Chatbots Reduce Health-Related Conspiracy Beliefs Not Because of but Despite Being Perceived as AI

Chatbots Reduce Health-Related Conspiracy Beliefs Not Because of but Despite Being Perceived as AI | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Chatbots Reduce Health-Related Conspiracy Beliefs Not Because of but Despite Being Perceived as AI Paul Ballot, Yana van de Sande, Hanna Schraffenberger, Philipp Schmid This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9428259/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Beliefs in conspiracy theories and their resistance to correction pose critical challenges to public health. A potential countermeasure to this is the use of Large Language Model (LLM) driven debates about conspiracy theories. Yet while this approach yields promising results for general conspiracies, its generalizability to the health domain as well as the underlying mechanism for its effectiveness remain unclear. This study investigates whether LLM-driven debates can reduce health-related conspiracy theories in the context of COVID-19. Furthermore, it examines attribution to artificial rather than human sources as a potential explanation. In an online experiment, 554 participants were randomly assigned to either a control condition or to debate an individual COVID-19 conspiracy theory with an AI-labelled LLM or a human-labelled LLM. Compared to the control, participants aware of their artificial conversation partner reported 7.88 percentage points less confidence in the conspiracy after the intervention ( d = 0.63, p < .001). Contrary to our expectation, however, this effect was stronger for those assuming a human conversation partner: They indicated a 13.76 percentage points larger confidence drop compared to the control ( d = 0.87, p < .001). This difference in effectiveness appears to be mediated by the perceived neutrality of the source. Biological sciences/Psychology Social science/Psychology Social science/Science technology and society Health misinformation persuasion infodemic belief-revision debunking generative AI Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Conspiracy theories constitute a critical challenge to public health, as those susceptible to such beliefs are, inter alia, more likely to defy restrictions, disregard protective measures and refuse vaccination [ 1 – 3 ]. Conspiracy theories are defined as “attempts to explain the ultimate causes of significant social and political events and circumstances with claims of secret plots by two or more powerful actors” [ 4 ]. As these theories are especially persuasive when psychological needs for safety and control are threatened, health-related crises such as pandemics provide fertile ground for their dissemination [ 5 ]. The consequence is a self-perpetuating loop: crises foster the spread of conspiracy beliefs, which in turn exacerbate the crises themselves [ 6 ]. To counteract this cycle, researchers from various fields have proposed interventions to mitigate the belief in conspiracy theories [ 7 – 9 ]. Despite these efforts, however, correcting established conspiracy beliefs remains challenging with most interventions yielding small to moderate effects only. A recent exception to this trend is a novel intervention leveraging Large Language Models (LLMs), generative Artificial Intelligence (genAI) models capable of producing synthetic text in response to natural language prompts [ 10 ]. After debating common conspiracies with an LLM-driven chatbot, U.S. study participants’ confidence in these theories was substantially reduced [ 11 ]. Yet, while promising, it remains unclear whether this intervention’s effectiveness generalizes to health-related conspiracy theories. Health-related conspiracy theories and misinformation appear particularly resilient to correction, with debunking effect sizes proving smaller than in other comparable domains [ 12 ]. This might be due to the heightened self-relevance of these beliefs: when the outcomes at stake are more personally relevant, people tend to be less susceptible to persuasive attempts directed at their prior beliefs [ 13 – 15 ]. So, the more self-relevant a conspiracy belief is, the harder it is to dispel. As Costello et al. rely on a randomly recruited general population sample and allow for any self-selected conspiracy to be discussed [ 11 ], it seems unlikely that their participants debated deeply self-relevant conspiracy beliefs with the chatbot. For instance, the most discussed conspiracy in their experiment (15.14%) focused on the assassination of John F. Kennedy. Having occurred over 60 years ago, this event is probably of limited personal relevance to most today. In contrast, health-related conspiracies make up only 8.25% of their sample. This casts doubt on whether Costello et al.’s findings capture the true effect of the LLM-driven debates for those conspiracies most urgently in need of an intervention. Thus, we set out to explore the effect size of LLM-driven debates within a health-related crisis context, specifically with Coronavirus Disease 2019 (COVID-19). Yet, we still predict that LLM-driven debates about health-related conspiracy theories will remain more effective in changing conspiracy beliefs compared to a neutral control condition. Additionally, while LLM corrections of health-related conspiracy theories may be less effective than in other domains, they could still outperform corrections from human sources and thus remain among the most effective interventions available. Specifically, we suggest that individuals might be more open to opposing views when the source is perceived as being artificial rather than human. We therefore propose three mechanisms through which artificial sources may overcome social barriers to persuasion: (1) weaker social responses, (2) reduced face-threats, and (3) machine heuristics. A common response to persuasive attempts is reactance – an amalgam of negative emotional and cognitive reactions (e.g., counterarguing) to perceived threats to freedom [ 16 ]. However, reactance decreases when reducing the number of social cues associated with the source agent [ 17 – 19 ]. While artificial sources are often perceived as social actors, they elicit weaker social responses [ 20 – 23 ]. As a result, individuals interacting with an AI-labelled intervention should be less motivated to counterargue compared to a human-labelled intervention, that is, they should generate less “thoughts that dispute or are inconsistent with the persuasive argument” [ 24 ]. In sum, weaker social responses to artificial sources should translate to lower perceived threat to freedom, less counterarguing and ultimately stronger intervention effects compared to human-labelled interventions. People may also interact with artificial sources without worrying about preserving their social status, that is, without worrying about losing face. According to politeness theory, conspiracy corrections can threaten both positive face (i.e., desire for approval) and negative face (i.e., desire for autonomy) [ 25 , 26 ]. Such face-threats prompt individuals to derogate the source and in turn the message itself [ 27 ]. By shifting the behavioural focus from cooperation towards competition, they “creat[e] a roadblock to agreement” [ 28 ]. However, whether face threats are triggered by corrections or not might depend on the source of the correction [ 29 ]. For instance, people do not fear evaluation (a threat to positive face) by a computer and thus desire less impression management [ 30 , 31 ]. This is mirrored in findings on conversational bots in mental health care and alcohol use counselling: The non-judgemental support provided by LLMs is an incentive that allows users to discuss stigmatized topics without fearing reputational damage [ 32 , 33 ]. As artificial agents are less likely to threaten an individual’s face, resistance to corrections might decrease and ultimately lead to stronger intervention effects compared to human-labelled interventions. Lastly, artificial sources might outperform human-labelled interventions because people might be influenced by machine heuristics, that is, the mental shortcut “wherein we attribute machine characteristics […] when making judgments about the outcome of an interaction” [ 34 ]. Partisan Republicans, for example, rate news and fact-checks credited to AI sources as less biased compared to those attributed to human sources [ 35 , 36 ]. Similarly, corrective messages on COVID-19 attributed to AI rather than human experts trigger less motivated reasoning [ 37 ]. And finally, AI source cues also reduced hostile media biases in a general US sample [ 38 ]. Therefore, conspiracy believers might categorize the artificial source as part of a neutral, non-competitive outgroup that induces less resistance to change and ultimately stronger intervention effects. This advantage of non-human sources for individuals’ general openness to opposing views has been confirmed in experimental research [ 39 ]. Thus, following research on (1) weaker social responses, (2) reduced face-threats, and (3) machine heuristics, we predict that an “AI” source label condition will be more effective in changing conspiracy beliefs compared to a “Human” source label condition. In sum, we hypothesize that while LLMs may be able to reduce beliefs in health-related conspiracy theories, the perceived source could be crucial. Specifically, LLM-generated responses should be more effective if the source is perceived to be non-human. To test these assumptions, we invited participants to debate a COVID-19-related conspiracy with a chatbot labelled as either human or AI. Thereby we expanded research on chatbot-driven interventions in three ways: First, by recruiting a sample with negative attitudes towards the COVID-19 vaccine to increase the likelihood of personally relevant conspiracy beliefs. Then, by exploring effects of the different labels. And finally, by measuring exploratory mediators such as threat to freedom, counterarguing, threat to face, perceived neutrality and perceived competition that might explain the predicted effect. Results Intervention Effectiveness Consistent with our pre-registered hypothesis, belief change was significantly stronger for participants in the AI-label treatment compared to the control ( b = 7.88, 95% CI [5.35,10.42], p < .001, t (370) = 6.11, d = 0.63; see Supplementary Table S2 for full model). Descriptively, participants in the AI-label condition decreased their confidence in their belief by 7.06 percentage points on average ( SD = 16.65), whereas belief confidence in the control condition slightly increased by 0.82 percentage points ( SD = 5.77; see Fig. 1 ). The predicted difference remained significant when excluding participants who failed an attention check and when conducting various robustness checks (see Supplementary Data Analysis and Supplementary Table S3–S6). Replicating Costello et al. [ 11 ], we further dichotomized post-treatment belief confidence at the scale midpoint. This threshold acted as the boundary between endorsement and rejection and thus allowed us to identify participants who transitioned from belief to scepticism. Logistic regression reveals that the AI-label treatment was significantly more effective in converting believers to sceptics ( b = 3.22, 95% CI [1.65, 6.11], p = .001; see Supplementary Table S7). While 12% of the participants in the AI-label treatment shifted toward scepticism, this was the case for fewer than 1% in the control group. Overall, this translates to a number needed to treat (NNT) of 9, indicating that for every 9 people receiving the intervention, one person more turns sceptic of conspiracies (95% CI [ 6 , 15 ]) compared to the control. Source Attribution Investigating the role of source attribution we then proceeded to compare the human-label condition to the AI-label and the control (see Fig. 1 ). In both cases we detected significant differences: Debates under the human-label results in a 13.76 percentage points larger confidence drop compared to the control (95% CI [-17.05, -10.47], t (551) = -8.21, p < .001, d = -0.87, see Supplementary Table S8). Contrary to our hypothesis, however, this drop was also 5.88 percentage points larger than in the AI-label condition (95% CI [-9.17, -2.58], t (551) = -3.50, p < .001, d = -0.30). Put differently, the human-label was nearly twice as effective as the AI-label. Again, these differences remained significant when excluding participants who failed the attention check and conducting various robustness checks (see Supplementary Data Analysis and Supplementary Table S9–S12). In terms of converting believers, 18% of the participants in the human-label condition fell below the scale midpoint after the treatment. For every 6 individuals receiving the intervention, one more becomes sceptic of conspiracies (95% CI [ 4 , 9 ]) compared to the control. However, compared to the AI-label, this difference in belief conversions was not significant ( b = 0.46, 95% CI [ -0.12, 1.06], p = .13; see Supplementary Table S13). Exploratory Analysis of Process Variables We now turn to how participants perceived the discussion with the LLM (see Fig. 2 ). The first pathway we proposed was via weaker social responses like reduced threat to freedom and consequently less counterarguing. Contrary to this account, however, individuals perceived the AI-label to be more threatening to freedom ( t (360) = 2.51, p = .01, d = 0.26) compared to the human-label. Furthermore, there was no evidence that counterarguing differed between conditions; t (360) = 1.87, p = .06, d = 0.2. The second pathway we explored was based on decreased face-threat. Yet, once more, the AI-label was seen as more threatening; t (365) = 2.26, p = 0.02, d = 0.24. Finally, as a third pathway, we explored whether machine heuristics increase perceived neutrality and decrease perceived competition. Among AI-label participants, correlations supported this mechanism: machine heuristics were moderately associated with both perceived neutrality ( r s (183) = .51, p < .001) and belief change ( r s (183) = .33, p < .001). However, this effect did not translate to the condition level. On the contrary, the AI-label intervention was perceived as less neutral than the human-label; t (365) = -4.94, p < .001, d = -0.52. While the correlation between machine heuristics and perceived competition was also significant ( r s (183) = .15, p = .05), the direction was reversed. Additionally, there was no evidence that the AI-label was seen as less competitive than the human-label; t (359) = 0.59, p = .56, d = 0.06. Given the fact that the human-label intervention was perceived as less threatening and more neutral, we proceeded by exploring whether any of these differences could explain the fact that the AI-label intervention was more persuasive than the human-label intervention. Specifically, we performed a parallel mediation analysis with 10,000 bootstrap samples comparing the intervention conditions on post-treatment confidence (controlling for baseline confidence) with threat to freedom, threat to face and neutrality as mediators (see Supplementary Table S16). In the human-label condition participants perceived their discussion partner as more neutral compared to the AI-label condition ( b = 0.68, p < .001). This higher neutrality, in turn, was associated with a larger confidence reduction ( b = -5.65, p < .001). The resulting indirect effect via neutrality was statistically significant ( b = − 3.86, 95% CI [-5.91, -2.14]). Threat to freedom was lower in the human-label condition ( b = -0.47, p = .008) and positively associated with post-treatment confidence ( b = 1.60, p = .05). However, there was no evidence of an indirect effect ( b = -0.75, 95% CI [-1.87, 0.02]). Threat to face had no significant association with post-treatment confidence ( b = -0.77, p = .37) and there was no evidence of an indirect effect ( b = 0.26, 95% CI [-0.28, 1.11]). Finally, to explore potential differences between perceptions of computer-mediated vs human-computer communication, we also explored whether the perceived agency of the discussion partner influenced belief change for those in the AI-label condition. However, no significant association was found ( r s (183) = .02, p = .80). Discussion Despite increasing interest in the use of conversational AI for persuasive purposes, the effectiveness of chatbots in the health communication domain remains largely unexplored. We addressed this gap by examining whether LLM-based conversational interventions can reduce health-related conspiracy beliefs, a particularly challenging and high-stakes context, and by testing source attribution as a key mechanism shaping their effectiveness. Our findings demonstrate that conversations with chatbots are indeed effective at fighting health-related conspiracy theories. The intervention produced a robust reduction in conspiracy beliefs. Even though effect sizes were at the lower end of the range reported by Costello et al., they align well with their results on “particularly entrenched beliefs” [ 11 ]. Moreover, compared to conventional debunking effects in health communication ( d = 0.40, 95% CI [0.25, 0.55]; [ 40 ]), chatbot induced reduction in conspiracy beliefs can be considered substantial. This illustrates that LLM-driven interventions can be effective even in highly self-relevant domains such as health. However, contrary to our expectations, this reduction occurred not because of but despite participants’ awareness of their artificial conversation partner. Post-debate, participants in the AI-label condition reported 7 percentage points less confidence in the target conspiracy. In contrast, those believing to debate another human indicated a confidence drop nearly twice as strong (13 percentage points). This difference appears to be mediated by the perceived neutrality of the conversation partner. Additionally and inconsistent with machine heuristics and previous work on the perceived bias in artificial versus human sources [ 35 , 36 ], artificial sources were also seen as more threatening to both autonomy and face compared to those labelled as human. This study adds to the growing body of evidence for the persuasive power of LLMs. It illustrates how LLMs can be used to deliver effective attitude-change communication at scale. Yet, it also reveals that source attribution appears to matter for health communication: People are influenced more easily when they believe to be debating with humans rather than LLMs. This rules out source attribution as the mechanism driving the effectiveness of these interventions. Trying to reconcile these findings with Boissin et al., who observed no meaningful difference between human and AI-labels in a similar setup [ 41 ], there are two relevant factors to look at: The heightened level of self-relevance for the health domain and the generalizability of insights on source-based reasoning. In established dual-process theories of persuasion, heightened self-relevance encourages argument-based processing and therefore diminishes the role of source characteristics [ 42 , 43 ]. Yet, considering such context clues might not always be a sign of peripheral processing. Instead, Lee’s integrative theory of human-machine communication argues that using source information in addition to message content actually signals a more effortful cognitive approach [ 44 ]. Put differently, people discussing things that matter to them use source cues not as a shortcut, but as an additional means of reaching a more accurate judgement. This elevates the importance of source cues and brings us to our second variable: their generalizability. Computers are often perceived as unbiased, objective and less threatening [ 34 ]. However, as our participants perceived the AI-label as less neutral and more threatening to freedom and face, our data suggests that these assumptions might be context specific. One reason for these findings might be the changing public narrative around AI. Disappointing experiences with LLMs in everyday use, the rise in negative news coverage following the release of ChatGPT [ 45 ], concerns about authenticity and slop, and the heightened awareness of manipulative and fallible chatbots might have shifted the valence of how people perceive LLMs by the time of our data collection. In addition, it remains unclear to what degree machine heuristics (as originally conceptualized: [ 46 ]) generalize to LLM-driven chatbots in the first place. In fact, the interactive nature and the affordances of the technology differs strongly from the work that established machine heuristics on news recommendation systems [ 46 , 47 ]. Thus, source-based evaluation for LLMs might not follow the same mechanisms. Future research should therefore attempt to test whether general machine heuristics apply to LLMs at all, or whether the technology has given rise to new LLM heuristics. To fully understand potential shifts in machine heuristics, longitudinal survey studies exploring changes in people’s perceptions of LLMs over time are necessary. As our study failed to identify a mechanism for the performance of LLM-driven interventions, additional studies should also probe into alternative explanations. This includes, inter alia, information density, the interactive presentation of information, or both politeness and confidence of the model. It is of course difficult to determine the direction of the effect, as we measured machine heuristics after the interaction: did weaker machine heuristics lead to smaller belief changes, or did participants report more negative responses because the LLM argued against their deeply held beliefs? Recent findings on increased negative attitudes towards AI after being outsmarted by an artificial agent might point to the latter [ 48 ]. Additionally, as we utilized identity threat as a proxy for positive face threat, we might have missed nuances of the concept relevant for the intervention at hand. Given the mediating effect of neutrality, the question also arises whether the difference between AI-label and human-label was driven by the lower perceived neutrality of the AI agent, or the heightened neutrality of our three deception framings (i.e., a communication student, a journalist or a social worker in training). A third limitation is the high deception-based screen-out quota for our human-label condition. It is possible that by excluding those who saw through the deception, we incidentally included only those participants most gullible to synthetic influence in the first place. Finally, in our attempt to increase the credibility of the deception while controlling for differences in message content we instructed the LLM to paraphrase its outputs in a human-like way across both intervention conditions. As a result, generated messages were less polite and conflict averse. Therefore, the differences found could be attributed to the violation of expectations in the AI-label condition. However, given that Boissin et al. found no interaction between source label and prompt type when utilizing a similar approach [ 41 ], this seems unlikely. Our findings indicate that LLMs are indeed effective persuaders in highly self-relevant domains. While this is good news for those trying to shape effective interventions in the health contexts, it also unveils equivalent risks of misuse and manipulation. Unlike other epistemic threats like disinformation and propaganda, chatbots are not in short demand: Day by day, millions of users worldwide choose to interact voluntarily with LLMs accountable only to the corporations owning them. If even one conversation with such a model can fundamentally shift deeply held beliefs, this raises important questions about the long-term consequences of repeated exposure. Potentially, these interactions could erode societal core values like the support for democracy or trust in science and modern medicine. Effects that, by propagating along individual and social networks, can cascade beyond single beliefs and single believers [ 49 ]. Yet, our results also show, that people are more open to human rather than artificial guidance. And while this creates further incentives for malign actors to hide the true origin of synthetic messages, it also illustrates that people still value – what they believe to be – genuine human interaction. Methods & Materials This pre-registered online experiment ( https://osf.io/yf6qs/overview ) follows a 3 (Treatment: Control vs “Human” label vs “AI” label; between) x 2 (Time: Pre-Treatment vs Post-Treatment; within) mixed design with belief-change (the delta between pre- and post-treatment measures) as the primary dependent variable. This study received approval by Radboud University’s Ethics Assessment Committee Humanities (2025–5907) and was conducted in accordance with its guidelines and regulations. Participants & Recruitment US and UK based participants with high approval rates (> 95%) were recruited via Prolific and compensated with £ 6 per hour. Due to the deceptive nature of the study (i.e., some participants were told they were interacting with another human when they were not), we relied on Prolific pre-screeners to target users that a) agreed on participating in deceptive studies and b) had some prior experience with chatbots (indicating at least a minimal level of acceptance for interacting with generative AI). Furthermore, we only targeted individuals with negative attitudes towards the COVID-19 vaccine. Our target sample size of 546 was based on a priori power analysis utilizing a Monte-Carlo simulation for a linear model with the belief delta as the dependent variable and dummy variables for both the control ( d = -0.53, based on the more conservative effect size identified by Costello et al. [ 11 ]) and the human condition ( d = -0.26) as predictors with 80% power at alpha = 0.05. During the experiment, participants were automatically excluded based on pre-registered criteria (e.g., failed to share a valid conspiracy belief; see Procedure). Furthermore, we sequentially reviewed whether they saw through the deception. In both cases, additional individuals were recruited to meet the required target sample size. Out of 876 participants initially recruited, 754 completed the experiment. Following our screen-out process (see Fig. 3 ), 554 individuals remained in the final sample (see Supplementary Table S1 ). Of those, 361 (65%) were women, 192 (35%) men and 1 (< 1%) non-binary. All of them were between 18 and 78 years old ( M = 42.96, SD = 12.51). As their highest level of education, most participants indicated finishing high school (50%) followed by Bachelor’s (38%) and Master’s (11%) degrees. Procedure Following the original design by Costello et al. [ 11 ], the experiment was conducted online via Qualtrics. Utilizing the randomization element, participants were assigned to one of three conditions: A control condition, the “AI” label treatment, and the “human” label treatment. After obtaining informed consent, participants across all conditions were instructed to describe a conspiracy theory endorsed by them. In a follow-up question, they further expanded on their previous response by including specific evidence and sources influencing their perspective. Utilizing the OpenRouter API, these responses were piped into an LLM (i.e., Claude Sonnet 4.0; prompt adapted from [ 41 ]; see Supplementary Table S15 for all prompts) instructed to assess whether the described belief qualifies as a conspiracy. If this threshold was not met, participants were asked for another conspiracy belief once before being screened out. Evaluating model performance post-hoc based on a balanced sample (i.e. 10% of the dataset, half classified as conspiracy and half not classified as such by the LLM), we observed substantial alignment with human-labels ( κ = 0.73). For valid conspiracies, the provided description was then summarized by the model into a single sentence. In the “human” label condition, this process was artificially delayed facilitating the deception. Then, subjects were asked to indicate their belief in the summarized conspiracy on a 101-point rating scale as well as whether the summary