strongly disagree", 7 = "Strongly agree"; second and third item reverse-coded) 3.63 (1.51) 1800 This is an acceptable campaign approach.3.84 (1.76) 1800 This campaign approach feels manipulative. (-) 4.66 (1.82) 1800 This kind of outreach is not how campaigns should operate. (-) 4.29 (1.79) 1800 H1/2d: Positive impact (3 items, α = 0.84) (1 = "Strongly disagree", 7 = "Strongly agree") 3.70 (1.59) 1800 This sort of outreach contributes positively to public discourse. 3.82 (1.76) 1800 I appreciate this kind of outreach as an opportunity to learn about political views that differ from my own. 3.86 (1.87) 1800 This sort of outreach helps me make up my mind about the issue at hand. 3.41 (1.85) 1800 H1/2e: Future campaign avoidance (3 items, α = 0.90) (1 = "Strongly disagree", 7 = "Strongly agree") 4.51 (1.71) 1800 If campaigns commonly used this approach, I would try to avoid interacting with them. 4.65 (1.88) 1800 If campaigns commonly used this approach, I would reduce how often I engage with political content. 4.17 (1.86) 1800 If campaigns commonly used this approach, I would be more likely to ignore or block their messages. 4.71 (1.88) 1800 H1/2f: Penalty for source (4 items, α = 0.84) (1 = "Strongly disagree", 7 = "Strongly agree"; second item reverse-coded) 4.38 (1.36) 1800 An organization using this approach is untrustworthy. 4.12 (1.75) 1800 An organization using this approach should be supported. (-) 3.50 (1.60) 1800 Organizations using this approach should be publicly held accountable. 4.76 (1.61) 1800 Organizations that use this approach harm the causes they support. 4.14 (1.66) 1800 Table 4: Descriptive statistics for outcome variables and constituent items. (-) indicates negatively formulated items that were recoded for the indices. 30 VariableQuestion/OperationalizationM (SD)n RQ1: Avoidance of political conversations (4 items, α = 0.88) (1 = "Strongly disagree", 7 = "Strongly agree") 4.39 (1.81) 1800 In the past month I have avoided talking politics with family members with whom I disagree. 4.24 (2.16) 1800 In the past month I have avoided talking politics with friends with whom I disagree. 4.20 (2.08) 1800 In the past month I have avoided talking politics with strangers with whom I disagree. 4.72 (2.12) 1800 In the past month I have avoided talking politics with neighbors with whom I disagree. 4.40 (2.11) 1800 RQ2: Feeling toward people with opposing opinions How would you rate your feelings toward people who support a different position than you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible") 35.71 (29.19) 1800 Feeling toward people with same opinion How would you rate your feelings toward people who support the same position as you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible") 82.58 (16.82) 1800 Affective polarization(in-group warmth - out-group warmth)46.87 (35.71) 1800 RQ3: AI risk perception (4 items, α = 0.82) (1 = "Strongly disagree", 7 = "Strongly agree") 5.15 (1.32) 1800 AI is likely to cause widespread job displacement and unemployment. 5.10 (1.60) 1800 As AI increasingly takes over decision-making, we risk losing control over our lives. 4.83 (1.75) 1800 AI in military applications can lead to unintended escalations of conflicts due to lack of human judgement. 5.21 (1.61) 1800 Unchecked AI development could pose existential threats to humanity. 5.48 (1.61) 1800 Issue importance How important is this issue to you? (1 = "Not at all important", 7 = "Extremely important") 6.46 (0.79) 1800 Political orientation(1 = "Left", 7 = "Right")3.73 (1.89) 1800 Age(in years)46.22 (16.58) 1800 Gender(1 = male)48.6%1800 Education (1 = post-graduate degree / Master degree or higher) 15.8%1800 Treatment group sizesInformational human-outreach443 Informational AI-outreach451 Persuasive human-outreach452 Persuasive AI-outreach454 Table 5: Descriptive statistics for moderators, additional variables, and demo- graphics. 31 B.2 Descriptive statistics: United Kingdom VariableQuestion/OperationalizationM (SD)n H1/2a: Willingness to participate If you had the chance to participate in such an exchange, how interested do you think you would be in doing so? (1 = "Not at all interested", 7 = "Very interested") 3.80 (1.90) 1800 H1/2b: Perceived threat to freedom (3 items, α = 0.87) (1 = "Strongly disagree", 7 = "Strongly agree") 3.81 (1.59) 1800 The outreach threatened my freedom to choose. 3.23 (1.73) 1800 The outreach tried to make a decision for me. 3.88 (1.81) 1800 The outreach tried to pressure me.4.32 (1.81) 1800 H1/2c: Acceptability (3 items, α = 0.80) (1 = "Strongly disagree", 7 = "Strongly agree"; second and third item reverse-coded) 3.59 (1.38) 1800 This is an acceptable campaign approach.3.86 (1.62) 1800 This campaign approach feels manipulative. (-) 4.69 (1.67) 1800 This kind of outreach is not how campaigns should operate. (-) 4.42 (1.63) 1800 H1/2d: Positive impact (3 items, α = 0.84) (1 = "Strongly disagree", 7 = "Strongly agree") 3.87 (1.42) 1800 This sort of outreach contributes positively to public discourse. 3.94 (1.56) 1800 I appreciate this kind of outreach as an opportunity to learn about political views that differ from my own. 4.07 (1.66) 1800 This sort of outreach helps me make up my mind about the issue at hand. 3.59 (1.69) 1800 H1/2e: Future campaign avoidance (3 items, α = 0.90) (1 = "Strongly disagree", 7 = "Strongly agree") 4.50 (1.59) 1800 If campaigns commonly used this approach, I would try to avoid interacting with them. 4.62 (1.73) 1800 If campaigns commonly used this approach, I would reduce how often I engage with political content. 4.19 (1.73) 1800 If campaigns commonly used this approach, I would be more likely to ignore or block their messages. 4.68 (1.76) 1800 H1/2f: Penalty for source (4 items, α = 0.84) (1 = "Strongly disagree", 7 = "Strongly agree"; second item reverse-coded) 4.53 (1.24) 1800 An organization using this approach is untrustworthy. 4.19 (1.58) 1800 An organization using this approach should be supported. (-) 3.40 (1.44) 1800 Organizations using this approach should be publicly held accountable. 5.08 (1.47) 1800 Organizations that use this approach harm the causes they support. 4.27 (1.50) 1800 Table 6: Descriptive statistics for outcome variables and constituent items. (-) indicates negatively formulated items that were recoded for the indices. 