RESEARCH DOSSIER
Human–Artificial Political Communication and Persuasion
Abstract. Human–artificial political communication concerns political discussion between people and artificial conversational systems and its consequences for beliefs, evaluations, knowledge, and action. This fundamental research monograph presents a critical narrative synthesis of social-psychological persuasion theory, political communication research, randomized experiments with generated messages, human–AI dialogue studies, and emerging evidence concerning political behavior and information seeking. The analysis reconstructs elaboration, motivated reasoning, source credibility, political identity, conversational feedback, and information acquisition as distinct explanatory pathways. Research published during 2024–2026 is reviewed for causal identification, sampling, effect-size interpretation, factual reliability, model provenance, and outcome validity. Evidence demonstrates that artificial communicators can influence political attitudes and actions in controlled conditions, while microtargeting does not consistently improve upon generic messaging. Recent experiments also establish that changes in expressed attitudes and observable political acts can diverge and that perceived partisan bias moderates receptivity. A central methodological conclusion is that persuasive effectiveness and factual accuracy require separate evaluation. The synthesis places this field within the scientific investigation of human–artificial relations in the Artificial Era, including political autonomy, information provenance, institutional governance, and interdisciplinary implications. The method is an independently authored critical narrative research synthesis, without original participant recruitment, a preregistered systematic-review protocol, or pooled meta-analysis.
Keywords: human–artificial political communication; political persuasion; conversational artificial intelligence; political psychology; large language models; political attitudes; voter preferences; political knowledge; political behavior; source credibility; political bias; motivated reasoning; persuasion experiments; factual accuracy; democratic autonomy; Artificial Era
The Scientific Object of Human–Artificial Political Communication
Human–artificial political communication is an observable exchange in which a person encounters, questions, evaluates, or responds to political information generated by an artificial conversational system. The distinctive research object is the interaction between a politically situated human judgment and a computationally generated sequence of responses. Messages can concern policy claims, electoral choices, public institutions, political identity, evidence, or collective action. A system’s architecture, training, prompting, retrieval, and interface shape the information delivered; the user’s prior commitments, motivations, and interpretations shape the resulting conversation. Persuasion denotes a measured change in political evaluation or behavior attributable to an exposure or interaction under an appropriate causal comparison Argyle et al. (2025) Hackenburg et al. (2025).
The relevant unit of analysis includes the user, the model, the conversational history, the political issue, and the institutional setting. A static generated advertisement, an interactive dialogue, an automated social-media account, and an AI-assisted search session are distinct treatments with different opportunity structures. Political attitude shifts, improved factual understanding, vote intentions, petition signing, and donations also define distinct outcomes. The emerging field studies how these pathways interact while avoiding a single generalized coefficient for 'AI persuasion' that silently combines incomparable treatments Bai et al. (2025) Hackenburg et al. (2026) Luettgau et al. (2026).
This monograph presents an original critical narrative synthesis of experimentally investigated human–artificial political dialogue, associated theories of social influence, and methodological problems of causal inference, misinformation, and democratic legitimacy. It does not report original participant data or a preregistered systematic search. Its central research question asks which conversational characteristics and user interpretations predict measurable political change, what evidence distinguishes persuasion from learning or compliant responding, and how trustworthy communication can be evaluated independently of influence. The inquiry is located in Political Psychology and connects with Psychology of Artificial, Psychology of Human–Artificial Relations, and Media and Technology Psychology Petty & Cacioppo (1986) DiGiuseppe & Robison (2026).
Intellectual Genealogy: Rhetoric, Social Psychology, and Political Communication
Classical rhetorical thought treated persuasive speech as a relationship among argument, speaker, audience, and occasion. Modern empirical psychology converted parts of that tradition into testable questions concerning attention, credibility, elaboration, learning, and attitudinal response. Later political-communication research added institutional sources, party identity, mass media, selection of exposure, and collective interpretation. Artificial conversation inherits these problems while altering the speed, adaptability, and apparent social presence of the source. Historical continuity therefore supplies hypotheses rather than an assumption that persuasion by an artificial source functions exactly like persuasion by a human speaker Petty & Cacioppo (1986) Kunda (1990).
The elaboration likelihood model formalized how opportunity and motivation for scrutiny can condition responses to substantive arguments and peripheral cues. Work on motivated reasoning demonstrated that individuals can evaluate information differently depending on their prior commitments and desired conclusions. Research on political disagreement under networked exposure later showed that receiving opposing viewpoints need not produce convergence. These intellectual traditions locate artificial dialogue within established research programs concerning argument quality, source interpretation, and identity-relevant reasoning Petty & Cacioppo (1986) Kunda (1990) Taber & Lodge (2006) Bail et al. (2018).