accurately captured their description. At this point, individuals rating their belief beneath the midpoint or the summary as inaccurate were automatically excluded from the study. For the “AI” label treatment, subjects were instructed to discuss their conspiracy beliefs with an AI agent. For the “human” label treatment, they were informed about conversing with either a communication student, a journalist or a social worker in training. To facilitate the deception in the “human” label condition, we included elaborate backstories (e.g., training journalists on how to engage around complicated topics), artificial delays and occasional spelling mistakes. Additionally, participation in the experiment was only available during “office hours”. In both treatment conditions, participants had three rounds of debate with an instance of Claude Sonnet 4.0. The LLM was prompted to reduce the user’s belief in the conspiracy based on their description of it as well as their previous responses and the rating on the belief scale. In both conditions, a second instance of the same model was then instructed to regenerate the answer in a human-like tone. Finally, the “humanified” response was evaluated by two more instances to check, whether the model refused to follow any instructions or admitted to being an LLM (for ethical reasons we explicitly prompted the model to admit to being an LLM when asked). In those cases, the initial response was displayed to the participant (see Fig. 4 ). For the control condition, participants were asked to discuss an unrelated topic – i.e., their previous experiences with the fire department – with the LLM. Post-treatment, participants were once more asked to rate their confidence in the initial machine-generated summary and to answer items on various process variables. Finally, all participants were debriefed and offered the opportunity to remove their data from the analysis. Measures Primary Outcome Variable Before and after interaction with the chatbot, participants were asked to rate their confidence in the target conspiracy theory on a 101-point scale. The item “On a scale of 0% to 100%, please indicate your level of confidence that this statement is true.” was adapted from [ 11 ]. Process Variables To identify possible mechanisms for the effects we collected data on various process variables. Measures and items were presented in randomized order. All items were rated on a 7-point Likert scale (1 = strongly disagree to 7 = strongly agree). For all scales mean scores were calculated. Counterarguing. The measure for counterarguing consisted of 4 items adapted from [ 50 , 51 ] (Cronbach’s α = .71). Examples are “While reading the responses, I sometimes found myself thinking of ways I disagreed with what was being presented.” and “I found myself looking for flaws in the way information was presented in the response.” Threat to Freedom. Threat to freedom was measured utilizing 5 items adapted from [ 50 , 52 ] (Cronbach’s α = .94). Specific wording for each item depends on the condition. Examples are “The AI / other person tried to make a decision for me.” and “The AI / other person tried to force their opinion on me.” Threat to Face. To assess perceived threat to face we adapted 4 items utilized by [ 50 ] (Cronbach’s α = .89). Although originally intended to measure social identity threat, these measures capture threats to self-image and as such align well with threats to positive face as conceptualized by [ 25 ]. Examples are “The responses I received undermined my sense of self-worth.” and “The responses I received made me feel less unique as a person.” Source Neutrality. For source neutrality 4 items were adapted from [ 53 , 54 ] (Cronbach’s α = .76). Specific wording for each item depends on the condition. Examples are “The AI’s / other person’s position on this issue is rooted in an objective analysis of the issue.” and “The AI’s / other person’s reasoning about this issue is influenced by bias.” Perceived Competition. We relied on 3 items adapted from [ 55 ] to measure perceived competition (Cronbach’s α = .91). Specific wording for each item depends on the condition. Examples are “I view the AI / other person as a competitor.” and “Competition towards the AI / other person is important to me.” Additional Measures Perceived Agency. To measure the perceived agency of the chatbot across both the AI-label and the control condition, we utilized 5 items from [ 56 ] (Cronbach’s α = .86). All items were presented randomized and rated on a 7-point Likert scale (1 = strongly disagree to 7 = strongly agree). Mean scores were calculated. Examples are “The AI intentionally completed its responses.” and “The AI could have chosen not to respond in this way.” Machine Heuristics. For machine heuristics across both the AI-label and the control condition we adapted 7 items from [ 57 ] (Cronbach’s α = .97). All items were presented randomized and rated on a 7-point Likert scale (1 = strongly disagree to 7 = strongly agree). Mean scores were calculated. Examples are “I trust AI more than humans when discussing my beliefs.” and “AI is better than humans at discussing my beliefs.” Demographics. Finally, we also collected demographic data on age, gender, and education. Deviation from Pre-Registration For enhanced clarity we reversed the presentation order of hypotheses from our pre-registration. Additionally, to assess the effectiveness of the baseline intervention we focused on data from the control and the AI-label treatment only. To increase interpretability for the analysis of source effects we changed the reference group from the AI-label to the human-label intervention. This allowed us to test two relevant comparisons (AI vs. human and human vs. control) in one model rather than using two models and was thus considered superior. However, results of the original model are in line with our findings (see Supplementary Table S14). Declarations Acknowledgements Acknowledgements will be described on publication. Additional Information Competing interests The authors declare no competing interests. Funding There is no funding associated with the work featured in this article. Author Contribution Supervision: H.S., P.S.;Conceptualization: P.B., Y.v.d.S., H.S., P.S.;Study Design: P.B., Y.v.d.S., P.S.;Data Collection: P.B.;Data Analysis: P.B., Y.v.d.S., P.S.;Writing – original draft: P.B.;Writing – review & editing: P.B., Y.v.d.S., H.S., P.S.; Acknowledgement The authors thank Mark Dingemanse for his expertise and guidance. Data Availability The data supporting the findings of this study was deposited via OSF and is available at https://osf.io/su9kp/overview?view_only=7a40bc5d15cd4d308b10b9c0cc97d600. References Regazzi, L., Lontano, A., Cadeddu, C., Di Padova, P. & Rosano, A. Conspiracy beliefs, COVID-19 vaccine uptake and adherence to public health interventions during the pandemic in Europe. Eur. J. Public. Health . 33 , 717–724 (2023). Romer, D. & Jamieson, K. H. Conspiracy theories as barriers to controlling the spread of COVID-19 in the U.S. Soc. Sci. Med. 263 , 113356 (2020). Zilinsky, J. & Theocharis, Y. Conspiracism and government distrust predict COVID-19 vaccine refusal. Humanit. Soc. Sci. Commun. 12 , 1002 (2025). Douglas, K. M. et al. Understanding Conspiracy Theories. Polit Psychol. 40 , 3–35 (2019). Douglas, K. M., Sutton, R. M. & Cichocka, A. The Psychology of Conspiracy Theories. Curr. Dir. Psychol. Sci. 26 , 538–542 (2017). Van Mulukom, V. et al. Antecedents and consequences of COVID-19 conspiracy beliefs: A systematic review. Soc. Sci. Med. 301 , 114912 (2022). O’Mahony, C., Brassil, M., Murphy, G. & Linehan, C. The efficacy of interventions in reducing belief in conspiracy theories: A systematic review. PLOS ONE . 18 , e0280902 (2023). Taubert, F., Meyer-Hoeven, G., Schmid, P., Gerdes, P. & Betsch, C. Conspiracy narratives and vaccine hesitancy: a scoping review of prevalence, impact, and interventions. BMC Public. Health . 24 , 3325 (2024). Stasielowicz, L. The Effectiveness of Interventions Addressing Conspiracy Beliefs: A Meta-Analysis. Eur. J. Soc. Psychol. 56 , 275–291 (2026). Banh, L. & Strobel, G. Generative artificial intelligence. Electron. Mark. 33 , 63 (2023). Costello, T. H., Pennycook, G. & Rand, D. G. Durably reducing conspiracy beliefs through dialogues with AI. Science 385 , eadq1814 (2024). Chan, M. S. & Albarracín, D. A meta-analysis of correction effects in science-relevant misinformation. Nat. Hum. Behav. 7 , 1514–1525 (2023). Darke, P. R. & Chaiken, S. The pursuit of self-interest: Self-interest bias in attitude judgment and persuasion. J. Pers. Soc. Psychol. 89 , 864–883 (2005). Falk, E., Scholz, C. & Persuasion Influence, and Value: Perspectives from Communication and Social Neuroscience. Annu. Rev. Psychol. 69 , 329–356 (2018). Johnson, B. T. & Eagly, A. H. Effects of involvement on persuasion: A meta-analysis. Psychol. Bull. 106 , 290–314 (1989). Steindl, C., Jonas, E., Sittenthaler, S., Traut-Mattausch, E. & Greenberg, J. Understanding Psychological Reactance: New Developments and Findings. Z. Für Psychol. 223 , 205–214 (2015). Ghazali, A. S., Ham, J., Barakova, E. & Markopoulos, P. The influence of social cues in persuasive social robots on psychological reactance and compliance. Comput. Hum. Behav. 87 , 58–65 (2018). Roubroeks, M., Ham, J. & Midden, C. When Artificial Social Agents Try to Persuade People: The Role of Social Agency on the Occurrence of Psychological Reactance. Int. J. Soc. Robot . 3 , 155–165 (2011). Roubroeks, M., Midden, C. & Ham, J. Does it make a difference who tells you what to do? exploring the effect of social agency on psychological reactance. in Proceedings of the 4th International Conference on Persuasive Technology 1–6 (ACM, Claremont California USA, (2009). 10.1145/1541948.1541970 Krämer, N. C. Social Communicative Effects of a Virtual Program Guide. in Intelligent Virtual Agents (ed (eds Panayiotopoulos, T. et al.) vol. 3661 442–453 (Springer Berlin Heidelberg, Berlin, Heidelberg, (2005). Morkes, J., Kernal, H. K. & Nass, C. Effects of Humor in Task-Oriented Human-Computer Interaction and Computer-Mediated Communication: A Direct Test of SRCT Theory. Human–Computer Interact. 14 , 395–435 (1999). Nass, C. & Moon, Y. Machines and Mindlessness: Social Responses to Computers. J. Soc. Issues . 56 , 81–103 (2000). Nass, C., Steuer, J. & Voices Boxes, and Sources of Messages.: Computers and Social Actors. Hum. Commun. Res. 19 , 504–527 (1993). Slater, M. D. & Rouner, D. Entertainment?Education and Elaboration Likelihood: Understanding the Processing of Narrative Persuasion. Commun. Theory . 12 , 173–191 (2002). Brown, P. & Levinson, S. C. Politeness: Some Universals in Language Usage (Cambridge University Press, 1987). 10.1017/CBO9780511813085 Schmid, P. Debunking health misinformation with empathy. Curr. Opin. Psychol. 67 , 102213 (2026). Jenkins, M. & Dragojevic, M. Explaining the Process of Resistance to Persuasion: A Politeness Theory-Based Approach. Commun. Res. 40 , 559–590 (2013). White, J. B., Tynan, R., Galinsky, A. D. & Thompson, L. Face threat sensitivity in negotiation: Roadblock to agreement and joint gain. Organ. Behav. Hum. Decis. Process. 94 , 102–124 (2004). Babel, F., Hock, P., Kraus, J. & Baumann, M. Human-Robot Conflict Resolution at an Elevator - The Effect of Robot Type, Request Politeness and Modality. in 17th ACM/IEEE International Conference on Human-Robot Interaction (HRI) 693–697 (IEEE, Sapporo, Japan, 2022). 693–697 (IEEE, Sapporo, Japan, 2022). (2022). 10.1109/HRI53351.2022.9889387 Gratch, J., Lucas, G. M., King, A. A. & Morency, L. P. It’s only a computer: the impact of human-agent interaction in clinical interviews. in Proceedings of the International Conference on Autonomous Agents and Multi-Agent Systems 85–92 (International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2014). 85–92 (International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2014). (2014). Weisband, S. & Kiesler, S. Self disclosure on computer forms: meta-analysis and implications. in Proceedings of the SIGCHI conference on Human factors in computing systems common ground - CHI ’96 3–10ACM Press, Vancouver, British Columbia, Canada, (1996). 10.1145/238386.238387 Ma, Z., Mei, Y. & Su, Z. Understanding the benefits and challenges of using large language model-based conversational agents for mental well-being support. in AMIA Annual Symposium Proceedings vol. 1105 (American Medical Informatics Association, 2023). (2023). Steenstra, I., Nouraei, F., Arjmand, M. & Bickmore, T. W. Virtual Agents for Alcohol Use Counseling: Exploring LLM-Powered Motivational Interviewing. Preprint at (2024). https://doi.org/10.1145/3652988.3673932 Sundar, S. S., Kim, J. M. & Heuristic When We Trust Computers More than Humans with Our Personal Information. in Proceedings of the CHI Conference on Human Factors in Computing Systems 1–9 (ACM, Glasgow Scotland Uk, 2019). 1–9 (ACM, Glasgow Scotland Uk, 2019). (2019). 10.1145/3290605.3300768 Chae, J. H. & Tewksbury, D. Perceiving AI intervention does not compromise the persuasive effect of fact-checking. New. Media Soc. 28 , 191–211 (2024). Waddell, T. F. Can an Algorithm Reduce the Perceived Bias of News? Testing the Effect of Machine Attribution on News Readers’ Evaluations of Bias, Anthropomorphism, and Credibility. J. Mass. Commun. Q. 96 , 82–100 (2019). Moon, W. K., Chung, M. & Jones-Jang, S. M. How Can We Fight Partisan Biases in the COVID-19 Pandemic? AI Source Labels on Fact-checking Messages Reduce Motivated Reasoning. Mass. Commun. Soc. 26 , 646–670 (2023). Craig, M. J. A. & Choi, M. The role of affective and cognitive involvement in the mitigating effects of AI source cues on hostile media bias. Telemat Inf. 88 , 102097 (2024). Lu, L., Tormala, Z. L. & Duhachek, A. How AI sources can increase openness to opposing views. Sci. Rep. 15 , 17170 (2025). Walter, N., Brooks, J. J., Saucier, C. J. & Suresh, S. Evaluating the Impact of Attempts to Correct Health Misinformation on Social Media: A Meta-Analysis. Health Commun. 36 , 1776–1784 (2021). Boissin, E., Costello, T. H., Spinoza-Martín, D., Rand, D. G. & Pennycook, G. Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human. PNAS Nexus . 4 , pgaf325 (2025). Chaiken, S. Heuristic versus systematic information processing and the use of source versus message cues in persuasion. J. Pers. Soc. Psychol. 39 , 752–766 (1980). Petty, R. E. & Cacioppo, J. T. The Elaboration Likelihood Model of Persuasion. in Advances Experimental Social Psychology 19 123–205 (Elsevier, 1986). Lee, E. J. Minding the source: toward an integrative theory of human–machine communication. Hum. Commun. Res. 50 , 184–193 (2024). Jain, A. & Ranganathan, S. A. Longitudinal Analysis of Artificial Intelligence Coverage in Technology-Focused News Media Using Latent Dirichlet Allocation and Sentiment Analysis. J. Media . 6 , 176 (2025). Sundar, S. S. & Nass, C. Conceptualizing sources in online news. J. Commun. 51 , 52–72 (2001). Anderl, C. et al. Conversational presentation mode increases credibility judgements during information search with ChatGPT. Sci. Rep. 14 , 17127 (2024). Schaap, G., Van De Sande, Y. & Schraffenberger, H. Outperformed by AI: Interacting with Superhuman AI Changes the Way We Perceive Ourselves. in Extended Abstracts of the CHI Conference on Human Factors in Computing Systems 1–7 (ACM, Honolulu HI USA, (2024). 10.1145/3613905.3650961 Guingrich, R. E., Mehta, D. & Bhatt, U. Belief Offloading in Human-AI Interaction. Preprint at (2026). https://doi.org/10.48550/arXiv.2602.08754 Ma, Y. & Hmielowski, J. D. Are You Threatening Me? Identity Threat, Resistance to Persuasion, and Boomerang Effects in Environmental Communication. Environ. Commun. 16 , 225–242 (2022). Nabi, R. L., Moyer-Gusé, E. & Byrne, S. All Joking Aside: A Serious Investigation into the Persuasive Effect of Funny Social Issue Messages. Commun. Monogr. 74 , 29–54 (2007). Dillard, J. P. & Shen, L. On the Nature of Reactance and its Role in Persuasive Health Communication. Commun. Monogr. 72 , 144–168 (2005). Kennedy, K. A. & Pronin, E. When Disagreement Gets Ugly: Perceptions of Bias and the Escalation of Conflict. Pers. Soc. Psychol. Bull. 34 , 833–848 (2008). Ziegler, R. & Diehl, M. Mood and Multiple Source Characteristics: Mood Congruency of Source Consensus Status and Source Trustworthiness as Determinants of Message Scrutiny. Pers. Soc. Psychol. Bull. 37 , 1016–1030 (2011). Yip, J. A., Schweitzer, M. E. & Nurmohamed, S. Trash-talking: Competitive incivility motivates rivalry, performance, and unethical behavior. Organ. Behav. Hum. Decis. Process. 144 , 125–144 (2018). Korman, J., Harrison, A., McCurry, M. & Trafton, G. Beyond Programming: Can Robots’ Norm-Violating Actions Elicit Mental State Attributions? in 14th ACM/IEEE International Conference on Human-Robot Interaction (HRI) 530–531 (2019). 530–531 (2019). (2019). 10.1109/HRI.2019.8673293 Yang, H. & Sundar, S. S. Machine heuristic: concept explication and development of a measurement scale. J. Comput. -Mediat Commun. 29 , zmae019 (2024). Additional Declarations No competing interests reported. Supplementary Files Supplement.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 13 May, 2026 Reviews received at journal 08 May, 2026 Reviews received at journal 08 May, 2026 Reviewers agreed at journal 25 Apr, 2026 Reviewers agreed at journal 24 Apr, 2026 Reviewers agreed at journal 23 Apr, 2026 Reviewers invited by journal 23 Apr, 2026 Editor assigned by journal 23 Apr, 2026 Submission checks completed at journal 19 Apr, 2026 First submitted to journal 19 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9428259","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":633625295,"identity":"f954bfdb-ad8c-4a2b-a122-9bd14d36848e","order_by":0,"name":"Paul Ballot","email":"","orcid":"","institution":"Radboud University","correspondingAuthor":false,"prefix":"","firstName":"Paul","middleName":"","lastName":"Ballot","suffix":""},{"id":633625296,"identity":"176e320a-febb-4289-a06f-92b8a57d6bd4","order_by":1,"name":"Yana van de Sande","email":"","orcid":"","institution":"Radboud University","correspondingAuthor":false,"prefix":"","firstName":"Yana","middleName":"van","lastName":"de Sande","suffix":""},{"id":633625297,"identity":"82bfc9e5-5fcd-4a92-acfb-760e78fd673a","order_by":2,"name":"Hanna Schraffenberger","email":"","orcid":"","institution":"Radboud University","correspondingAuthor":false,"prefix":"","firstName":"Hanna","middleName":"","lastName":"Schraffenberger","suffix":""},{"id":633625300,"identity":"65c1b6e8-3718-454e-acca-0c17abf3285a","order_by":3,"name":"Philipp Schmid","email":"data:image/png;base64,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","orcid":"","institution":"University of Erfurt","correspondingAuthor":true,"prefix":"","firstName":"Philipp","middleName":"","lastName":"Schmid","suffix":""}],"badges":[],"createdAt":"2026-04-15 14:23:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9428259/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9428259/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108736389,"identity":"61449670-12ef-4e4c-b4f8-4d1b84427278","added_by":"auto","created_at":"2026-05-07 20:15:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":530067,"visible":true,"origin":"","legend":"\u003cp\u003ePre-Post Confidence Change by Condition. Error bars represent standard deviations for the raw values and bootstrapped 95% CIs for the mean differences.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/75c11d9c46c9a66f01abac08.png"},{"id":108807188,"identity":"4e352e08-100e-4d6e-aaa9-7684f29fe924","added_by":"auto","created_at":"2026-05-08 15:30:17","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":986145,"visible":true,"origin":"","legend":"\u003cp\u003eProcess Variables by Condition\u003cstrong\u003e. \u003c/strong\u003eError bars represent 95% CIs.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/a16b12b9f02d5f4e39e38173.png"},{"id":108806353,"identity":"c912e1cc-7415-45e3-8d74-d71b62f8b31f","added_by":"auto","created_at":"2026-05-08 15:28:20","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":243593,"visible":true,"origin":"","legend":"\u003cp\u003eFlow of Participants\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/9f4e989217dcbe794c463e1e.png"},{"id":108736391,"identity":"0ed31f3a-f1f6-44ed-93c7-e1a7888f4d6b","added_by":"auto","created_at":"2026-05-07 20:15:39","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":330370,"visible":true,"origin":"","legend":"\u003cp\u003eResponse Generation\u003cstrong\u003e. \u003c/strong\u003eAfter paraphrasing the response generated by the first model call, another model evaluates the response based on whether LLM2 complied with the instructions. If so, the output by LLM2 is presented to the user. Otherwise, the output by LLM1 is used.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/d45e975fdb2a04dcb0d5b700.png"},{"id":108810299,"identity":"4bd9bf3d-26b4-4a7e-a9db-4cc3978e6f74","added_by":"auto","created_at":"2026-05-08 15:58:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2419604,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/fd9032ff-2158-4aba-be3b-ee22feb871b7.pdf"},{"id":108806812,"identity":"7ce0fdaa-5d92-4256-aa31-00d045884077","added_by":"auto","created_at":"2026-05-08 15:29:31","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":61102,"visible":true,"origin":"","legend":"","description":"","filename":"Supplement.docx","url":"https://assets-eu.researchsquare.com/files/rs-9428259/v1/08949e0834d4fb623f8bf905.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Chatbots Reduce Health-Related Conspiracy Beliefs Not Because of but Despite Being Perceived as AI","fulltext":[{"header":"Introduction","content":"\u003cp\u003eConspiracy theories constitute a critical challenge to public health, as those susceptible to such beliefs are, inter alia, more likely to defy restrictions, disregard protective measures and refuse vaccination [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Conspiracy theories are defined as \u0026ldquo;attempts to explain the ultimate causes of significant social and political events and circumstances with claims of secret plots by two or more powerful actors\u0026rdquo; [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. As these theories are especially persuasive when psychological needs for safety and control are threatened, health-related crises such as pandemics provide fertile ground for their dissemination [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The consequence is a self-perpetuating loop: crises foster the spread of conspiracy beliefs, which in turn exacerbate the crises themselves [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. To counteract this cycle, researchers from various fields have proposed interventions to mitigate the belief in conspiracy theories [\u003cspan additionalcitationids=\"CR8\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Despite these efforts, however, correcting established conspiracy beliefs remains challenging with most interventions yielding small to moderate effects only. A recent exception to this trend is a novel intervention leveraging Large Language Models (LLMs), generative Artificial Intelligence (genAI) models capable of producing synthetic text in response to natural language prompts [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. After debating common conspiracies with an LLM-driven chatbot, U.S. study participants\u0026rsquo; confidence in these theories was substantially reduced [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Yet, while promising, it remains unclear whether this intervention\u0026rsquo;s effectiveness generalizes to health-related conspiracy theories.\u003c/p\u003e \u003cp\u003eHealth-related conspiracy theories and misinformation appear particularly resilient to correction, with debunking effect sizes proving smaller than in other comparable domains [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This might be due to the heightened self-relevance of these beliefs: when the outcomes at stake are more personally relevant, people tend to be less susceptible to persuasive attempts directed at their prior beliefs [\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. So, the more self-relevant a conspiracy belief is, the harder it is to dispel. As Costello et al. rely on a randomly recruited general population sample and allow for any self-selected conspiracy to be discussed [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], it seems unlikely that their participants debated deeply self-relevant conspiracy beliefs with the chatbot. For instance, the most discussed conspiracy in their experiment (15.14%) focused on the assassination of John F. Kennedy. Having occurred over 60 years ago, this event is probably of limited personal relevance to most today. In contrast, health-related conspiracies make up only 8.25% of their sample. This casts doubt on whether Costello et al.\u0026rsquo;s findings capture the true effect of the LLM-driven debates for those conspiracies most urgently in need of an intervention. Thus, we set out to explore the effect size of LLM-driven debates within a health-related crisis context, specifically with Coronavirus Disease 2019 (COVID-19). Yet, we still predict that LLM-driven debates about health-related conspiracy theories will remain more effective in changing conspiracy beliefs compared to a neutral control condition.\u003c/p\u003e \u003cp\u003eAdditionally, while LLM corrections of health-related conspiracy theories may be less effective than in other domains, they could still outperform corrections from human sources and thus remain among the most effective interventions available. Specifically, we suggest that individuals might be more open to opposing views when the source is perceived as being artificial rather than human. We therefore propose three mechanisms through which artificial sources may overcome social barriers to persuasion: (1) weaker social responses, (2) reduced face-threats, and (3) machine heuristics.\u003c/p\u003e \u003cp\u003eA common response to persuasive attempts is reactance \u0026ndash; an amalgam of negative emotional and cognitive reactions (e.g., counterarguing) to perceived threats to freedom [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. However, reactance decreases when reducing the number of social cues associated with the source agent [\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. While artificial sources are often perceived as social actors, they elicit weaker social responses [\u003cspan additionalcitationids=\"CR21 CR22\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. As a result, individuals interacting with an AI-labelled intervention should be less motivated to counterargue compared to a human-labelled intervention, that is, they should generate less \u0026ldquo;thoughts that dispute or are inconsistent with the persuasive argument\u0026rdquo; [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. In sum, weaker social responses to artificial sources should translate to lower perceived threat to freedom, less counterarguing and ultimately stronger intervention effects compared to human-labelled interventions.\u003c/p\u003e \u003cp\u003ePeople may also interact with artificial sources without worrying about preserving their social status, that is, without worrying about losing face. According to politeness theory, conspiracy corrections can threaten both positive face (i.e., desire for approval) and negative face (i.e., desire for autonomy) [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Such face-threats prompt individuals to derogate the source and in turn the message itself [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. By shifting the behavioural focus from cooperation towards competition, they \u0026ldquo;creat[e] a roadblock to agreement\u0026rdquo; [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. However, whether face threats are triggered by corrections or not might depend on the source of the correction [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. For instance, people do not fear evaluation (a threat to positive face) by a computer and thus desire less impression management [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. This is mirrored in findings on conversational bots in mental health care and alcohol use counselling: The non-judgemental support provided by LLMs is an incentive that allows users to discuss stigmatized topics without fearing reputational damage [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. As artificial agents are less likely to threaten an individual\u0026rsquo;s face, resistance to corrections might decrease and ultimately lead to stronger intervention effects compared to human-labelled interventions.