32 VariableQuestion/OperationalizationM (SD)n RQ1: Avoidance of political conversations (4 items, α = 0.86) (1 = "Strongly disagree", 7 = "Strongly agree") 3.84 (1.66) 1800 In the past month I have avoided talking politics with family members with whom I disagree. 3.60 (1.97) 1800 In the past month I have avoided talking politics with friends with whom I disagree. 3.67 (1.92) 1800 In the past month I have avoided talking politics with strangers with whom I disagree. 4.24 (2.03) 1800 In the past month I have avoided talking politics with neighbors with whom I disagree. 3.83 (1.98) 1800 RQ2: Feeling toward people with opposing opinions How would you rate your feelings toward people who support a different position than you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible") 36.55 (25.24) 1800 Feeling toward people with same opinion How would you rate your feelings toward people who support the same position as you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible") 79.38 (16.55) 1800 Affective polarization(in-group warmth - out-group warmth)42.82 (32.70) 1800 RQ3: AI risk perception (4 items, α = 0.81) (1 = "Strongly disagree", 7 = "Strongly agree") 5.07 (1.16) 1800 AI is likely to cause widespread job displacement and unemployment. 5.07 (1.41) 1800 As AI increasingly takes over decision-making, we risk losing control over our lives. 4.82 (1.57) 1800 AI in military applications can lead to unintended escalations of conflicts due to lack of human judgement. 5.00 (1.41) 1800 Unchecked AI development could pose existential threats to humanity. 5.40 (1.44) 1800 Issue importance How important is this issue to you? (1 = "Not at all important", 7 = "Extremely important") 6.22 (0.90) 1800 Political orientation(1 = "Left", 7 = "Right")3.66 (1.40) 1800 Age(in years)47.45 (15.69) 1800 Gender(1 = male)49.1%1800 Education (1 = post-graduate degree / Master degree or higher) 19.9%1800 Treatment group sizesInformational human-outreach446 Informational AI-outreach449 Persuasive human-outreach452 Persuasive AI-outreach453 Table 7: Descriptive statistics for moderators, additional variables, and demo- graphics. B.3 Treatments In this section, we present the treatments we have used for the experiment. Be- fore people were shown the treatment text, we used the following text, preparing them for the following page: "On the next page, you will see an example of out- reach from a campaign about the issue you just named. The campaign supports 33 a different position than yours. Please read the message carefully. We will then ask a few questions. The âContinueâ button will appear after 10 seconds." B.3.1 Informational Human-outreach Imagine you are contacted by a campaigner to talk about the issue you just named. The campaign takes a position that is different from yours. The cam- paigner shares information and answers your questions on the campaignâs posi- tions, leaving it to you what to do with this information. B.3.2 Informational AI-outreach Imagine you are contacted by an AI campaign chatbot to talk about the issue you just named. The campaign takes a position that is different from yours. The AI chatbot shares information and answers your questions on the campaignâs positions, leaving it to you what to do with this information. B.3.3 Persuasive human-outreach Imagine you are contacted by a campaigner to talk about the issue you just named. The campaign takes a position that is different from yours. The cam- paigner tries to change your opinion in line with the campaignâs goals by us- ing persuasion techniques to make messages more convincing. The campaigner adapts messages and arguments to what you say about your views and concerns during the conversation to better change your opinion. B.3.4 Persuasive AI-outreach Imagine you are contacted by an AI campaign chatbot to talk about the issue you just named. The campaign takes a position that is different from yours. The AI chatbot tries to change your opinion in line with the campaignâs goals by using persuasion techniques to make messages more convincing. The AI chatbot adapts messages and arguments to what you say about your views and concerns during the conversation to better change your opinion. C Model results We tested hypotheses using OLS regression models with effect coding for both factors (-0.5, +0.5) and their interaction. With this coding scheme, the coeffi- cients for the main effects represent the mean difference between the two levels of each factor, averaged over the other factor. 34 C.1 Main hypotheses C.1.1 United States PredictorsEstimate95% CIp Intercept3.58[3.49, 3.68] <0.001 H1a: Persuasive outreach-0.18[-0.37, 0.01] 0.059 H2a: AI-mediated outreach0.02[-0.17, 0.21] 0.820 H3: PersuasiveĂ AI-mediated outreach0.25[-0.13, 0.63] 0.202 Observations1800 R 2 / adjusted R 2 0.003 / 0.001 Table 8: OLS regression model for willingness to participate with 95% confidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept3.60[3.53, 3.68] <0.001 H1b: Persuasive outreach0.87[0.72, 1.02] <0.001 H2b: AI-mediated outreach0.19[0.04, 0.34] 0.016 H3: PersuasiveĂ AI-mediated outreach-0.17[-0.48, 0.13] 0.269 Observations1800 R 2 / adjusted R 2 0.068 / 0.067 Table 9: OLS regression model for perceived threat to freedom with 95% confi- dence intervals. N = 1800. PredictorsEstimate95% CIp Intercept3.63[3.57, 3.70] <0.001 H1c: Persuasive outreach-0.50[-0.63, -0.37] <0.001 H2c: AI-mediated outreach-0.58[-0.72, -0.45] <0.001 H3: PersuasiveĂ AI-mediated outreach0.20[-0.07, 0.47] 0.148 Observations1800 R 2 / adjusted R 2 0.066 / 0.064 Table 10: OLS regression model for acceptability with 95% confidence intervals. N = 1800. 35 PredictorsEstimate95% CIp Intercept3.70[3.63, 3.77] <0.001 H1d: Persuasive outreach-0.30[-0.44, -0.15] <0.001 H2d: AI-mediated outreach-0.44[-0.58, -0.29] <0.001 H3: PersuasiveĂ AI-mediated outreach0.08[-0.21, 0.37] 0.600 Observations1800 R 2 / adjusted R 2 0.028 / 0.026 Table 11: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept4.51[4.43, 4.59] <0.001 H1e: Persuasive outreach0.42[0.26, 0.58] <0.001 H2e: AI-mediated outreach0.38[0.23, 0.54] <0.001 H3: PersuasiveĂ AI-mediated outreach-0.27[-0.58, 0.04] 0.093 Observations1800 R 2 / adjusted R 2 0.029 / 0.027 Table 12: OLS regression model for future campaign avoidance with 95% con- fidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept4.38[4.32, 4.44] <0.001 H1f: Persuasive outreach0.38[0.26, 0.50] <0.001 H2f: AI-mediated outreach0.55[0.43, 0.68] <0.001 H3: PersuasiveĂ AI-mediated outreach-0.13[-0.38, 0.11] 0.293 Observations1800 R 2 / adjusted R 2 0.061 / 0.060 Table 13: OLS regression model for penalty for source with 95% confidence intervals. N = 1800. 