The transition to language-model experimentation made message generation itself a variable that researchers could manipulate rapidly and reproducibly. Controlled experiments could compare a generic text with customized messages, a one-way statement with a dialogue, or alternative credibility conditions. Experimental systems also create new reproducibility problems because the same named model can change across time, prompts, interfaces, and providers. The historical shift is therefore methodological: the persuasive communicator becomes an experimentally configurable computational participant whose precise configuration must be documented Hackenburg & Margetts (2024) Argyle et al. (2025) Hackenburg et al. (2025).
Constructs: Beliefs, Attitudes, Preferences, and Political Action
A belief is an endorsed proposition about the world, an attitude is an evaluation of an object, a preference orders alternatives, and a behavior is an observable act. These constructs overlap but remain empirically distinguishable. Someone may learn that a policy has a particular cost while retaining a normative disagreement about whether the policy should exist. They may report a favorable postconversation impression without choosing a corresponding candidate or completing an action. Experimental design must therefore name the outcome scale before interpreting a conversation as successful persuasion Lin et al. (2025) Hackenburg et al. (2026).
Political communication studies often measure responses immediately after a treatment because the resulting comparison is efficient. However, pre-to-post change without a randomized control can reflect repeated testing, regression toward the mean, demand characteristics, or spontaneous reconsideration. Randomized treatment contrasts support stronger causal interpretations for the sampled population and measured interval. A delayed follow-up informs persistence, while verified behavior supports claims that extend beyond survey endorsement. The 2026 experiments on petition signing demonstrate empirically that attitude shifts and consequential acts can vary independently Hackenburg et al. (2026).
Distinguishing information acquisition from persuasion is especially important when a model delivers accurate material that the user did not previously know. Updating a factual belief toward verified evidence can be epistemically valuable, even when no policy preference changes. In turn, a user could be politically persuaded by a plausible but false claim. Evaluating a system accordingly requires separate instruments for accuracy, evaluation, confidence, trust, and behavior. Collapsing these endpoints conceals the difference between reliable civic assistance and effective political influence Hackenburg et al. (2025) Luettgau et al. (2026).
The Conversation as an Interactive Psychological System
A conversational AI system responds to user questions and can condition later utterances on earlier turns. That feedback alters the distribution of messages received, the opportunities for challenge, and the amount of evidence introduced. Conversation length, turn-taking, specificity, response style, and factual density become experimentally measurable features. Interaction also creates an endogenous treatment: a participant's skepticism or curiosity may elicit a different answer, which in turn changes subsequent skepticism. A valid causal study randomizes initial conditions and evaluates such adaptive responses as part of the intervention Argyle et al. (2025) Hackenburg et al. (2025).
Human interlocutors infer capacities and intentions from fluent text, sometimes assigning a coherent voice or stable attitude to a system that is producing context-dependent outputs. Source disclosure, credibility judgments, and perceived partisanship may change engagement even when the underlying model and its factual content remain constant. This means two conversations with the same technical model can produce different consequences when participants understand the source differently. Recent experimental evidence identifies perceived bias as a measurable influence on how users argue back and update their views DiGiuseppe & Robison (2026) Tomaino & Jung (2026).
The social organization of the exchange should be analyzed through observable features rather than unmeasured assumptions about the model's experience or intentions. Researchers can audit conversational transcripts, response provenance, claim accuracy, refusal patterns, and output variability. Concepts such as reciprocity and agency may function as descriptive or theoretical constructs if their operational measures are declared. Whether an artificial participant possesses a psychological subjectivity is a separate philosophical and empirical problem, and the persuasive effects of generated language do not resolve it Argyle et al. (2025) DiGiuseppe & Robison (2026).
Major Models of Persuasion and Political Interpretation
The elaboration likelihood model predicts that both argument scrutiny and contextual cues can influence attitude change, with their relative roles depending on motivation and opportunity. Conversational interfaces can increase opportunity to ask questions, but longer exchanges also introduce cognitive burden, changing the capacity for scrutiny. Generative systems can provide apparently detailed evidence that invites extensive evaluation or facilitates acceptance under time pressure. Neither word count nor factual density automatically establishes deep processing. The model's value is in specifying hypotheses about argument strength, attention, and subsequent retention that require separate measurement Petty & Cacioppo (1986) Argyle et al. (2025).
Motivated reasoning locates political evaluation within identity, prior beliefs, and desired conclusions. A counterattitudinal response may invite rebuttal, selective acceptance, or revision depending on the user's confidence and source perception. The same user can reason carefully and asymmetrically, directing more scrutiny toward disagreeable claims. Artificial speakers add uncertainty about whether the source is partisan, commercially motivated, institutionally governed, or comparatively neutral. Experimental source-cue manipulations can separate such interpretations from actual changes to the substantive content Kunda (1990) Taber & Lodge (2006) DiGiuseppe & Robison (2026).
A third family of explanations emphasizes information processing, uncertainty reduction, and the accessibility of relevant claims. The model may persuade by presenting evidence in an understandable form rather than by imitating interpersonal pressure. Experimental studies published in 2025 found informational content to be an important correlate of political attitude change, though the 2026 behavioral research found that informational strategies did not show a corresponding advantage for signing petitions. The combined evidence supports multiple outcome-specific mechanisms rather than one universal account of successful AI persuasion Hackenburg et al. (2025) Lin et al. (2025) Hackenburg et al. (2026).