\u003c/p\u003e \u003cp\u003eLastly, artificial sources might outperform human-labelled interventions because people might be influenced by machine heuristics, that is, the mental shortcut \u0026ldquo;wherein we attribute machine characteristics [\u0026hellip;] when making judgments about the outcome of an interaction\u0026rdquo; [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Partisan Republicans, for example, rate news and fact-checks credited to AI sources as less biased compared to those attributed to human sources [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Similarly, corrective messages on COVID-19 attributed to AI rather than human experts trigger less motivated reasoning [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. And finally, AI source cues also reduced hostile media biases in a general US sample [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Therefore, conspiracy believers might categorize the artificial source as part of a neutral, non-competitive outgroup that induces less resistance to change and ultimately stronger intervention effects. This advantage of non-human sources for individuals\u0026rsquo; general openness to opposing views has been confirmed in experimental research [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Thus, following research on (1) weaker social responses, (2) reduced face-threats, and (3) machine heuristics, we predict that an \u0026ldquo;AI\u0026rdquo; source label condition will be more effective in changing conspiracy beliefs compared to a \u0026ldquo;Human\u0026rdquo; source label condition.\u003c/p\u003e \u003cp\u003eIn sum, we hypothesize that while LLMs may be able to reduce beliefs in health-related conspiracy theories, the perceived source could be crucial. Specifically, LLM-generated responses should be more effective if the source is perceived to be non-human. To test these assumptions, we invited participants to debate a COVID-19-related conspiracy with a chatbot labelled as either human or AI. Thereby we expanded research on chatbot-driven interventions in three ways: First, by recruiting a sample with negative attitudes towards the COVID-19 vaccine to increase the likelihood of personally relevant conspiracy beliefs. Then, by exploring effects of the different labels. And finally, by measuring exploratory mediators such as threat to freedom, counterarguing, threat to face, perceived neutrality and perceived competition that might explain the predicted effect.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eIntervention Effectiveness\u003c/h2\u003e \u003cp\u003eConsistent with our pre-registered hypothesis, belief change was significantly stronger for participants in the AI-label treatment compared to the control (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;7.88, 95% CI [5.35,10.42], \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, \u003cem\u003et\u003c/em\u003e(370)\u0026thinsp;=\u0026thinsp;6.11, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.63; see Supplementary Table S2 for full model). Descriptively, participants in the AI-label condition decreased their confidence in their belief by 7.06 percentage points on average (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;16.65), whereas belief confidence in the control condition slightly increased by 0.82 percentage points (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.77; see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The predicted difference remained significant when excluding participants who failed an attention check and when conducting various robustness checks (see Supplementary Data Analysis and Supplementary Table S3\u0026ndash;S6).\u003c/p\u003e \u003cp\u003eReplicating Costello et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], we further dichotomized post-treatment belief confidence at the scale midpoint. This threshold acted as the boundary between endorsement and rejection and thus allowed us to identify participants who transitioned from belief to scepticism. Logistic regression reveals that the AI-label treatment was significantly more effective in converting believers to sceptics (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.22, 95% CI [1.65, 6.11], \u003cem\u003ep\u003c/em\u003e = .001; see\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSupplementary Table S7). While 12% of the participants in the AI-label treatment shifted toward scepticism, this was the case for fewer than 1% in the control group. Overall, this translates to a number needed to treat (NNT) of 9, indicating that for every 9 people receiving the intervention, one person more turns sceptic of conspiracies (95% CI [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]) compared to the control.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSource Attribution\u003c/h3\u003e\n\u003cp\u003eInvestigating the role of source attribution we then proceeded to compare the human-label condition to the AI-label and the control (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In both cases we detected significant differences: Debates under the human-label results in a 13.76 percentage points larger confidence drop compared to the control (95% CI [-17.05, -10.47], \u003cem\u003et\u003c/em\u003e(551) = -8.21, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, \u003cem\u003ed\u003c/em\u003e = -0.87, see Supplementary Table S8). Contrary to our hypothesis, however, this drop was also 5.88 percentage points larger than in the AI-label condition (95% CI [-9.17, -2.58], \u003cem\u003et\u003c/em\u003e(551) = -3.50, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, \u003cem\u003ed\u003c/em\u003e = -0.30). Put differently, the human-label was nearly twice as effective as the AI-label. Again, these differences remained significant when excluding participants who failed the attention check and conducting various robustness checks (see Supplementary Data Analysis and Supplementary Table S9\u0026ndash;S12).\u003c/p\u003e \u003cp\u003eIn terms of converting believers, 18% of the participants in the human-label condition fell below the scale midpoint after the treatment. For every 6 individuals receiving the intervention, one more becomes sceptic of conspiracies (95% CI [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]) compared to the control. However, compared to the AI-label, this difference in belief conversions was not significant (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.46, 95% CI [ -0.12, 1.06], \u003cem\u003ep\u003c/em\u003e = .13; see Supplementary Table S13).\u003c/p\u003e\n\u003ch3\u003eExploratory Analysis of Process Variables\u003c/h3\u003e\n\u003cp\u003eWe now turn to how participants perceived the discussion with the LLM (see Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The first pathway we proposed was via weaker social responses like reduced threat to freedom and consequently less counterarguing. Contrary to this account, however, individuals perceived the AI-label to be more threatening to freedom (\u003cem\u003et\u003c/em\u003e(360)\u0026thinsp;=\u0026thinsp;2.51, \u003cem\u003ep\u003c/em\u003e = .01, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.26) compared to the human-label. Furthermore, there was no evidence that counterarguing differed between conditions; \u003cem\u003et\u003c/em\u003e(360)\u0026thinsp;=\u0026thinsp;1.87, \u003cem\u003ep\u003c/em\u003e = .06, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.2. The second pathway we explored was based on decreased face-threat. Yet, once more, the AI-label was seen as more threatening; \u003cem\u003et\u003c/em\u003e(365)\u0026thinsp;=\u0026thinsp;2.26, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.02, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.24. Finally, as a third pathway, we explored whether machine heuristics increase perceived neutrality and decrease perceived competition. Among AI-label participants, correlations supported this mechanism: machine heuristics were moderately associated with both perceived neutrality (\u003cem\u003er\u003c/em\u003e\u003csub\u003e\u003cem\u003es\u003c/em\u003e\u003c/sub\u003e(183) = .51, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001) and belief change (\u003cem\u003er\u003c/em\u003e\u003csub\u003e\u003cem\u003es\u003c/em\u003e\u003c/sub\u003e(183) = .33, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). However, this effect did not translate to the condition level.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eOn the contrary, the AI-label intervention was perceived as less neutral than the human-label; \u003cem\u003et\u003c/em\u003e(365) = -4.94, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, \u003cem\u003ed\u003c/em\u003e = -0.52. While the correlation between machine heuristics and perceived competition was also significant (\u003cem\u003er\u003c/em\u003e\u003csub\u003e\u003cem\u003es\u003c/em\u003e\u003c/sub\u003e(183) = .15, \u003cem\u003ep\u003c/em\u003e = .05), the direction was reversed. Additionally, there was no evidence that the AI-label was seen as less competitive than the human-label; \u003cem\u003et\u003c/em\u003e(359)\u0026thinsp;=\u0026thinsp;0.59, \u003cem\u003ep\u003c/em\u003e = .56, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.06.\u003c/p\u003e \u003cp\u003eGiven the fact that the human-label intervention was perceived as less threatening and more neutral, we proceeded by exploring whether any of these differences could explain the fact that the AI-label intervention was more persuasive than the human-label intervention. Specifically, we performed a parallel mediation analysis with 10,000 bootstrap samples comparing the intervention conditions on post-treatment confidence (controlling for baseline confidence) with threat to freedom, threat to face and neutrality as mediators (see Supplementary Table S16). In the human-label condition participants perceived their discussion partner as more neutral compared to the AI-label condition (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.68, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). This higher neutrality, in turn, was associated with a larger confidence reduction (\u003cem\u003eb\u003c/em\u003e = -5.65, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). The resulting indirect effect via neutrality was statistically significant (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;3.86, 95% CI [-5.91, -2.14]). Threat to freedom was lower in the human-label condition (\u003cem\u003eb\u003c/em\u003e = -0.47, \u003cem\u003ep\u003c/em\u003e = .008) and positively associated with post-treatment confidence (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.60, \u003cem\u003ep\u003c/em\u003e = .05). However, there was no evidence of an indirect effect (\u003cem\u003eb\u003c/em\u003e = -0.75, 95% CI [-1.87, 0.02]). Threat to face had no significant association with post-treatment confidence (\u003cem\u003eb\u003c/em\u003e = -0.77, \u003cem\u003ep\u003c/em\u003e = .37) and there was no evidence of an indirect effect (\u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.26, 95% CI [-0.28, 1.11]).\u003c/p\u003e \u003cp\u003eFinally, to explore potential differences between perceptions of computer-mediated vs human-computer communication, we also explored whether the perceived agency of the discussion partner influenced belief change for those in the AI-label condition. However, no significant association was found (\u003cem\u003er\u003c/em\u003e\u003csub\u003e\u003cem\u003es\u003c/em\u003e\u003c/sub\u003e(183) = .02, \u003cem\u003ep\u003c/em\u003e = .80).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDespite increasing interest in the use of conversational AI for persuasive purposes, the effectiveness of chatbots in the health communication domain remains largely unexplored. We addressed this gap by examining whether LLM-based conversational interventions can reduce health-related conspiracy beliefs, a particularly challenging and high-stakes context, and by testing source attribution as a key mechanism shaping their effectiveness.\u003c/p\u003e \u003cp\u003eOur findings demonstrate that conversations with chatbots are indeed effective at fighting health-related conspiracy theories. The intervention produced a robust reduction in conspiracy beliefs. Even though effect sizes were at the lower end of the range reported by Costello et al., they align well with their results on \u0026ldquo;particularly entrenched beliefs\u0026rdquo; [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Moreover, compared to conventional debunking effects in health communication (\u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.40, 95% CI [0.25, 0.55]; [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]), chatbot induced reduction in conspiracy beliefs can be considered substantial. This illustrates that LLM-driven interventions can be effective even in highly self-relevant domains such as health. However, contrary to our expectations, this reduction occurred not because of but despite participants\u0026rsquo; awareness of their artificial conversation partner. Post-debate, participants in the AI-label condition reported 7 percentage points less confidence in the target conspiracy. In contrast, those believing to debate another human indicated a confidence drop nearly twice as strong (13 percentage points). This difference appears to be mediated by the perceived neutrality of the conversation partner. Additionally and inconsistent with machine heuristics and previous work on the perceived bias in artificial versus human sources [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e], artificial sources were also seen as more threatening to both autonomy and face compared to those labelled as human.