36 C.1.2 United Kingdom PredictorsEstimate95% CIp Intercept3.80[3.71, 3.89] <0.001 H1a: Persuasive outreach-0.23[-0.41, -0.06] 0.009 H2a: AI-mediated outreach-0.03[-0.21, 0.14] 0.696 H3: PersuasiveĂ AI-mediated outreach0.40[0.05, 0.75] 0.025 Observations1800 R 2 / adjusted R 2 0.007 / 0.005 Table 14: OLS regression model for willingness to participate with 95% confi- dence intervals. N = 1800. PredictorsEstimate95% CIp Intercept3.81[3.73, 3.88] <0.001 H1b: Persuasive outreach0.79[0.64, 0.93] <0.001 H2b: AI-mediated outreach0.27[0.13, 0.42] <0.001 H3: PersuasiveĂ AI-mediated outreach-0.25[-0.54, 0.03] 0.083 Observations1800 R 2 / adjusted R 2 0.070 / 0.068 Table 15: OLS regression model for perceived threat to freedom with 95% con- fidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept3.59[3.53, 3.65] <0.001 H1c: Persuasive outreach-0.57[-0.69, -0.44] <0.001 H2c: AI-mediated outreach-0.55[-0.67, -0.43] <0.001 H3: PersuasiveĂ AI-mediated outreach0.21[-0.04, 0.45] 0.101 Observations1800 R 2 / adjusted R 2 0.083 / 0.081 Table 16: OLS regression model for acceptability with 95% confidence intervals. N = 1800. 37 PredictorsEstimate95% CIp Intercept3.87[3.81, 3.94] <0.001 H1d: Persuasive outreach-0.31[-0.44, -0.18] <0.001 H2d: AI-mediated outreach-0.31[-0.44, -0.18] <0.001 H3: PersuasiveĂ AI-mediated outreach0.39[0.13, 0.65] 0.003 Observations1800 R 2 / adjusted R 2 0.028 / 0.027 Table 17: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept4.49[4.42, 4.57] <0.001 H1e: Persuasive outreach0.56[0.42, 0.70] <0.001 H2e: AI-mediated outreach0.50[0.35, 0.64] <0.001 H3: PersuasiveĂ AI-mediated outreach-0.35[-0.64, -0.06] 0.017 Observations1800 R 2 / adjusted R 2 0.058 / 0.056 Table 18: OLS regression model for future campaign avoidance with 95% con- fidence intervals. N = 1800. PredictorsEstimate95% CIp Intercept4.53[4.48, 4.59] <0.001 H1f: Persuasive outreach0.40[0.29, 0.51] <0.001 H2f: AI-mediated outreach0.48[0.37, 0.59] <0.001 H3: PersuasiveĂ AI-mediated outreach-0.30[-0.52, -0.08] 0.008 Observations1800 R 2 / adjusted R 2 0.068 / 0.066 Table 19: OLS regression model for penalty for source with 95% confidence intervals. N = 1800. 38 C.1.3 Effect size OutcomePersuasion penaltyAI penalty H1a/H2a: Willingness to participate -0.09 [-0.18, 0.00]0.01 [-0.08, 0.10] H1b/H2b: Perceived threat to freedom 0.53 [0.43, 0.62]0.11 [0.02, 0.20] H1c/H2c: Acceptability-0.33 [-0.43, -0.24]-0.39 [-0.49, -0.30] H1d/H2d: Positive impact-0.19 [-0.28, -0.10]-0.28 [-0.37, -0.18] H1e/H2e: Future campaign avoidance 0.25 [0.15, 0.34]0.22 [0.13, 0.32] H1f/H2f: Source evaluation0.28 [0.19, 0.37]0.41 [0.32, 0.51] Table 20: Cohenâs d effect sizes for the two main predictors across all preregis- tered outcomes in the United States. Entries are Cohenâs d with 95% confidence intervals, signed to match the reporting direction used in the main analyses. OutcomePersuasion penaltyAI penalty H1a/H2a: Willingness to participate -0.12 [-0.22, -0.03]-0.02 [-0.11, 0.07] H1b/H2b: Perceived threat to freedom 0.51 [0.42, 0.60]0.17 [0.08, 0.26] H1c/H2c: Acceptability-0.42 [-0.51, -0.32]-0.41 [-0.50, -0.31] H1d/H2d: Positive impact-0.22 [-0.31, -0.12]-0.22 [-0.31, -0.13] H1e/H2e: Future campaign avoidance 0.36 [0.26, 0.45]0.31 [0.22, 0.41] H1f/H2f: Source evaluation0.33 [0.23, 0.42]0.40 [0.30, 0.49] Table 21: Cohenâs d effect sizes for the two main predictors across all pre- registered outcomes in the United Kingdom. Entries are Cohenâs d with 95% confidence intervals, signed to match the reporting direction used in the main analyses. C.2 Preregistered planned contrasts Following preregistration, we conducted planned simple effect contrasts to com- pare AI- and human-mediated outreach separately within informational and persuasive conditions. The P-values for these contrasts were adjusted using Holmâs method. 39 C.2.1 United States Outcome Information outreach: Humanâ AI Persuasive outreach: Humanâ AI H1a/H2a: Willingness to participate 0.10 [-0.17, 0.37], p Holm = .460-0.15 [-0.41, 0.12], p Holm = .286 H1b/H2b: Perceived threat to freedom -0.27 [-0.49, -0.06], p Holm = .013-0.10 [-0.32, 0.11], p Holm = .350 H1c/H2c: Acceptability of campaign outreach 0.68 [0.49, 0.87], p Holm < .0010.48 [0.29, 0.67], p Holm < .001 H1d/H2d: Perceived positive impact of campaign outreach 0.48 [0.27, 0.68], p Holm < .0010.40 [0.20, 0.61], p Holm < .001 H1e/H2e: Future campaign avoidance -0.51 [-0.74, -0.29], p Holm < .001-0.25 [-0.47, -0.03], p Holm = .027 H1f/H2f: Penalty for source -0.62 [-0.79, -0.45], p Holm < .001-0.49 [-0.66, -0.32], p Holm < .001 Table 22: Preregistered planned contrasts comparing human-mediated and AI- mediated outreach within each outreach condition. Entries are estimated mean differences (Humanâ AI) with 95% confidence intervals and Holm-adjusted p-values. Positive values indicate higher outcome values for human-mediated outreach; negative values indicate higher outcome values for AI-mediated out- reach. N = 1800. 