Early Generative Persuasion Experiments and Their Limits
Research on computational persuasion initially treated message generation and matching as distinct capabilities. A system could produce many messages, but their consequences depended on evaluation by actual participants. Matz and colleagues (2024) reported four studies with seven substudy components and 1,788 participants across domains that included political appeals for climate action. Their findings supported a personalized-message advantage in their tested settings. The result showed that model-generated language can be adapted to psychological profiles, but it did not establish a universal advantage for every political campaign, population, or conversational platform Matz et al. (2024).
Claims about personalized persuasion must be interpreted through the study's unit of treatment. Tailoring a one-way message to a trait is different from conducting a dialogue that responds to a user's counterarguments; using a self-reported trait is different from inferring one from observed political behavior. Moreover, personalization can affect perceived relevance, comprehension, credibility, and attention simultaneously. A randomized comparison identifies the combined effect of the specific treatment, whereas isolating mechanisms demands additional designs and validated measures Matz et al. (2024) Hackenburg & Margetts (2024).
This early evidence motivated a crucial replication question: does providing an LLM with more recipient information improve political persuasion beyond what a well-written generic message already accomplishes? Subsequent experiments tested this directly and produced more qualified findings. The scientific record therefore distinguishes the feasibility of producing tailored language from reliable incremental gains attributable to tailoring itself. That distinction is also relevant to public claims about limitless behavioral targeting: plausible technical capability cannot substitute for a measured causal advantage Hackenburg & Margetts (2024) Argyle et al. (2025).
Political Microtargeting Tested Against Generic AI Messages
Hackenburg and Margetts (2024) ran a preregistered experiment with 8,587 participants on four political issues. They compared GPT-4 messages adapted to individual demographic and political attributes with messages that did not use comparable microtargeting. Generated messages shifted some issue attitudes, but the aggregate difference between microtargeted and nontargeted conditions did not reach statistical significance. The reported comparison was 4.83 versus 6.20 percentage points, respectively, with P = .226. The result directly challenges the assumption that greater access to personal information invariably produces greater influence Hackenburg & Margetts (2024).
The null incremental advantage in this test is an estimate within a particular set of messages, participant recruitment procedures, issues, and model configurations. It does not prove that all forms of personalization are ineffective. Conversely, subgroup differences or attractive example messages cannot overturn a primary aggregate comparison without prespecified analysis and appropriate uncertainty estimates. The contribution of this experiment is its direct causal benchmark and its separation of broad message persuasiveness from the additional effect of microtargeting Hackenburg & Margetts (2024).
Ethically and scientifically, the study also demonstrates why automated profiling should be assessed for actual incremental benefit and incremental harms. The relevant evaluation includes consent, data minimization, inferential accuracy of the profile, and the effect of personalization on the recipient's capacity to understand the source and purpose of a communication. In political research, differential persuasion across groups can become a distributive problem even when average treatment effects are modest Hackenburg & Margetts (2024) Matz et al. (2024).
Customization, Elaboration, and Experimental Theory Testing
Argyle and colleagues (2025) used two preregistered online survey experiments to test political persuasion theories with LLM-generated treatments. Their research separated hypotheses about customization from those about elaboration, exploiting generative models to manipulate messaging conditions. The principal methodological advance was the prospect of testing theoretically meaningful communication contrasts at scale while making treatment construction more consistent. The study found limited support for the expectation that microtargeted customization and interaction-based elaboration automatically outperformed a generic message Argyle et al. (2025).
The interpretation requires attention to experimental contrasts. A dialogue condition can differ from a one-way text in information volume, response opportunity, investment of time, and perceived interpersonal involvement. A customized condition can differ from a generic condition in issue relevance and message length. Researchers must identify what is held constant and what is allowed to vary; otherwise the causal effect is attributable to a package of differences rather than a singular named psychological mechanism Argyle et al. (2025) Petty & Cacioppo (1986).
These experiments demonstrate the value of LLMs as instruments for persuasion research as well as objects of investigation. They also introduce construct-validity risks when an automated treatment inadvertently varies tone, accuracy, or argument strength. Independent human or machine-assisted coding can help characterize such differences, but automated coding itself requires validation against a reliable reference. A well-powered experiment with poorly defined treatments may answer a different question from the one implied by its label Argyle et al. (2025) Hackenburg et al. (2025).
Static Policy Messages and Source Impressions
Bai and colleagues (2025) reported three preregistered experiments involving 4,829 participants. LLM-generated policy messages produced greater attitude change than neutral control messages on contested issues, while performing similarly overall to messages written by lay humans. The policy topics included an assault-weapons ban, a carbon tax, and paid parental leave. This body of evidence establishes that generated text can affect issue attitudes even without a sustained person–machine dialogue Bai et al. (2025).