\u003c/p\u003e \u003cp\u003eThis study adds to the growing body of evidence for the persuasive power of LLMs. It illustrates how LLMs can be used to deliver effective attitude-change communication at scale. Yet, it also reveals that source attribution appears to matter for health communication: People are influenced more easily when they believe to be debating with humans rather than LLMs. This rules out source attribution as the mechanism driving the effectiveness of these interventions. Trying to reconcile these findings with Boissin et al., who observed no meaningful difference between human and AI-labels in a similar setup [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], there are two relevant factors to look at: The heightened level of self-relevance for the health domain and the generalizability of insights on source-based reasoning.\u003c/p\u003e \u003cp\u003eIn established dual-process theories of persuasion, heightened self-relevance encourages argument-based processing and therefore diminishes the role of source characteristics [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. Yet, considering such context clues might not always be a sign of peripheral processing. Instead, Lee\u0026rsquo;s integrative theory of human-machine communication argues that using source information in addition to message content actually signals a more effortful cognitive approach [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Put differently, people discussing things that matter to them use source cues not as a shortcut, but as an additional means of reaching a more accurate judgement. This elevates the importance of source cues and brings us to our second variable: their generalizability.\u003c/p\u003e \u003cp\u003eComputers are often perceived as unbiased, objective and less threatening [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. However, as our participants perceived the AI-label as less neutral and more threatening to freedom and face, our data suggests that these assumptions might be context specific. One reason for these findings might be the changing public narrative around AI. Disappointing experiences with LLMs in everyday use, the rise in negative news coverage following the release of ChatGPT [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e], concerns about authenticity and slop, and the heightened awareness of manipulative and fallible chatbots might have shifted the valence of how people perceive LLMs by the time of our data collection. In addition, it remains unclear to what degree machine heuristics (as originally conceptualized: [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]) generalize to LLM-driven chatbots in the first place. In fact, the interactive nature and the affordances of the technology differs strongly from the work that established machine heuristics on news recommendation systems [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. Thus, source-based evaluation for LLMs might not follow the same mechanisms.\u003c/p\u003e \u003cp\u003eFuture research should therefore attempt to test whether general machine heuristics apply to LLMs at all, or whether the technology has given rise to new LLM heuristics. To fully understand potential shifts in machine heuristics, longitudinal survey studies exploring changes in people\u0026rsquo;s perceptions of LLMs over time are necessary. As our study failed to identify a mechanism for the performance of LLM-driven interventions, additional studies should also probe into alternative explanations. This includes, inter alia, information density, the interactive presentation of information, or both politeness and confidence of the model.\u003c/p\u003e \u003cp\u003eIt is of course difficult to determine the direction of the effect, as we measured machine heuristics after the interaction: did weaker machine heuristics lead to smaller belief changes, or did participants report more negative responses because the LLM argued against their deeply held beliefs? Recent findings on increased negative attitudes towards AI after being outsmarted by an artificial agent might point to the latter [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. Additionally, as we utilized identity threat as a proxy for positive face threat, we might have missed nuances of the concept relevant for the intervention at hand. Given the mediating effect of neutrality, the question also arises whether the difference between AI-label and human-label was driven by the lower perceived neutrality of the AI agent, or the heightened neutrality of our three deception framings (i.e., a communication student, a journalist or a social worker in training). A third limitation is the high deception-based screen-out quota for our human-label condition. It is possible that by excluding those who saw through the deception, we incidentally included only those participants most gullible to synthetic influence in the first place. Finally, in our attempt to increase the credibility of the deception while controlling for differences in message content we instructed the LLM to paraphrase its outputs in a human-like way across both intervention conditions. As a result, generated messages were less polite and conflict averse. Therefore, the differences found could be attributed to the violation of expectations in the AI-label condition. However, given that Boissin et al. found no interaction between source label and prompt type when utilizing a similar approach [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], this seems unlikely.\u003c/p\u003e \u003cp\u003eOur findings indicate that LLMs are indeed effective persuaders in highly self-relevant domains. While this is good news for those trying to shape effective interventions in the health contexts, it also unveils equivalent risks of misuse and manipulation. Unlike other epistemic threats like disinformation and propaganda, chatbots are not in short demand: Day by day, millions of users worldwide choose to interact voluntarily with LLMs accountable only to the corporations owning them. If even one conversation with such a model can fundamentally shift deeply held beliefs, this raises important questions about the long-term consequences of repeated exposure. Potentially, these interactions could erode societal core values like the support for democracy or trust in science and modern medicine. Effects that, by propagating along individual and social networks, can cascade beyond single beliefs and single believers [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Yet, our results also show, that people are more open to human rather than artificial guidance. And while this creates further incentives for malign actors to hide the true origin of synthetic messages, it also illustrates that people still value \u0026ndash; what they believe to be \u0026ndash; genuine human interaction.\u003c/p\u003e"},{"header":"Methods \u0026 Materials","content":"\u003cp\u003eThis pre-registered online experiment (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://osf.io/yf6qs/overview\u003c/span\u003e\u003cspan address=\"https://osf.io/yf6qs/overview\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) follows a 3 (Treatment: Control vs \u0026ldquo;Human\u0026rdquo; label vs \u0026ldquo;AI\u0026rdquo; label; between) x 2 (Time: Pre-Treatment vs Post-Treatment; within) mixed design with belief-change (the delta between pre- and post-treatment measures) as the primary dependent variable. This study received approval by Radboud University\u0026rsquo;s Ethics Assessment Committee Humanities (2025\u0026ndash;5907) and was conducted in accordance with its guidelines and regulations.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eParticipants \u0026amp; Recruitment\u003c/h2\u003e \u003cp\u003eUS and UK based participants with high approval rates (\u0026gt;\u0026thinsp;95%) were recruited via Prolific and compensated with \u0026pound; 6 per hour. Due to the deceptive nature of the study (i.e., some participants were told they were interacting with another human when they were not), we relied on Prolific pre-screeners to target users that a) agreed on participating in deceptive studies and b) had some prior experience with chatbots (indicating at least a minimal level of acceptance for interacting with generative AI). Furthermore, we only targeted individuals with negative attitudes towards the COVID-19 vaccine. Our target sample size of 546 was based on a priori power analysis utilizing a Monte-Carlo simulation for a linear model with the belief delta as the dependent variable and dummy variables for both the control (\u003cem\u003ed\u003c/em\u003e = -0.53, based on the more conservative effect size identified by Costello et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]) and the human condition (\u003cem\u003ed\u003c/em\u003e = -0.26) as predictors with 80% power at alpha\u0026thinsp;=\u0026thinsp;0.05.\u003c/p\u003e \u003cp\u003eDuring the experiment, participants were automatically excluded based on pre-registered criteria (e.g., failed to share a valid conspiracy belief; see Procedure). Furthermore, we sequentially reviewed whether they saw through the deception. In both cases, additional individuals were recruited to meet the required target sample size. Out of 876 participants initially recruited, 754 completed the experiment. Following our screen-out process (see Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), 554 individuals remained in the final sample (see Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). Of those, 361 (65%) were women, 192 (35%) men and 1 (\u0026lt;\u0026thinsp;1%) non-binary. All of them were between 18 and 78 years old (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;42.96, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;12.51). As their highest level of education, most participants indicated finishing high school (50%) followed by Bachelor\u0026rsquo;s (38%) and Master\u0026rsquo;s (11%) degrees.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eProcedure\u003c/h3\u003e\n\u003cp\u003eFollowing the original design by Costello et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], the experiment was conducted online via Qualtrics. Utilizing the randomization element, participants were assigned to one of three conditions: A control condition, the \u0026ldquo;AI\u0026rdquo; label treatment, and the \u0026ldquo;human\u0026rdquo; label treatment. After obtaining informed consent, participants across all conditions were instructed to describe a conspiracy theory endorsed by them. In a follow-up question, they further expanded on their previous response by including specific evidence and sources influencing their perspective. Utilizing the OpenRouter API, these responses were piped into an LLM (i.e., Claude Sonnet 4.0; prompt adapted from [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]; see Supplementary Table S15 for all prompts) instructed to assess whether the described belief qualifies as a conspiracy. If this threshold was not met, participants were asked for another conspiracy belief once before being screened out. Evaluating model performance post-hoc based on a balanced sample (i.e. 10% of the dataset, half classified as conspiracy and half not classified as such by the LLM), we observed substantial alignment with human-labels (\u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.73). For valid conspiracies, the provided description was then summarized by the model into a single sentence. In the \u0026ldquo;human\u0026rdquo; label condition, this process was artificially delayed facilitating the deception. Then, subjects were asked to indicate their belief in the summarized conspiracy on a 101-point rating scale as well as whether the summary accurately captured their description. At this point, individuals rating their belief beneath the midpoint or the summary as inaccurate were automatically excluded from the study.\u003c/p\u003e \u003cp\u003eFor the \u0026ldquo;AI\u0026rdquo; label treatment, subjects were instructed to discuss their conspiracy beliefs with an AI agent. For the \u0026ldquo;human\u0026rdquo; label treatment, they were informed about conversing with either a communication student, a journalist or a social worker in training. To facilitate the deception in the \u0026ldquo;human\u0026rdquo; label condition, we included elaborate backstories (e.g., training journalists on how to engage around complicated topics), artificial delays and occasional spelling mistakes. Additionally, participation in the experiment was only available during \u0026ldquo;office hours\u0026rdquo;. In both treatment conditions, participants had three rounds of debate with an instance of Claude Sonnet 4.0. The LLM was prompted to reduce the user\u0026rsquo;s belief in the conspiracy based on their description of it as well as their previous responses and the rating on the belief scale. In both conditions, a second instance of the same model was then instructed to regenerate the answer in a human-like tone. Finally, the \u0026ldquo;humanified\u0026rdquo; response was evaluated by two more instances to check, whether the model refused to follow any instructions or admitted to being an LLM (for ethical reasons we explicitly prompted the model to admit to being an LLM when asked). In those cases, the initial response was displayed to the participant (see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). For the control condition, participants were asked to discuss an unrelated topic \u0026ndash; i.e., their previous experiences with the fire department \u0026ndash; with the LLM. Post-treatment, participants were once more asked to rate their confidence in the initial machine-generated summary and to answer items on various process variables. Finally, all participants were debriefed and offered the opportunity to remove their data from the analysis.