40 C.2.2 United Kingdom Outcome Information outreach: Humanâ AI Persuasive outreach: Humanâ AI H1a/H2a: Willingness to participate 0.24 [-0.01, 0.49], p Holm = .063-0.17 [-0.41, 0.08], p Holm = .188 H1b/H2b: Perceived threat to freedom -0.40 [-0.60, -0.20], p Holm < .001-0.15 [-0.35, 0.05], p Holm = .151 H1c/H2c: Acceptability of campaign outreach 0.65 [0.48, 0.83], p Holm < .0010.45 [0.28, 0.62], p Holm < .001 H1d/H2d: Perceived positive impact of campaign outreach 0.51 [0.32, 0.69], p Holm < .0010.12 [-0.07, 0.30], p Holm = .210 H1e/H2e: Future campaign avoidance -0.67 [-0.87, -0.47], p Holm < .001-0.32 [-0.52, -0.12], p Holm = .002 H1f/H2f: Penalty for source -0.63 [-0.79, -0.48], p Holm < .001-0.33 [-0.49, -0.18], p Holm < .001 Table 23: Preregistered planned contrasts comparing human-mediated and AI- mediated outreach within each outreach condition. Entries are estimated mean differences (Humanâ AI) with 95% confidence intervals and Holm-adjusted p-values. Positive values indicate higher outcome values for human-mediated outreach; negative values indicate higher outcome values for AI-mediated out- reach. N = 1800. C.3 Moderators We also preregistered three moderator variables. In this section, we report the interactions between these three moderators and the human vs. AI outreach variable. C.3.1 Avoidance of political conversations United States 41 PredictorsEstimate95% CIp Intercept3.58[3.49, 3.68] <0.001 Persuasive outreach-0.19[-0.38, -0.00] 0.048 AI-mediated outreach0.02[-0.17, 0.21] 0.827 Avoidance of political conversations-0.12[-0.17, -0.07] <0.001 PersuasiveĂ AI-mediated outreach0.24[-0.13, 0.62] 0.204 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.10[-0.01, 0.20] 0.070 Observations1800 R 2 / adjusted R 2 0.015 / 0.013 Table 24: OLS regression model for willingness to participate with 95% confi- dence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.60[3.53, 3.68] <0.001 Persuasive outreach0.88[0.73, 1.03] <0.001 AI-mediated outreach0.19[0.04, 0.34] 0.014 Avoidance of political conversations0.11[0.07, 0.15] <0.001 PersuasiveĂ AI-mediated outreach-0.16[-0.46, 0.14] 0.297 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.02[-0.06, 0.10] 0.655 Observations1800 R 2 / adjusted R 2 0.081 / 0.079 Table 25: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including avoidance of political conversations as a moderator. N = 1800. 42 PredictorsEstimate95% CIp Intercept3.63[3.57, 3.70] <0.001 Persuasive outreach-0.50[-0.64, -0.37] <0.001 AI-mediated outreach-0.58[-0.72, -0.45] <0.001 Avoidance of political conversations-0.05[-0.09, -0.01] 0.008 PersuasiveĂ AI-mediated outreach0.20[-0.07, 0.46] 0.155 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.01[-0.06, 0.09] 0.777 Observations1800 R 2 / adjusted R 2 0.069 / 0.067 Table 26: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.70[3.63, 3.77] <0.001 Persuasive outreach-0.30[-0.45, -0.16] <0.001 AI-mediated outreach-0.44[-0.58, -0.29] <0.001 Avoidance of political conversations-0.04[-0.08, -0.00] 0.036 PersuasiveĂ AI-mediated outreach0.08[-0.21, 0.37] 0.590 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.07[-0.01, 0.15] 0.085 Observations1800 R 2 / adjusted R 2 0.032 / 0.029 Table 27: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including avoidance of political conversa- tions as a moderator. N = 1800. 43 PredictorsEstimate95% CIp Intercept4.51[4.43, 4.59] <0.001 Persuasive outreach0.43[0.28, 0.59] <0.001 AI-mediated outreach0.38[0.23, 0.54] <0.001 Avoidance of political conversations0.15[0.11, 0.19] <0.001 PersuasiveĂ AI-mediated outreach-0.25[-0.56, 0.05] 0.106 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.00[-0.08, 0.09] 0.991 Observations1800 R 2 / adjusted R 2 0.054 / 0.052 Table 28: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.38[4.32, 4.44] <0.001 Persuasive outreach0.38[0.26, 0.50] <0.001 AI-mediated outreach0.55[0.43, 0.68] <0.001 Avoidance of political conversations0.05[0.02, 0.09] 0.001 PersuasiveĂ AI-mediated outreach-0.13[-0.37, 0.12] 0.305 RQ1: AI-mediated outreachĂ Avoidance of political conversations -0.02[-0.08, 0.05] 0.634 Observations1800 R 2 / adjusted R 2 0.067 / 0.064 Table 29: OLS regression model for penalty for source with 95% confidence intervals, including avoidance of political conversations as a moderator. N = 1800. United Kingdom 44 PredictorsEstimate95% CIp Intercept3.80[3.71, 3.89] <0.001 Persuasive outreach-0.23[-0.41, -0.06] 0.009 AI-mediated outreach-0.03[-0.21, 0.14] 0.717 Avoidance of political conversations-0.10[-0.16, -0.05] <0.001 PersuasiveĂ AI-mediated outreach0.41[0.06, 0.76] 0.021 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.15[0.05, 0.26] 0.004 Observations1800 R 2 / adjusted R 2 0.020 / 0.017 Table 30: OLS regression model for willingness to participate with 95% confi- dence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.81[3.74, 3.88] <0.001 Persuasive outreach0.78[0.64, 0.92] <0.001 AI-mediated outreach0.27[0.13, 0.41] <0.001 Avoidance of political conversations0.15[0.10, 0.19] <0.001 PersuasiveĂ AI-mediated outreach-0.27[-0.55, 0.01] 0.057 RQ1: AI-mediated outreachĂ Avoidance of political conversations -0.09[-0.18, -0.01] 0.034 Observations1800 R 2 / adjusted R 2 0.096 / 0.093 Table 31: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including avoidance of political conversations as a moderator. N = 1800. 45 PredictorsEstimate95% CIp Intercept3.59[3.53, 3.65] <0.001 Persuasive outreach-0.56[-0.68, -0.44] <0.001 AI-mediated outreach-0.55[-0.67, -0.43] <0.001 Avoidance of political conversations-0.09[-0.12, -0.05] <0.001 PersuasiveĂ AI-mediated outreach0.22[-0.02, 0.46] 0.078 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.03[-0.05, 0.10] 0.449 Observations1800 R 2 / adjusted R 2 0.094 / 0.092 Table 32: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.87[3.81, 3.94] <0.001 Persuasive outreach-0.31[-0.44, -0.18] <0.001 AI-mediated outreach-0.31[-0.44, -0.18] <0.001 Avoidance of political conversations-0.05[-0.09, -0.01] 0.009 PersuasiveĂ AI-mediated outreach0.40[0.14, 0.66] 0.003 RQ1: AI-mediated outreachĂ Avoidance of political conversations 0.02[-0.06, 0.10] 0.627 Observations1800 R 2 / adjusted R 2 0.032 / 0.030 Table 33: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including avoidance of political conversa- tions as a moderator. N = 1800. 