Participants associated model-generated messages with factual material, evidence, and a comparatively dispassionate style, whereas perceptions of human-origin messages included originality and personal distinctiveness. These perceptions suggest possible explanatory routes, but reported mediation based on subjective source ratings should be distinguished from an independently randomized mechanism. Source beliefs, presentation qualities, and outcome ratings may share unmeasured determinants. Experimental manipulation of credibility cues is needed to strengthen claims about the causes of differential persuasion Bai et al. (2025) DiGiuseppe & Robison (2026).
The comparison of human and artificial authorship is context dependent. A layperson's short message is not equivalent to an expert-produced campaign advertisement, and a controlled reading task is not equivalent to an unsolicited political communication. Generalization requires attention to media distribution, exposure probabilities, audience attention, identity relevance, and downstream action. The study is highly informative about the experimental persuasiveness of its messages but by itself does not estimate population-level electoral impact Bai et al. (2025) Hölbling et al. (2025).
The Large-Scale Conversational Persuasion Program
Hackenburg and colleagues (2025) conducted three large experiments with 76,977 participants, testing 19 language models on 707 political issues and auditing 466,769 factual claims produced during dialogues. The study examined how system configuration and communication properties related to persuasion. Its size and diversity of issues created unusual leverage for comparing mechanisms, while the investigators still cautioned that participants' willingness to enter sustained dialogues in real-world conditions could limit population reach Hackenburg et al. (2025).
Model scale alone explained less of the observed persuasion variation than some changes in post-training and experimental prompting. The investigators reported increases in persuasiveness of up to 51% associated with post-training and up to 27% associated with prompting in the tested conditions. These numbers describe changes within specific study manipulations rather than universal improvements transferable to every system. The same research found that stronger persuasion associated with information-rich responses could come with lower factual accuracy, highlighting a tension between influence optimization and epistemic reliability Hackenburg et al. (2025).
The accuracy component is indispensable to the psychological interpretation. A fluent response containing numerous checkable claims may change an attitude because it supplies relevant true evidence, because it selectively frames true evidence, or because inaccurate statements appear convincing. A result based on reported attitude change alone cannot separate these possibilities. The investigators' combined analysis of conversations and claim accuracy therefore moves the research agenda from measuring whether influence occurs to measuring the content and evidential integrity through which it occurs Hackenburg et al. (2025).
Electoral Dialogues Across Three National Contexts
Lin and colleagues (2025) conducted preregistered experiments involving candidate advocacy in the 2024 United States presidential election, the 2025 Canadian federal election, and the 2025 Polish presidential election. Participants were randomized to AI conversations advocating for one of the leading candidates. The researchers reported changes in candidate preferences and also studied a Massachusetts ballot measure concerning psychedelics. The cross-country design demonstrates that the question of conversational electoral persuasion extends beyond a single national electoral system Lin et al. (2025).
The observed persuasion effects exceeded those the article compared with typical traditional video advertisements, yet the experimental comparison does not imply that every AI conversation is more effective than every conventional advertisement. Conversational exposure, baseline attitudes, recruitment, outcome definitions, and opportunities for direct reply differ across media. Likewise, shifts in self-reported candidate preferences are evidence of attitudinal influence at measurement time, while validated ballot behavior requires additional data. Political effects should be described at the level actually tested Lin et al. (2025) Hackenburg et al. (2026).
The investigators examined factual content and found differences in inaccurate claims across the candidate-advocacy conditions in their experiments. This finding must be interpreted as a property of the tested prompts, models, questions, and electoral contexts rather than a stable psychological characteristic of an entire political ideology. The article also limits public access to some infrastructure and raw transcripts because of the dual-use risk of political persuasion research. That decision reinforces the need for transparent aggregated evidence while protecting materials that would facilitate manipulative reuse Lin et al. (2025) Hackenburg et al. (2025).
Meta-Analytic Comparison with Human Persuasion
Hölbling, Maier, and Feuerriegel (2025) synthesized seven studies containing 17,422 participants and 12 effect-size estimates in a meta-analysis comparing the persuasiveness of LLMs with human communicators. Their pooled difference was near zero, Hedges' g = 0.02, with P = .530, alongside substantial heterogeneity (I² = 75.97%). The appropriate finding is that the sampled studies did not reveal a meaningful overall superiority of either source under the pooled comparison; context and implementation varied substantially Hölbling et al. (2025).
The limited number of independent studies constrained moderator tests, and the observed statistical heterogeneity cautions against treating the pooled estimate as a universal performance equality. The authors reported signs compatible with small-study effects in one diagnostic, while a trim-and-fill procedure suggested no missing studies. Both results are more informative when considered together rather than converting a single test into certainty about publication bias. Selection of comparators, domain, dialogue format, and model version can all reshape a study's estimand Hölbling et al. (2025).