\u003c/p\u003e\n\u003ch3\u003eMeasures\u003c/h3\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003ePrimary Outcome Variable\u003c/h2\u003e \u003cp\u003eBefore and after interaction with the chatbot, participants were asked to rate their confidence in the target conspiracy theory on a 101-point scale. The item \u0026ldquo;On a scale of 0% to 100%, please indicate your level of confidence that this statement is true.\u0026rdquo; was adapted from [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eProcess Variables\u003c/h2\u003e \u003cp\u003eTo identify possible mechanisms for the effects we collected data on various process variables. Measures and items were presented in randomized order. All items were rated on a 7-point Likert scale (1\u0026thinsp;=\u0026thinsp;strongly disagree to 7\u0026thinsp;=\u0026thinsp;strongly agree). For all scales mean scores were calculated.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCounterarguing.\u003c/b\u003e The measure for counterarguing consisted of 4 items adapted from [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.71). Examples are \u0026ldquo;While reading the responses, I sometimes found myself thinking of ways I disagreed with what was being presented.\u0026rdquo; and \u0026ldquo;I found myself looking for flaws in the way information was presented in the response.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003eThreat to Freedom.\u003c/b\u003e Threat to freedom was measured utilizing 5 items adapted from [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.94). Specific wording for each item depends on the condition. Examples are \u0026ldquo;The AI / other person tried to make a decision for me.\u0026rdquo; and \u0026ldquo;The AI / other person tried to force their opinion on me.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003eThreat to Face.\u003c/b\u003e To assess perceived threat to face we adapted 4 items utilized by [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.89). Although originally intended to measure social identity threat, these measures capture threats to self-image and as such align well with threats to positive face as conceptualized by [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Examples are \u0026ldquo;The responses I received undermined my sense of self-worth.\u0026rdquo; and \u0026ldquo;The responses I received made me feel less unique as a person.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003eSource Neutrality.\u003c/b\u003e For source neutrality 4 items were adapted from [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.76). Specific wording for each item depends on the condition. Examples are \u0026ldquo;The AI\u0026rsquo;s / other person\u0026rsquo;s position on this issue is rooted in an objective analysis of the issue.\u0026rdquo; and \u0026ldquo;The AI\u0026rsquo;s / other person\u0026rsquo;s reasoning about this issue is influenced by bias.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003ePerceived Competition.\u003c/b\u003e We relied on 3 items adapted from [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e] to measure perceived competition (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.91). Specific wording for each item depends on the condition. Examples are \u0026ldquo;I view the AI / other person as a competitor.\u0026rdquo; and \u0026ldquo;Competition towards the AI / other person is important to me.\u0026rdquo;\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eAdditional Measures\u003c/h2\u003e \u003cp\u003e\u003cb\u003ePerceived Agency.\u003c/b\u003e To measure the perceived agency of the chatbot across both the AI-label and the control condition, we utilized 5 items from [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.86). All items were presented randomized and rated on a 7-point Likert scale (1\u0026thinsp;=\u0026thinsp;strongly disagree to 7\u0026thinsp;=\u0026thinsp;strongly agree). Mean scores were calculated. Examples are \u0026ldquo;The AI intentionally completed its responses.\u0026rdquo; and \u0026ldquo;The AI could have chosen not to respond in this way.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003eMachine Heuristics.\u003c/b\u003e For machine heuristics across both the AI-label and the control condition we adapted 7 items from [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e] (Cronbach\u0026rsquo;s α\u0026thinsp;=\u0026thinsp;.97). All items were presented randomized and rated on a 7-point Likert scale (1\u0026thinsp;=\u0026thinsp;strongly disagree to 7\u0026thinsp;=\u0026thinsp;strongly agree). Mean scores were calculated. Examples are \u0026ldquo;I trust AI more than humans when discussing my beliefs.\u0026rdquo; and \u0026ldquo;AI is better than humans at discussing my beliefs.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003cb\u003eDemographics.\u003c/b\u003e Finally, we also collected demographic data on age, gender, and education.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eDeviation from Pre-Registration\u003c/h2\u003e \u003cp\u003eFor enhanced clarity we reversed the presentation order of hypotheses from our pre-registration. Additionally, to assess the effectiveness of the baseline intervention we focused on data from the control and the AI-label treatment only. To increase interpretability for the analysis of source effects we changed the reference group from the AI-label to the human-label intervention. This allowed us to test two relevant comparisons (AI vs. human and human vs. control) in one model rather than using two models and was thus considered superior. However, results of the original model are in line with our findings (see Supplementary Table S14).\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eAcknowledgements will be described on publication.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003e \u003cb\u003eAdditional Information\u003c/b\u003e \u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eCompeting interests\u003c/strong\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThere is no funding associated with the work featured in this article.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eSupervision: H.S., P.S.;Conceptualization: P.B., Y.v.d.S., H.S., P.S.;Study Design: P.B., Y.v.d.S., P.S.;Data Collection: P.B.;Data Analysis: P.B., Y.v.d.S., P.S.;Writing \u0026ndash; original draft: P.B.;Writing \u0026ndash; review \u0026amp; editing: P.B., Y.v.d.S., H.S., P.S.;\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors thank Mark Dingemanse for his expertise and guidance.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data supporting the findings of this study was deposited via OSF and is available at https://osf.io/su9kp/overview?view_only=7a40bc5d15cd4d308b10b9c0cc97d600.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eRegazzi, L., Lontano, A., Cadeddu, C., Di Padova, P. \u0026amp; Rosano, A. Conspiracy beliefs, COVID-19 vaccine uptake and adherence to public health interventions during the pandemic in Europe. \u003cem\u003eEur. J. Public. Health\u003c/em\u003e. \u003cb\u003e33\u003c/b\u003e, 717\u0026ndash;724 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRomer, D. \u0026amp; Jamieson, K. H. Conspiracy theories as barriers to controlling the spread of COVID-19 in the U.S. \u003cem\u003eSoc. Sci. Med.\u003c/em\u003e \u003cb\u003e263\u003c/b\u003e, 113356 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZilinsky, J. \u0026amp; Theocharis, Y. Conspiracism and government distrust predict COVID-19 vaccine refusal. \u003cem\u003eHumanit. Soc. Sci. Commun.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 1002 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDouglas, K. M. et al. Understanding Conspiracy Theories. \u003cem\u003ePolit Psychol.\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e, 3\u0026ndash;35 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDouglas, K. M., Sutton, R. M. \u0026amp; Cichocka, A. The Psychology of Conspiracy Theories. \u003cem\u003eCurr. Dir. Psychol. Sci.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e, 538\u0026ndash;542 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Mulukom, V. et al. Antecedents and consequences of COVID-19 conspiracy beliefs: A systematic review. \u003cem\u003eSoc. Sci. Med.\u003c/em\u003e \u003cb\u003e301\u003c/b\u003e, 114912 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO\u0026rsquo;Mahony, C., Brassil, M., Murphy, G. \u0026amp; Linehan, C. The efficacy of interventions in reducing belief in conspiracy theories: A systematic review. \u003cem\u003ePLOS ONE\u003c/em\u003e. \u003cb\u003e18\u003c/b\u003e, e0280902 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTaubert, F., Meyer-Hoeven, G., Schmid, P., Gerdes, P. \u0026amp; Betsch, C. Conspiracy narratives and vaccine hesitancy: a scoping review of prevalence, impact, and interventions. \u003cem\u003eBMC Public. Health\u003c/em\u003e. \u003cb\u003e24\u003c/b\u003e, 3325 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStasielowicz, L. The Effectiveness of Interventions Addressing Conspiracy Beliefs: A Meta-Analysis. \u003cem\u003eEur. J. Soc. Psychol.\u003c/em\u003e \u003cb\u003e56\u003c/b\u003e, 275\u0026ndash;291 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanh, L. \u0026amp; Strobel, G. Generative artificial intelligence. \u003cem\u003eElectron. Mark.\u003c/em\u003e \u003cb\u003e33\u003c/b\u003e, 63 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCostello, T. H., Pennycook, G. \u0026amp; Rand, D. G. Durably reducing conspiracy beliefs through dialogues with AI. \u003cem\u003eScience\u003c/em\u003e \u003cb\u003e385\u003c/b\u003e, eadq1814 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan, M. S. \u0026amp; Albarrac\u0026iacute;n, D. A meta-analysis of correction effects in science-relevant misinformation. \u003cem\u003eNat. Hum. Behav.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 1514\u0026ndash;1525 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDarke, P. R. \u0026amp; Chaiken, S. The pursuit of self-interest: Self-interest bias in attitude judgment and persuasion. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cb\u003e89\u003c/b\u003e, 864\u0026ndash;883 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFalk, E., Scholz, C. \u0026amp; Persuasion Influence, and Value: Perspectives from Communication and Social Neuroscience. \u003cem\u003eAnnu. Rev. Psychol.\u003c/em\u003e \u003cb\u003e69\u003c/b\u003e, 329\u0026ndash;356 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson, B. T. \u0026amp; Eagly, A. H. Effects of involvement on persuasion: A meta-analysis. \u003cem\u003ePsychol. Bull.\u003c/em\u003e \u003cb\u003e106\u003c/b\u003e, 290\u0026ndash;314 (1989).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSteindl, C., Jonas, E., Sittenthaler, S., Traut-Mattausch, E. \u0026amp; Greenberg, J. Understanding Psychological Reactance: New Developments and Findings. \u003cem\u003eZ. F\u0026uuml;r Psychol.\u003c/em\u003e \u003cb\u003e223\u003c/b\u003e, 205\u0026ndash;214 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhazali, A. S., Ham, J., Barakova, E. \u0026amp; Markopoulos, P. The influence of social cues in persuasive social robots on psychological reactance and compliance. \u003cem\u003eComput. Hum. Behav.\u003c/em\u003e \u003cb\u003e87\u003c/b\u003e, 58\u0026ndash;65 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoubroeks, M., Ham, J. \u0026amp; Midden, C. When Artificial Social Agents Try to Persuade People: The Role of Social Agency on the Occurrence of Psychological Reactance. \u003cem\u003eInt. J. Soc. Robot\u003c/em\u003e. \u003cb\u003e3\u003c/b\u003e, 155\u0026ndash;165 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoubroeks, M., Midden, C. \u0026amp; Ham, J. Does it make a difference who tells you what to do? exploring the effect of social agency on psychological reactance. in \u003cem\u003eProceedings of the 4th International Conference on Persuasive Technology\u003c/em\u003e 1\u0026ndash;6 (ACM, Claremont California USA, (2009). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/1541948.1541970\u003c/span\u003e\u003cspan address=\"10.1145/1541948.1541970\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKr\u0026auml;mer, N. C. Social Communicative Effects of a Virtual Program Guide. in Intelligent Virtual Agents (ed (eds Panayiotopoulos, T. et al.) vol. 3661 442\u0026ndash;453 (Springer Berlin Heidelberg, Berlin, Heidelberg, (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorkes, J., Kernal, H. K. \u0026amp; Nass, C. Effects of Humor in Task-Oriented Human-Computer Interaction and Computer-Mediated Communication: A Direct Test of SRCT Theory. \u003cem\u003eHuman\u0026ndash;Computer Interact.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 395\u0026ndash;435 (1999).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNass, C. \u0026amp; Moon, Y. Machines and Mindlessness: Social Responses to Computers. \u003cem\u003eJ. Soc. Issues\u003c/em\u003e. \u003cb\u003e56\u003c/b\u003e, 81\u0026ndash;103 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNass, C., Steuer, J. \u0026amp; Voices Boxes, and Sources of Messages.: Computers and Social Actors. \u003cem\u003eHum. Commun. Res.\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e, 504\u0026ndash;527 (1993).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSlater, M. D. \u0026amp; Rouner, D. Entertainment?Education and Elaboration Likelihood: Understanding the Processing of Narrative Persuasion. \u003cem\u003eCommun. Theory\u003c/em\u003e. \u003cb\u003e12\u003c/b\u003e, 173\u0026ndash;191 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown, P. \u0026amp; Levinson, S. C. \u003cem\u003ePoliteness: Some Universals in Language Usage\u003c/em\u003e (Cambridge University Press, 1987). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1017/CBO9780511813085\u003c/span\u003e\u003cspan address=\"10.1017/CBO9780511813085\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmid, P. Debunking health misinformation with empathy. \u003cem\u003eCurr. Opin. Psychol.