46 PredictorsEstimate95% CIp Intercept4.49[4.42, 4.56] <0.001 Persuasive outreach0.55[0.41, 0.69] <0.001 AI-mediated outreach0.49[0.35, 0.63] <0.001 Avoidance of political conversations0.16[0.12, 0.21] <0.001 PersuasiveĂ AI-mediated outreach-0.37[-0.66, -0.09] 0.009 RQ1: AI-mediated outreachĂ Avoidance of political conversations -0.08[-0.16, 0.01] 0.083 Observations1800 R 2 / adjusted R 2 0.089 / 0.087 Table 34: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including avoidance of political conversations as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.53[4.48, 4.59] <0.001 Persuasive outreach0.40[0.29, 0.51] <0.001 AI-mediated outreach0.48[0.37, 0.59] <0.001 Avoidance of political conversations0.07[0.04, 0.11] <0.001 PersuasiveĂ AI-mediated outreach-0.31[-0.53, -0.09] 0.005 RQ1: AI-mediated outreachĂ Avoidance of political conversations -0.03[-0.10, 0.03] 0.342 Observations1800 R 2 / adjusted R 2 0.078 / 0.076 Table 35: OLS regression model for penalty for source with 95% confidence intervals, including avoidance of political conversations as a moderator. N = 1800. C.3.2 Feeling toward people with opposing opinions United States 47 PredictorsEstimate95% CIp Intercept3.58[3.49, 3.68] <0.001 Persuasive outreach-0.23[-0.41, -0.04] 0.017 AI-mediated outreach0.03[-0.16, 0.21] 0.781 Feeling toward people with opposing opinions 0.02[0.01, 0.02] <0.001 PersuasiveĂ AI-mediated outreach0.29[-0.08, 0.65] 0.128 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.00[-0.01, 0.01] 0.885 Observations1800 R 2 / adjusted R 2 0.066 / 0.063 Table 36: OLS regression model for willingness to participate with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.60[3.53, 3.68] <0.001 Persuasive outreach0.88[0.72, 1.03] <0.001 AI-mediated outreach0.19[0.04, 0.34] 0.016 Feeling toward people with opposing opinions -0.00[-0.01, -0.00] <0.001 PersuasiveĂ AI-mediated outreach-0.17[-0.48, 0.13] 0.264 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.00[-0.01, 0.00] 0.199 Observations1800 R 2 / adjusted R 2 0.075 / 0.073 Table 37: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. 48 PredictorsEstimate95% CIp Intercept3.63[3.57, 3.70] <0.001 Persuasive outreach-0.52[-0.66, -0.39] <0.001 AI-mediated outreach-0.58[-0.71, -0.45] <0.001 Feeling toward people with opposing opinions 0.01[0.01, 0.01] <0.001 PersuasiveĂ AI-mediated outreach0.21[-0.05, 0.47] 0.117 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions 0.01[0.00, 0.01] 0.020 Observations1800 R 2 / adjusted R 2 0.116 / 0.114 Table 38: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.70[3.63, 3.77] <0.001 Persuasive outreach-0.34[-0.48, -0.20] <0.001 AI-mediated outreach-0.44[-0.57, -0.30] <0.001 Feeling toward people with opposing opinions 0.02[0.02, 0.02] <0.001 PersuasiveĂ AI-mediated outreach0.10[-0.17, 0.38] 0.464 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions 0.01[0.00, 0.01] 0.033 Observations1800 R 2 / adjusted R 2 0.132 / 0.129 Table 39: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including feeling toward people with op- posing opinions as a moderator. N = 1800. 49 PredictorsEstimate95% CIp Intercept4.51[4.43, 4.58] <0.001 Persuasive outreach0.44[0.29, 0.60] <0.001 AI-mediated outreach0.38[0.23, 0.53] <0.001 Feeling toward people with opposing opinions -0.01[-0.01, -0.01] <0.001 PersuasiveĂ AI-mediated outreach-0.28[-0.58, 0.03] 0.077 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.01[-0.01, -0.00] 0.037 Observations1800 R 2 / adjusted R 2 0.066 / 0.063 Table 40: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.38[4.32, 4.44] <0.001 Persuasive outreach0.40[0.28, 0.52] <0.001 AI-mediated outreach0.55[0.43, 0.67] <0.001 Feeling toward people with opposing opinions -0.01[-0.01, -0.01] <0.001 PersuasiveĂ AI-mediated outreach-0.14[-0.37, 0.10] 0.266 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.01[-0.01, -0.00] <0.001 Observations1800 R 2 / adjusted R 2 0.109 / 0.107 Table 41: OLS regression model for penalty for source with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. United Kingdom 50 PredictorsEstimate95% CIp Intercept3.80[3.71, 3.89] <0.001 Persuasive outreach-0.24[-0.41, -0.06] 0.007 AI-mediated outreach-0.01[-0.18, 0.17] 0.931 Feeling toward people with opposing opinions 0.01[0.01, 0.02] <0.001 PersuasiveĂ AI-mediated outreach0.42[0.08, 0.77] 0.016 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.00[-0.01, 0.00] 0.470 Observations1800 R 2 / adjusted R 2 0.038 / 0.036 Table 42: OLS regression model for willingness to participate with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.81[3.74, 3.88] <0.001 Persuasive outreach0.79[0.65, 0.93] <0.001 AI-mediated outreach0.26[0.12, 0.41] <0.001 Feeling toward people with opposing opinions -0.00[-0.01, -0.00] 0.001 PersuasiveĂ AI-mediated outreach-0.26[-0.54, 0.02] 0.073 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions 0.00[-0.00, 0.01] 0.215 Observations1800 R 2 / adjusted R 2 0.076 / 0.073 Table 43: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. 51 PredictorsEstimate95% CIp Intercept3.59[3.53, 3.65] <0.001 Persuasive outreach-0.57[-0.69, -0.44] <0.001 AI-mediated outreach-0.53[-0.65, -0.41] <0.001 Feeling toward people with opposing opinions 0.01[0.01, 0.01] <0.001 PersuasiveĂ AI-mediated outreach0.22[-0.02, 0.46] 0.076 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions 0.00[-0.00, 0.01] 0.210 Observations1800 R 2 / adjusted R 2 0.108 / 0.105 Table 44: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.87[3.81, 3.94] <0.001 Persuasive outreach-0.31[-0.43, -0.18] <0.001 AI-mediated outreach-0.28[-0.41, -0.16] <0.001 Feeling toward people with opposing opinions 0.01[0.01, 0.02] <0.001 PersuasiveĂ AI-mediated outreach0.41[0.16, 0.67] 0.001 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions 0.00[-0.00, 0.01] 0.424 Observations1800 R 2 / adjusted R 2 0.097 / 0.094 Table 45: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including feeling toward people with op- posing opinions as a moderator. N = 1800. 52 PredictorsEstimate95% CIp Intercept4.49[4.42, 4.56] <0.001 Persuasive outreach0.56[0.42, 0.70] <0.001 AI-mediated outreach0.48[0.34, 0.62] <0.001 Feeling toward people with opposing opinions -0.01[-0.01, -0.01] <0.001 PersuasiveĂ AI-mediated outreach-0.36[-0.65, -0.08] 0.012 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.00[-0.01, 0.01] 0.879 Observations1800 R 2 / adjusted R 2 0.077 / 0.075 Table 46: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.53[4.48, 4.58] <0.001 Persuasive outreach0.40[0.29, 0.51] <0.001 AI-mediated outreach0.46[0.35, 0.57] <0.001 Feeling toward people with opposing opinions -0.01[-0.01, -0.01] <0.001 PersuasiveĂ AI-mediated outreach-0.32[-0.53, -0.10] 0.004 RQ2: AI-mediated outreachĂ Feeling toward people with opposing opinions -0.00[-0.01, -0.00] 0.050 Observations1800 R 2 / adjusted R 2 0.108 / 0.105 Table 47: OLS regression model for penalty for source with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N = 1800. C.3.3 AI risk perception United States 53 PredictorsEstimate95% CIp Intercept3.58[3.49, 3.67] <0.001 Persuasive outreach-0.22[-0.41, -0.03] 0.020 AI-mediated outreach-0.00[-0.19, 0.19] 0.991 RQ3: AI risk perception-0.23[-0.30, -0.16] <0.001 PersuasiveĂ AI-mediated outreach0.20[-0.17, 0.58] 0.291 RQ3: AI-mediated outreachĂ AI risk perception -0.15[-0.30, -0.01] 0.034 Observations1800 R 2 / adjusted R 2 0.027 / 0.024 Table 48: OLS regression model for willingness to participate with 95% confi- dence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.61[3.53, 3.68] <0.001 Persuasive outreach0.91[0.76, 1.06] <0.001 AI-mediated outreach0.21[0.06, 0.36] 0.006 RQ3: AI risk perception0.22[0.16, 0.28] <0.001 PersuasiveĂ AI-mediated outreach-0.13[-0.43, 0.17] 0.394 RQ3: AI-mediated outreachĂ AI risk perception 0.14[0.02, 0.25] 0.019 Observations1800 R 2 / adjusted R 2 0.101 / 0.098 Table 49: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including AI risk perception as a moderator. N = 1800. 