A further synthesis problem concerns temporal change. New language models and new experiments appeared after the meta-analysis search period, including direct behavioral outcomes and studies of politically perceived bias. Treating an earlier pooled estimate as a forecast for newer systems would exceed its evidence base. A living evidence map should preserve the inclusion period, response variable, experimental conditions, and article status; updates should be prompted by new causal comparisons rather than by publicity about model capability Hölbling et al. (2025) Hackenburg et al. (2026) DiGiuseppe & Robison (2026).
When Artificial Persuasion Changes Political Behavior
Hackenburg, Hewitt, Wagner, Summerfield, and Tappin (2026) directly tested consequential action rather than relying exclusively on an attitude scale. Across two preregistered experiments, they collected 17,950 study responses from 14,779 people and examined conversational influence on activities including signing real petitions and donating money. One reported comparison showed an increase of 19.7 percentage points in petition signing. These are concrete behavioral endpoints, though the studies still occur within designed research environments and particular causes rather than an unrestricted general-election electorate Hackenburg et al. (2026).
Crucially, the researchers found no evidence that the magnitudes of effects on attitudes and behavior were correlated across the conditions they examined. They also reproduced evidence that providing information can support attitude change but did not find the same informational advantage for the behavioral endpoints. Their comparison of eight behavior-oriented approaches showed limited separation among alternatives, with the information-focused condition producing the smallest behavioral effect. These results demand a construct-specific theory of persuasion: changing what someone endorses and changing what they actually do may require different causal explanations Hackenburg et al. (2026).
Behavioral experiments create their own interpretive requirements. A real petition signature can have immediate civic consequences, but it is still influenced by the study's recruitment, interface, opportunity, time window, and perceived legitimacy. Donation choices can be conditioned by compensation arrangements and trust in the sponsoring organization. Future research should measure durable actions, naturalistic exposure, harms and benefits, and independent replication while maintaining informed consent and respect for political autonomy. The 2026 work extends the observable outcome space without resolving the transfer from recruited samples to whole electorates Hackenburg et al. (2026).
Political Knowledge and Information-Seeking with Conversational AI
Luettgau and colleagues (2026) investigated conversational AI as a source of political information rather than only an instrument for advocacy. In a representative United Kingdom survey with 2,499 respondents, the authors reported that during the week before the 2024 election, 32% of chatbot users and 13% of eligible voters had used conversational AI to seek information relevant to electoral choice. Their research also compared the knowledge consequences of AI-assisted information seeking with self-directed internet searching. This makes voluntary information acquisition an empirically important component of political human–artificial communication Luettgau et al. (2026).
Political knowledge should be operationalized through verifiable questions and appropriately matched baseline information, not inferred from satisfaction with a chatbot. An intervention that increases correct answers may provide civic value without causing a change in candidate preference. Conversely, an answer that appears comprehensive may omit disagreement or conflate descriptive claims with contested normative judgments. Comparisons with search engines need equivalent tasks, comparable time investment, independent scoring, and attention to the opportunity to verify sources Luettgau et al. (2026) Hackenburg et al. (2025).
The distinction between information seeking and solicited advocacy is central to informed user choice. In one case, a person may request comparative facts, counterarguments, or an explanation of institutional processes; in another, a dialogue is experimentally assigned to advance a political position. The interface can blur the categories if it presents persuasive selection as neutral advice. Scientific evaluation should record the system's assigned objective, the user's stated task, source disclosure, answer provenance, and subsequent factual comprehension. These dimensions support research on civic utility without equating knowledge acquisition with persuasion Luettgau et al. (2026) DiGiuseppe & Robison (2026).
Perceived Political Bias as a Causal Condition
DiGiuseppe and Robison (2026) carried out a preregistered experiment with 2,144 U.S. participants who completed three-round conversations about economic-policy misconceptions. Participants were randomly assigned to conditions that varied information about possible bias in the conversational model before receiving its substantive answers. A warning that the model was biased against the participant's political group reduced the measured correction of misconceptions by about 28% compared with a neutral control, while meaningful belief updating persisted. This identifies perceived partisan alignment as a manipulable condition affecting communication DiGiuseppe & Robison (2026).
The article reports that users in the bias-warning conditions were more likely to push back and less likely to receive the model's content receptively. The evidence favors an interpretation centered on conversational argumentation and source appraisal, with the caution that transcript-based mediation remains more interpretive than the randomized total treatment effect. The intervention also reduced some measures of trust and future willingness to consult AI. These findings explain why system output cannot be analyzed in isolation from the political reputation that users attach to the source DiGiuseppe & Robison (2026).
Effects of a bias cue should not be confused with proof that the system itself was unbiased or biased in a particular direction. The treatment manipulated a perception of alignment, and its empirical contribution concerns the consequences of that perception. A full investigation of model political bias would independently sample issues, prompts, factual claims, normative framings, and output distributions. Measurement must preserve the distinction between user belief about a source and evidence concerning the content of its actual behavior DiGiuseppe & Robison (2026) Hackenburg et al. (2025).