\u003c/em\u003e \u003cb\u003e67\u003c/b\u003e, 102213 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJenkins, M. \u0026amp; Dragojevic, M. Explaining the Process of Resistance to Persuasion: A Politeness Theory-Based Approach. \u003cem\u003eCommun. Res.\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e, 559\u0026ndash;590 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWhite, J. B., Tynan, R., Galinsky, A. D. \u0026amp; Thompson, L. Face threat sensitivity in negotiation: Roadblock to agreement and joint gain. \u003cem\u003eOrgan. Behav. Hum. Decis. Process.\u003c/em\u003e \u003cb\u003e94\u003c/b\u003e, 102\u0026ndash;124 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBabel, F., Hock, P., Kraus, J. \u0026amp; Baumann, M. Human-Robot Conflict Resolution at an Elevator - The Effect of Robot Type, Request Politeness and Modality. in \u003cem\u003e17th ACM/IEEE International Conference on Human-Robot Interaction (HRI)\u003c/em\u003e 693\u0026ndash;697 (IEEE, Sapporo, Japan, 2022). 693\u0026ndash;697 (IEEE, Sapporo, Japan, 2022). (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/HRI53351.2022.9889387\u003c/span\u003e\u003cspan address=\"10.1109/HRI53351.2022.9889387\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGratch, J., Lucas, G. M., King, A. A. \u0026amp; Morency, L. P. It\u0026rsquo;s only a computer: the impact of human-agent interaction in clinical interviews. in \u003cem\u003eProceedings of the International Conference on Autonomous Agents and Multi-Agent Systems\u003c/em\u003e 85\u0026ndash;92 (International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2014). 85\u0026ndash;92 (International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2014). (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeisband, S. \u0026amp; Kiesler, S. Self disclosure on computer forms: meta-analysis and implications. in \u003cem\u003eProceedings of the SIGCHI conference on Human factors in computing systems common ground - CHI \u0026rsquo;96\u003c/em\u003e 3\u0026ndash;10ACM Press, Vancouver, British Columbia, Canada, (1996). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/238386.238387\u003c/span\u003e\u003cspan address=\"10.1145/238386.238387\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa, Z., Mei, Y. \u0026amp; Su, Z. Understanding the benefits and challenges of using large language model-based conversational agents for mental well-being support. in \u003cem\u003eAMIA Annual Symposium Proceedings\u003c/em\u003e vol. 1105 (American Medical Informatics Association, 2023). (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSteenstra, I., Nouraei, F., Arjmand, M. \u0026amp; Bickmore, T. W. Virtual Agents for Alcohol Use Counseling: Exploring LLM-Powered Motivational Interviewing. Preprint at (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3652988.3673932\u003c/span\u003e\u003cspan address=\"10.1145/3652988.3673932\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSundar, S. S., Kim, J. M. \u0026amp; Heuristic When We Trust Computers More than Humans with Our Personal Information. in \u003cem\u003eProceedings of the CHI Conference on Human Factors in Computing Systems\u003c/em\u003e 1\u0026ndash;9 (ACM, Glasgow Scotland Uk, 2019). 1\u0026ndash;9 (ACM, Glasgow Scotland Uk, 2019). (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3290605.3300768\u003c/span\u003e\u003cspan address=\"10.1145/3290605.3300768\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChae, J. H. \u0026amp; Tewksbury, D. Perceiving AI intervention does not compromise the persuasive effect of fact-checking. \u003cem\u003eNew. Media Soc.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 191\u0026ndash;211 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWaddell, T. F. Can an Algorithm Reduce the Perceived Bias of News? Testing the Effect of Machine Attribution on News Readers\u0026rsquo; Evaluations of Bias, Anthropomorphism, and Credibility. \u003cem\u003eJ. Mass. Commun. Q.\u003c/em\u003e \u003cb\u003e96\u003c/b\u003e, 82\u0026ndash;100 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoon, W. K., Chung, M. \u0026amp; Jones-Jang, S. M. How Can We Fight Partisan Biases in the COVID-19 Pandemic? AI Source Labels on Fact-checking Messages Reduce Motivated Reasoning. \u003cem\u003eMass. Commun. Soc.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e, 646\u0026ndash;670 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCraig, M. J. A. \u0026amp; Choi, M. The role of affective and cognitive involvement in the mitigating effects of AI source cues on hostile media bias. \u003cem\u003eTelemat Inf.\u003c/em\u003e \u003cb\u003e88\u003c/b\u003e, 102097 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu, L., Tormala, Z. L. \u0026amp; Duhachek, A. How AI sources can increase openness to opposing views. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 17170 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalter, N., Brooks, J. J., Saucier, C. J. \u0026amp; Suresh, S. Evaluating the Impact of Attempts to Correct Health Misinformation on Social Media: A Meta-Analysis. \u003cem\u003eHealth Commun.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e, 1776\u0026ndash;1784 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoissin, E., Costello, T. H., Spinoza-Mart\u0026iacute;n, D., Rand, D. G. \u0026amp; Pennycook, G. Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human. \u003cem\u003ePNAS Nexus\u003c/em\u003e. \u003cb\u003e4\u003c/b\u003e, pgaf325 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChaiken, S. Heuristic versus systematic information processing and the use of source versus message cues in persuasion. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cb\u003e39\u003c/b\u003e, 752\u0026ndash;766 (1980).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePetty, R. E. \u0026amp; Cacioppo, J. T. The Elaboration Likelihood Model of Persuasion. in \u003cem\u003eAdvances Experimental Social Psychology\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e 123\u0026ndash;205 (Elsevier, 1986).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, E. J. Minding the source: toward an integrative theory of human\u0026ndash;machine communication. \u003cem\u003eHum. Commun. Res.\u003c/em\u003e \u003cb\u003e50\u003c/b\u003e, 184\u0026ndash;193 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJain, A. \u0026amp; Ranganathan, S. A. Longitudinal Analysis of Artificial Intelligence Coverage in Technology-Focused News Media Using Latent Dirichlet Allocation and Sentiment Analysis. \u003cem\u003eJ. Media\u003c/em\u003e. \u003cb\u003e6\u003c/b\u003e, 176 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSundar, S. S. \u0026amp; Nass, C. Conceptualizing sources in online news. \u003cem\u003eJ. Commun.\u003c/em\u003e \u003cb\u003e51\u003c/b\u003e, 52\u0026ndash;72 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnderl, C. et al. Conversational presentation mode increases credibility judgements during information search with ChatGPT. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 17127 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchaap, G., Van De Sande, Y. \u0026amp; Schraffenberger, H. Outperformed by AI: Interacting with Superhuman AI Changes the Way We Perceive Ourselves. in \u003cem\u003eExtended Abstracts of the CHI Conference on Human Factors in Computing Systems\u003c/em\u003e 1\u0026ndash;7 (ACM, Honolulu HI USA, (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3613905.3650961\u003c/span\u003e\u003cspan address=\"10.1145/3613905.3650961\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuingrich, R. E., Mehta, D. \u0026amp; Bhatt, U. Belief Offloading in Human-AI Interaction. Preprint at (2026). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2602.08754\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2602.08754\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa, Y. \u0026amp; Hmielowski, J. D. Are You Threatening Me? Identity Threat, Resistance to Persuasion, and Boomerang Effects in Environmental Communication. \u003cem\u003eEnviron. Commun.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e, 225\u0026ndash;242 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNabi, R. L., Moyer-Gus\u0026eacute;, E. \u0026amp; Byrne, S. All Joking Aside: A Serious Investigation into the Persuasive Effect of Funny Social Issue Messages. \u003cem\u003eCommun. Monogr.\u003c/em\u003e \u003cb\u003e74\u003c/b\u003e, 29\u0026ndash;54 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDillard, J. P. \u0026amp; Shen, L. On the Nature of Reactance and its Role in Persuasive Health Communication. \u003cem\u003eCommun. Monogr.\u003c/em\u003e \u003cb\u003e72\u003c/b\u003e, 144\u0026ndash;168 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKennedy, K. A. \u0026amp; Pronin, E. When Disagreement Gets Ugly: Perceptions of Bias and the Escalation of Conflict. \u003cem\u003ePers. Soc. Psychol. Bull.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e, 833\u0026ndash;848 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZiegler, R. \u0026amp; Diehl, M. Mood and Multiple Source Characteristics: Mood Congruency of Source Consensus Status and Source Trustworthiness as Determinants of Message Scrutiny. \u003cem\u003ePers. Soc. Psychol. Bull.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e, 1016\u0026ndash;1030 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYip, J. A., Schweitzer, M. E. \u0026amp; Nurmohamed, S. Trash-talking: Competitive incivility motivates rivalry, performance, and unethical behavior. \u003cem\u003eOrgan. Behav. Hum. Decis. Process.\u003c/em\u003e \u003cb\u003e144\u003c/b\u003e, 125\u0026ndash;144 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKorman, J., Harrison, A., McCurry, M. \u0026amp; Trafton, G. Beyond Programming: Can Robots\u0026rsquo; Norm-Violating Actions Elicit Mental State Attributions? in \u003cem\u003e14th ACM/IEEE International Conference on Human-Robot Interaction (HRI)\u003c/em\u003e 530\u0026ndash;531 (2019). 530\u0026ndash;531 (2019). (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/HRI.2019.8673293\u003c/span\u003e\u003cspan address=\"10.1109/HRI.2019.8673293\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, H. \u0026amp; Sundar, S. S. Machine heuristic: concept explication and development of a measurement scale. \u003cem\u003eJ. Comput. -Mediat Commun.\u003c/em\u003e \u003cb\u003e29\u003c/b\u003e, zmae019 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Health misinformation, persuasion, infodemic, belief-revision, debunking, generative AI","lastPublishedDoi":"10.21203/rs.3.rs-9428259/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9428259/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBeliefs in conspiracy theories and their resistance to correction pose critical challenges to public health. A potential countermeasure to this is the use of Large Language Model (LLM) driven debates about conspiracy theories. Yet while this approach yields promising results for general conspiracies, its generalizability to the health domain as well as the underlying mechanism for its effectiveness remain unclear. This study investigates whether LLM-driven debates can reduce health-related conspiracy theories in the context of COVID-19. Furthermore, it examines attribution to artificial rather than human sources as a potential explanation. In an online experiment, 554 participants were randomly assigned to either a control condition or to debate an individual COVID-19 conspiracy theory with an AI-labelled LLM or a human-labelled LLM. Compared to the control, participants aware of their artificial conversation partner reported 7.88 percentage points less confidence in the conspiracy after the intervention (\u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.63, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Contrary to our expectation, however, this effect was stronger for those assuming a human conversation partner: They indicated a 13.76 percentage points larger confidence drop compared to the control (\u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.87, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). This difference in effectiveness appears to be mediated by the perceived neutrality of the source.\u003c/p\u003e","manuscriptTitle":"Chatbots Reduce Health-Related Conspiracy Beliefs Not Because of but Despite Being Perceived as AI","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-07 20:15:30","doi":"10.21203/rs.3.rs-9428259/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-13T06:59:26+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T07:30:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T04:46:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"194134858196921591501398812975558945716","date":"2026-04-25T15:46:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"111335175868707412566697908055430104754","date":"2026-04-24T07:12:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"190693786026108096668724433669602064458","date":"2026-04-23T22:24:41+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-23T13:25:44+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-23T06:52:00+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-19T09:53:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-04-19T09:49:15+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"24bc0762-73ad-4657-ab2f-4b9847e9957a","owner":[],"postedDate":"May 7th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-13T06:59:26+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T07:30:45+00:00","index":21,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T04:46:50+00:00","index":20,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":67437940,"name":"Biological sciences/Psychology"},{"id":67437941,"name":"Social science/Psychology"},{"id":67437942,"name":"Social science/Science technology and society"}],"tags":[],"updatedAt":"2026-05-13T07:12:31+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-07 20:15:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9428259","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9428259","identity":"rs-9428259","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}