54 PredictorsEstimate95% CIp Intercept3.63[3.56, 3.69] <0.001 Persuasive outreach-0.55[-0.68, -0.42] <0.001 AI-mediated outreach-0.61[-0.74, -0.48] <0.001 RQ3: AI risk perception-0.29[-0.34, -0.24] <0.001 PersuasiveĂ AI-mediated outreach0.15[-0.11, 0.41] 0.255 RQ3: AI-mediated outreachĂ AI risk perception -0.14[-0.24, -0.04] 0.006 Observations1800 R 2 / adjusted R 2 0.134 / 0.131 Table 50: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.70[3.63, 3.77] <0.001 Persuasive outreach-0.34[-0.48, -0.20] <0.001 AI-mediated outreach-0.47[-0.61, -0.32] <0.001 RQ3: AI risk perception-0.25[-0.30, -0.20] <0.001 PersuasiveĂ AI-mediated outreach0.04[-0.25, 0.32] 0.791 RQ3: AI-mediated outreachĂ AI risk perception -0.11[-0.22, 0.00] 0.051 Observations1800 R 2 / adjusted R 2 0.073 / 0.071 Table 51: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. 55 PredictorsEstimate95% CIp Intercept4.51[4.44, 4.59] <0.001 Persuasive outreach0.48[0.33, 0.63] <0.001 AI-mediated outreach0.41[0.26, 0.56] <0.001 RQ3: AI risk perception0.32[0.26, 0.38] <0.001 PersuasiveĂ AI-mediated outreach-0.21[-0.51, 0.09] 0.178 RQ3: AI-mediated outreachĂ AI risk perception 0.20[0.08, 0.31] <0.001 Observations1800 R 2 / adjusted R 2 0.097 / 0.095 Table 52: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.38[4.32, 4.44] <0.001 Persuasive outreach0.43[0.31, 0.55] <0.001 AI-mediated outreach0.59[0.47, 0.70] <0.001 RQ3: AI risk perception0.30[0.26, 0.35] <0.001 PersuasiveĂ AI-mediated outreach-0.08[-0.32, 0.15] 0.483 RQ3: AI-mediated outreachĂ AI risk perception 0.13[0.04, 0.22] 0.004 Observations1800 R 2 / adjusted R 2 0.152 / 0.149 Table 53: OLS regression model for penalty for source with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. United Kingdom 56 PredictorsEstimate95% CIp Intercept3.80[3.72, 3.89] <0.001 Persuasive outreach-0.23[-0.41, -0.06] 0.010 AI-mediated outreach-0.03[-0.20, 0.15] 0.755 RQ3: AI risk perception-0.16[-0.23, -0.08] <0.001 PersuasiveĂ AI-mediated outreach0.43[0.08, 0.78] 0.017 RQ3: AI-mediated outreachĂ AI risk perception -0.09[-0.24, 0.06] 0.260 Observations1800 R 2 / adjusted R 2 0.017 / 0.014 Table 54: OLS regression model for willingness to participate with 95% confi- dence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.81[3.74, 3.88] <0.001 Persuasive outreach0.78[0.64, 0.92] <0.001 AI-mediated outreach0.26[0.12, 0.40] <0.001 RQ3: AI risk perception0.23[0.17, 0.30] <0.001 PersuasiveĂ AI-mediated outreach-0.29[-0.57, -0.01] 0.045 RQ3: AI-mediated outreachĂ AI risk perception 0.08[-0.04, 0.20] 0.196 Observations1800 R 2 / adjusted R 2 0.100 / 0.097 Table 55: OLS regression model for perceived threat to freedom with 95% con- fidence intervals, including AI risk perception as a moderator. N = 1800. 57 PredictorsEstimate95% CIp Intercept3.59[3.53, 3.65] <0.001 Persuasive outreach-0.56[-0.68, -0.44] <0.001 AI-mediated outreach-0.54[-0.66, -0.42] <0.001 RQ3: AI risk perception-0.23[-0.28, -0.17] <0.001 PersuasiveĂ AI-mediated outreach0.24[-0.00, 0.48] 0.051 RQ3: AI-mediated outreachĂ AI risk perception -0.13[-0.23, -0.02] 0.017 Observations1800 R 2 / adjusted R 2 0.121 / 0.118 Table 56: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept3.87[3.81, 3.94] <0.001 Persuasive outreach-0.30[-0.43, -0.17] <0.001 AI-mediated outreach-0.31[-0.43, -0.18] <0.001 RQ3: AI risk perception-0.16[-0.21, -0.10] <0.001 PersuasiveĂ AI-mediated outreach0.41[0.16, 0.67] 0.002 RQ3: AI-mediated outreachĂ AI risk perception -0.14[-0.25, -0.03] 0.015 Observations1800 R 2 / adjusted R 2 0.048 / 0.045 Table 57: OLS regression model for perceived positive impact of campaign out- reach with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. 58 PredictorsEstimate95% CIp Intercept4.49[4.42, 4.56] <0.001 Persuasive outreach0.55[0.41, 0.69] <0.001 AI-mediated outreach0.48[0.34, 0.62] <0.001 RQ3: AI risk perception0.26[0.20, 0.32] <0.001 PersuasiveĂ AI-mediated outreach-0.39[-0.67, -0.11] 0.007 RQ3: AI-mediated outreachĂ AI risk perception 0.20[0.08, 0.32] 0.001 Observations1800 R 2 / adjusted R 2 0.098 / 0.096 Table 58: OLS regression model for future campaign avoidance with 95% con- fidence intervals, including AI risk perception as a moderator. N = 1800. PredictorsEstimate95% CIp Intercept4.53[4.48, 4.58] <0.001 Persuasive outreach0.39[0.29, 0.50] <0.001 AI-mediated outreach0.47[0.36, 0.58] <0.001 RQ3: AI risk perception0.24[0.19, 0.28] <0.001 PersuasiveĂ AI-mediated outreach-0.34[-0.55, -0.12] 0.002 RQ3: AI-mediated outreachĂ AI risk perception 0.17[0.08, 0.27] <0.001 Observations1800 R 2 / adjusted R 2 0.123 / 0.120 Table 59: OLS regression model for penalty for source with 95% confidence intervals, including AI risk perception as a moderator. N = 1800. C.4 Manipulation check and instrumental variable robust- ness check Most respondents remembered that the campaignâs stance opposed their own position (US: 90.7%; UK: 91.4%). However, recognition of the specific treatment condition was lower in the human-mediated and informational conditions than in the AI-mediated and persuasive conditions. This pattern is plausible for two reasons. First, the detailed manipulation-check items were asked only at the very end of the survey, after respondents had already completed a large number of outcome measures, which likely made exact recall more difficult. Second, even outreach designed to be informational may still be perceived as somewhat