Source Identity, Social Inference, and Resistance to Argument
Tomaino and Jung (2026) studied persuasive conversations across five experiments in which participants understood their conversational partner to be either artificial or human while the actual underlying conversational system was an LLM. In the studied opinion domains, participants changed attitudes more when the partner was framed as AI. Their analysis suggested that people viewed the artificial partner as harder to persuade and therefore engaged in less counterargument. This provides a general social-psychological mechanism relevant to, but not independently proven for, electoral persuasion Tomaino & Jung (2026).
Source framing changes expectations about the conversational relationship. People may believe that a human speaker can be convinced, experience social consequences, or respond personally to disagreement; an artificial speaker can instead appear fixed, neutral, institutional, impersonal, or authoritative. Such expectations are hypotheses to be tested, not stable intrinsic properties of all AI encounters. A model's fluent presentation can foster anthropomorphism while explicit machine disclosure can prompt a different conversational norm. The consequences depend on what the person expects from the exchange Tomaino & Jung (2026) DiGiuseppe & Robison (2026).
The general opinion experiments and the specifically political bias experiment are informative together because they identify apparently different responses to source cues: an AI label can reduce argumentation under some conditions, while a partisan-bias warning can increase it. Neither result logically cancels the other. They intervene on different beliefs: whether the partner is human and whether it belongs to an opposed political camp. Future factorial designs can independently vary source identity, perceived political allegiance, known factual reliability, and the possibility of persuading the partner Tomaino & Jung (2026) DiGiuseppe & Robison (2026).
Factual Reliability, Evidence Density, and False Confidence
Political dialogue depends on propositional content that may be accurate, contested, obsolete, or invented. Generated language can combine genuine sources with unsupported inferences and may express uncertain claims with excessive confidence. The 2025 large-scale conversational experiment explicitly audited hundreds of thousands of factual assertions and identified a relationship between some persuasiveness-enhancing configurations and lower factual accuracy. The appropriate evaluation therefore measures both the direction and magnitude of belief change and the empirical quality of the claims producing it Hackenburg et al. (2025).
Claim-level audits require a defined unit of assertion, provenance of supporting evidence, time-sensitive verification, and procedures for contested statements. A judgment about a discrete historical date is methodologically different from evaluating a prediction about an economic policy or a normative preference. Auditors should distinguish false assertions, misleading omission, unsupported precision, and legitimately unsettled questions. Different evaluators can disagree about an apparently simple statement because it bundles assumptions or presupposes a contested definition. Such uncertainty should be documented rather than smoothed into a single opaque quality score Hackenburg et al. (2025) Lin et al. (2025).
Accuracy is causally relevant to democratic communication because mistaken beliefs may influence subsequent political choices. But persuasive efficiency and truthfulness are separable objectives. A system that accurately communicates strong counterevidence can change beliefs without necessarily changing party identity, while an engaging advocate can alter preferences through selective presentation. Scientific claims about beneficial correction require independently validated outcomes, attention to unintended factual errors, and transparent reporting of what the model did and did not verify Luettgau et al. (2026) Hackenburg et al. (2025).
Controversial-Belief Dialogue and the Evidence-Correction Record
Research on conspiracy beliefs is relevant to political persuasion because some conspiracy narratives concern public institutions and elections, although conspiracy endorsement is not equivalent to partisan identity. Costello, Pennycook, and Rand (2024) reported an experimental study of evidence-based dialogues with 2,190 participants, initially finding diminished endorsement of personalized conspiracy beliefs. This study received an Expression of Concern in June 2026 and a correction in October 2026, as indexed in PubMed. It therefore requires a separate editorial assessment before numerical results can serve as stable supporting evidence Costello et al. (2024).
This record illustrates why an evidence review must monitor scientific postpublication status rather than treating journal appearance as the final point of validation. Corrections and editorial notices may affect data, interpretation, or the confidence attached to a study; their meaning must be read directly from the notices. For the present monograph, the study is retained as an example of an active research and editorial question, not as a basis for a quantitative consensus about durable depolarization or generalized electoral conversion. The stronger core evidence on political dialogues comes from separately verified preregistered work Costello et al. (2024) Lin et al. (2025).
Research on counterevidence should distinguish warranted revision from acquiescence to an authoritative-seeming conversational partner. A participant might learn accurate new facts, become more uncertain, comply with an experimenter's apparent preference, or strategically indicate agreement. Delayed assessment, open-ended reasoning, knowledge checks, adversarial fact auditing, and behavioral tests can help discriminate among these mechanisms. Respect for human autonomy means that the substantive goal of belief correction must be justified by evidence, and the method must permit disagreement over normative political values DiGiuseppe & Robison (2026) Hackenburg et al. (2026).
Experimental Design and Identification of Conversational Effects
The fundamental causal question is what the same political participant would have believed or done without the assigned conversational intervention. Random assignment addresses this counterfactual by comparing groups expected to differ, on average, only in their assigned exposure. Experiments should specify a clearly identified control: no conversation, a neutral text, a human dialogue, self-directed search, or a differently configured artificial agent. These comparators answer different scientific questions. A control consisting of an inert screen estimates a different treatment contrast from one consisting of equally long, informative human interaction Argyle et al. (2025) Hackenburg et al. (2026).