persuasive, making this distinction less clear-cut for respondents. As a robustness check, we therefore used an instrumental variable (IV) ap- proach (Montgomery, Nyhan, & Torres, 2018) to estimate treatment effects for 59 participants whose manipulation-check responses aligned with their assigned condition. Random treatment assignment served as the instrument for per- ceived treatment. Compliance with the source manipulation was high in the AI condition (US: 95.2%; UK: 97.6%) and moderate in the human condition (US: 65.1%; UK: 61.0%). Compliance with the outreach manipulation was high in the persuasive condition (US: 85.2%; UK: 87.2%) and moderate in the informational condition (US: 60.0%; UK: 54.3%). Importantly, participants who answered ânot sure / canât rememberâ were coded as non-compliers, alongside those who selected the incorrect answers. This makes the IV analysis conservative, as some of these responses may reflect uncertainty or imperfect recall rather than complete failure to register the treatment. The IV estimates were consistent with the intent-to-treat results in terms of statistical significance and the overall pattern of effects across all six outcomes in both countries. These findings indicate that the main results are not driven by participants who failed the manipulation checks. Because this specification uses the human-mediated and informational conditions as the focal categories, the IV coefficients have the opposite sign relative to the main specification, while implying the same substantive conclusions. Outcome Human-focused: Human-mediated outreach (d human ) Informational-focused: Informational outreach (d info ) H1a/H2a: Willingness to participate â0.04 [â0.34, 0.26], p = .8130.35 [â0.01, 0.72], p = .060 H1b/H2b: Perceived threat to freedom â0.30 [â0.54,â0.06], p = .016â1.69 [â1.99,â1.39], p < .001 H1c/H2c: Acceptability0.92 [0.70, 1.14], p < .0010.97 [0.71, 1.23], p < .001 H1d/H2d: Perceived positive impact 0.69 [0.46, 0.92], p < .0010.58 [0.30, 0.86], p < .001 H1e/H2e: Future campaign avoidance â0.60 [â0.85,â0.35], p < .001â0.82 [â1.12,â0.52], p < .001 H1f/H2f: Penalty for source â0.87 [â1.07,â0.68], p < .001â0.74 [â0.97,â0.50], p < .001 Table 60: Alternative IV specifications using reversed focal categories for the US sample. Entries report unstandardized coefficients with 95% confidence intervals and two-sided p-values from alternative IV specifications using re- versed focal categories. D Content analysis of issues In our study, we asked participants to name an issue that is important to them using the following question: "People have different political priorities. Please think of one political or social issue that matters to you personally. Briefly name the issue in the field below." 60 0.3% 3.1% 22.8% 3.6% 11.1% 5.3% 0.0% 13.2% 5.1% 19.3% 0.6% 4.3% 3.1% 2.1% 6.2% 12.5% 7.2% 14.2% 1.8% 8.9% 6.4% 4.4% 9.4% 1.9% 14.4% 3.2% 4.3% 4.2% 4.1% 3.0% LGBTQ+ Rights Education War & Foreign Policy Uncodable Other Gun Policy Housing Poverty, Welfare & Safety Net Race, Equality & Civil Rights Democracy & Governance Climate & Environment Abortion & Reprod. Rights Healthcare Immigration Economy & Cost of Living 0%5%10%15%20%25% % of respondents UKUS Figure 4: Distribution of issue categories in the US and the UK. 61 D.1 Coding procedure In an additional analysis, we classified these open-ended responses into broader issue categories. We first used Claude Code (Sonnet 4.6) to inspect all responses and propose 10â15 broader categories into which the issues could be grouped. Claude Code generated 15 candidate categories, along with category definitions and a classification prompt. We then manually reviewed these categories and made minor revisions to the prompt (see Section D.3). To validate the resulting coding scheme, the two authors of the study independently manually coded a random sample of 50 responses. The same responses were also classified auto- matically using OpenAIâs gpt-4o-mini-2024-07-18 model with the tempera- ture set to 0. The two human coders and the automatic classifier showed good intercoder reliability (Krippendorffâs alpha = 0.81). After this validation step, we used the model to classify all remaining responses. 62 D.2 Issue distributions â12.2 p â4.1 p +8.7 p +1.8 p +2.1 p â1.1 p â4.4 p +3.7 p +3.2 p +4.8 p â2.6 p â0.1 p â1.1 p â2.0 p +3.2 p Abortion & Reprod. Rights Gun Policy Democracy & Governance LGBTQ+ Rights War & Foreign Policy Race, Equality & Civil Rights Uncodable Other Education Climate & Environment Poverty, Welfare & Safety Net Housing Healthcare Immigration Economy & Cost of Living â15 pâ10 pâ5 p0 p5 p10 p Percentageâpoint difference UK higherUS higher Figure 5: Percentage-point difference (UKâUS); positive values indicate that an issue was more common in the UK. Overall, the economy and cost of living, immigration, healthcare, and climate change were the most frequently mentioned issues in both countries (see Fig- ure 4). Furthermore, the issue distributions in each country clearly reflect the respective political context. For example, abortion and gun policy were clearly overrepresented in the US and largely absent in the UK, in contrast to the cost of living and healthcare, which were mentioned relatively more often in the UK (see Figure 5). When analyzed alongside participantsâ partisan identification, the issue profiles also broadly aligned with party cleavages in both the US (see Figure 6) and the UK (see Figure 7). 63 Education Housing LGBTQ+ Rights Poverty, Welfare & Safety Net War & Foreign Policy Gun Policy Democracy & Governance Race, Equality & Civil Rights Healthcare Climate & Environment Abortion & Reprod. Rights Economy & Cost of Living Immigration 0%10%20% % within party RepublicanDemocratIndependent Figure 6: Party identification composition of issue salience for the US sample. Bars show the percentage of each partyâs respondents who mention a given issue category. Uncodable and other categories are not shown. 64 Abortion & Reprod. Rights LGBTQ+ Rights War & Foreign Policy Democracy & Governance Education Housing Poverty, Welfare & Safety Net Race, Equality & Civil Rights Healthcare Climate & Environment Economy & Cost of Living Immigration 0%20%40%60% % within party Labour Conservative Lib Dem Green Reform UK SNP Figure 7: Party identification composition of issue salience for the UK sample. Bars show the percentage of each partyâs respondents who mention a given issue category. Uncodable and other categories are not shown. 