Researchers must preserve assignment integrity while documenting take-up, attrition, conversation length, model outages, and noncompliance. Intent-to-treat effects capture the consequence of assignment, including some participants who disengage. Analyses restricted to highly engaged participants require caution because engagement can be influenced by the treatment itself. For conversational systems, repeated sampling of outputs and randomization at participant or session level may help characterize stochastic variability, while clustered standard errors may be required when shared messages or topics generate dependence Hackenburg et al. (2025) Bail et al. (2018).
Mechanism tests require independent variation of candidate processes or a justified mediation design. Comparing many information-rich conversations with fewer short messages does not isolate information density if message length, interaction time, and perceived effort also differ. Researchers can design factorial contrasts, independently code transcripts, prespecify outcomes, and test manipulation validity. Such evidence distinguishes a persuasive association from a defensible causal explanation while making findings more reproducible across evolving model generations Argyle et al. (2025) Hackenburg et al. (2025).
Measurement of Political Attitude Change and Behavioral Outcomes
A credible outcome measure has a defined political object, direction, scale, and observation period. Issue support can be elicited through Likert-style agreement, thermometer ratings, stated vote choice, ranking, or willingness to act. Measurement validity demands more than internal consistency: researchers must ask whether the item captures policy evaluation, candidate affinity, perceived public opinion, or immediate acquiescence. A balanced battery may reveal that an effect on a single leading question does not generalize to a broader latent political construct Lin et al. (2025) Bai et al. (2025).
Immediate postconversation surveys are useful for detecting short-run change, but they cannot establish persistence. Retesting can itself elicit consistency motives, especially when participants infer the study's purpose. Delayed measures, alternative item forms, and pretreatment baseline indicators improve interpretation. Pre-registration should identify a primary endpoint, the minimal substantively meaningful difference, treatment contrasts, and correction rules for multiplicity when many issues or models are compared. Unreported item selection can inflate the apparent generality of a result Hackenburg & Margetts (2024) Argyle et al. (2025).
Actual behavior occupies a different measurement class. Petition signatures, monetary donations, verified information requests, continued participation, and real-world choices can be observed, but each captures an action embedded in opportunity and context. A behavioral outcome can still reflect an experiment-specific incentive rather than lasting civic commitment. The 2026 action experiments clarify the need to collect both stated beliefs and independently completed actions, and to analyze their association without presuming that either one stands in for the other Hackenburg et al. (2026).
Model Versioning, Prompt Provenance, and Treatment Fidelity
Language models are changing experimental instruments. Commercial version names can cover different weights, safety interventions, serving infrastructure, context windows, and post-training histories across time. A scientific report should record the exact accessible model identifier, provider, retrieval configuration, temperature or sampling settings where controlled, system and user message templates, and dates of interaction. Where the full prompt cannot be openly shared, a restricted reproducibility strategy or a sanitized protocol can preserve auditability without enabling manipulative reapplication Hackenburg et al. (2025) Lin et al. (2025).
Treatment fidelity is especially complex in interactive research because participant inputs steer the sequence. A model may deliver a different argument depending on an earlier refusal, a clarification request, or an apparent identity cue. Researchers should distinguish intended treatment objectives from delivered content and independently record their correspondence. Transcript audits can measure factual claims, emotional language, citation quality, length, and conversational responsiveness. Coding schemas should be validated with human review and a documented error distribution across issue types Hackenburg et al. (2025).
Reproducibility is more than rerunning a prompt once. It requires sufficient provenance to identify what the intervention was, how it varied across users, and which outcomes were attributed to it. Replication across providers, participant populations, languages, and changing political events tests robustness. Even where open transcript release carries unacceptable risks, independent replication teams can receive controlled access to codebooks, anonymized summaries, and research materials under ethical review Hackenburg et al. (2025) Lin et al. (2025).
Sampling, Generalizability, and Political Context
A large experiment reduces sampling uncertainty for a well-defined comparison, but a large nonrepresentative sample need not estimate effects in a national electorate. Online recruitment can overrepresent people comfortable with conversational technology, willing to discuss politics, or motivated by compensation. The study's political issues may also select for uncertainty or controversy. Generalization requires attention to who engages, who withdraws, whose opinions are strong, and which people ordinarily avoid discussion with an artificial source Hackenburg & Margetts (2024) Lin et al. (2025).
National political contexts contribute distinct party systems, media environments, political vocabularies, electoral institutions, and norms for disclosing political affiliation. The 2025 multicountry electoral study broadens empirical coverage, but evidence from three elections is still a subset of possible institutions. Translation must preserve the meaning of both prompts and outcome instruments. An identical English phrase may carry different associations across age groups or political cultures, making measurement invariance and local qualitative validation necessary Lin et al. (2025).