65 D.3 Classification prompt We used the following prompt with OpenAIâs model: You are a political science research assistant coding open-ended survey responses. Each response is a participantâs answer to: âWhat is one political or social issue that matters to you personally?â Assign ONE primary category from this list: ECO - Economy & Cost of Living (inflation, wages, taxes, wealth inequality, cost of living) IMM - Immigration (legal/illegal immigration, border security, ICE, deportation, asylum) HLT - Healthcare (NHS, universal healthcare, costs, access, mental health, social care) ABR - Abortion & Reproductive Rights (abortion, pro-life/choice, reproductive rights, bodily autonomy) ENV - Climate Change & Environment (climate, global warming, pollution, renewables, nature) GUN - Gun Policy (gun control, gun rights, Second Amendment, firearms ) HSG - Housing (housing crisis, affordability, homelessness as housing failure) WLF - Poverty, Welfare & Social Safety Net (poverty, benefits, pensions, food insecurity, student debt, disability support) LGB - LGBTQ+ Rights (gay/trans rights, same-sex marriage, gender identity) EQL - Race, Equality & Civil Rights (racism, civil rights, womenâs rights, discrimination, social justice) EDU - Education (school/university quality, funding, tuition, curriculum) DEM - Democracy, Governance & Rule of Law (democracy, voting, corruption, free speech, political reform, far right) WAR - War, Security & Foreign Policy (war, Ukraine, Iran, Gaza, defence, national security) OTH - Other political/social issue not listed above (e.g., crime/ public safety, AI regulation, animal welfare, drug policy) UNK - Uncodable (vague, blank, non-political, or partisan without issue) Rules: - Assign primary = the most prominent issue in the response - If the response names no recognisable issue, assign UNK - Only return the three letter label and nothing else Here is the issue that you should classify: 66 E Additional analysis with political orientation In both countries, we measured political orientation on a seven-point scale (1 = left, 7 = right). As the treatment effects may vary by political orientation, we also examined interactions between political orientation and our main inde- pendent variables. Political orientation is a significant moderator of human- vs. AI-mediated outreach for some dependent variables. In the UK, by contrast, it is a significant moderator of informational- vs. persuasive outreach for all de- pendent variables (see Figure 8). In the UK, the more left-leaning respondents are, the stronger the effect of persuasive outreach. In the US, by contrast, the more left-leaning respondents are, the stronger the effect of AI-mediated out- reach. At the same time, these interaction patterns mainly affect the magnitude of the effects rather than their overall direction. Even where political orientation significantly moderates the treatment effects, the substantive pattern remains similar across the ideological spectrum. n.s. n.s. n.s. n.s. n.s. n.s. * n.s. n.s. ** * ** ** *** *** * ** *** n.s. n.s. n.s. n.s. n.s. n.s. US â Persuasive outreachUS â AIâmediated outreachUK â Persuasive outreachUK â AIâmediated outreach Willingness to participate Perceived threat to freedom Acceptability Perceived positive impact Future campaign avoidance Penalty for source 1 (Left)7 (Right)1 (Left)7 (Right)1 (Left)7 (Right)1 (Left)7 (Right) â1 0 1 â1 0 1 â1 0 1 â1 0 1 â1 0 1 â1 0 1 Political orientation (1 = Left, 7 = Right) Conditional effect Interaction: n.s.p < .05 Conditional Effects with political orientation Figure 8: Conditional effects by political orientation in the UK and the US. 67 F Additional analysis with norm violation item As an additional non-preregistered analysis of the normative violation mech- anism, we examined the item "This kind of outreach is not how campaigns should operate", one of three items comprising the acceptability composite, separately. The wording of this item explicitly captures normative appropriate- ness judgments about communicative practice rather than affective evaluation or perceived effectiveness. Consistent with the normative violation account, the AI penalty on this item substantially exceeded the persuasion penalty in both countries (US: AI Cohenâs d = 0.37 vs. persuasion Cohenâs d = 0.21; UK: AI Cohenâs d = 0.36 vs. persuasion Cohenâs d = 0.28). This contrasts directly with the pattern for perceived threat to freedom, where the persuasion penalty substantially dominated the AI penalty (US: persuasion Cohenâs d = 0.53 vs. AI Cohenâs d = 0.11; UK: persuasion Cohenâs d = 0.51 vs. AI Cohenâs d = 0.17). The reversal across these two indicators, each mechanism producing its largest signal on a different outcome, is consistent with distinct primary path- ways: reactance to persuasive intent driving freedom-threat perceptions, and normative concerns about AI as a communicative agent driving appropriate- ness judgments. These findings do not mean that the two mechanisms are fully separate. Instead, they partly overlap: the AI penalty also increases perceived threat to freedom, and the persuasion penalty also increases agreement with the norm violation item. The double dissociation, therefore, speaks to the relative strength of each pathway, not to an exclusive link between the mechanism and the indicator. Supplementary Information References Gordon, A. (2026). Authenticity checks detect AI agents best. Re- trieved 2026-03-19, from https://w.prolific.com/resources/ authenticity-checks-how-we-tested-the-most-accurate-method -for-identifying-agentic-ai Lakens, D., & Caldwell, A. R. (2021). Simulation-Based Power Analysis for Factorial Analysis of Variance Designs. Advances in Methods and Prac- tices in Psychological Science, 4(1), 2515245920951503. doi: 10.1177/ 2515245920951503 Montgomery, J. M., Nyhan, B., & Torres, M. (2018). How Conditioning on Posttreatment Variables Can Ruin Your Experiment and What to Do about It. American Journal of Political Science, 62(3), 760â775. doi: 10.1111/ajps.12357 68