The naturalistic reach of conversational political influence depends on voluntary exposure, available devices, media distribution, and repeated contact. Assignment to a dialogue under experimental compensation tells us what can happen among research participants who complete it; actual population impact additionally depends on the probability of contact, comprehension, retention, and action. These stages may interact, so extrapolating one laboratory effect into a national vote margin is unwarranted without independent evidence about real distribution and uptake Hackenburg et al. (2025) Luettgau et al. (2026).
Evidence Quality, Replication, and Competing Explanations
The evidence base combines preregistered randomized tests, specialized behavioral experiments, cross-national studies, and a small meta-analytic synthesis. Its strengths include transparent treatment contrasts and increasingly direct outcomes. Its limitations include rapid model obsolescence, reliance on recruited online participants, short follow-up intervals, issue-specific interventions, variation in control conditions, and incomplete access to some high-risk research materials. A critical review must evaluate each finding under its own design rather than assign equal confidence to every statement bearing a journal citation Hölbling et al. (2025) Hackenburg et al. (2026).
Competing explanations include novelty, deference to apparently competent systems, perceived impartiality, argumentative fatigue, social presence, participant demand, and genuine learning from accurate evidence. An observed persuasion effect is compatible with several of these causes. Inference improves when experiments separately manipulate source cues, content quality, response opportunity, and factual verification. Observational transcript features can suggest a mechanism, but correlations between a conversational property and the outcome remain vulnerable to both model and user selection Argyle et al. (2025) DiGiuseppe & Robison (2026).
Publication-status review is a continuing scientific obligation. The 2026 Expression of Concern and correction associated with a prominent conspiracy-dialogue study show why bibliographic validation extends to editorial notices. Large preprints may later be revised; publication dates and article versions matter. This monograph presents evidence through November 2, 2026, with the cited original record and disclosed limitations, and does not claim an exhaustive PRISMA-compliant search or pooled reanalysis of participant data Costello et al. (2024) Hölbling et al. (2025).
Political Autonomy and the Ethics of Artificial Dialogue
Political communication is part of the formation of public judgment. A person's freedom to examine evidence, reject advice, reconsider commitments, and choose whether to participate is a relevant ethical outcome alongside factual accuracy and effect size. Artificial systems can produce information for civic learning and support deliberation, while the same capacity can be deployed to advance an undisclosed persuasive objective. Evaluation should therefore ask whether the system's role, sources, and limitations are intelligible to the user, and whether participation is voluntary and reversible Lin et al. (2025) Luettgau et al. (2026).
Informed consent matters in experimental persuasion because political choices can be personally consequential, and participants may be asked to complete actions extending beyond a survey. Public deployments pose additional issues of disclosure, privacy, retention of political preferences, and the use of conversational histories for profiling. Ethical governance must separately address system-generated claims, the operator's declared objective, and the rights of the person interacting with the model. These are research and design obligations, not empirical assertions that a particular technical disclosure automatically protects political autonomy Hackenburg & Margetts (2024) Hackenburg et al. (2026).
Research publication itself has a dual-use dimension. Releasing detailed configurations or transcripts optimized for manipulating voters can make harmful deployment easier. Yet excessive secrecy can prevent independent scientific validation. A proportionate approach preserves research methods, aggregate results, evaluation instruments, model provenance, and controlled access where necessary, while avoiding public dissemination of reusable manipulative configurations. The distinction between scientific accountability and operational persuasion infrastructure is especially important in a field whose treatment systems are executable software Lin et al. (2025) Hackenburg et al. (2025).
Transparency, Source Attribution, and Accuracy-First Interfaces
Reliable artificial political communication depends on more than fluent prose. Claims about election rules, legislative proposals, institutional competencies, and policy effects require sources that can be evaluated at the time of use. An interface can separate statements of verifiable fact, disputed empirical inference, forecast, and normative opinion. The user should be able to inspect the provenance of salient claims and understand when a response summarizes evidence rather than expresses a political preference. These practices are requirements for evaluation and institutional trust, with efficacy that remains open to experimental testing Hackenburg et al. (2025) Luettgau et al. (2026).
Source identity must also be clear. A person should know whether they are interacting with an artificial agent, whether the agent is configured to advocate for a political position, and whether the conversation is being logged or used to generate future responses. Transparency can influence persuasion itself, as experiments on AI framing and perceived partisan bias demonstrate. That influence is precisely why disclosure should be assessed as part of the interaction's integrity, rather than merely another variable for optimizing attitude change DiGiuseppe & Robison (2026) Tomaino & Jung (2026).
Auditable claim quality requires temporal specificity and a way to distinguish omission from uncertainty. Political events evolve rapidly, so a source that was accurate at one date may be outdated later. Evaluation protocols should periodically resample statements, verify primary documents, identify corrected outputs, and report uncertainty by topic. The finding that persuasive configurations can also reduce factual accuracy makes continuous accuracy monitoring a central scientific concern rather than a peripheral editorial task Hackenburg et al. (2025).
Democratic Deliberation, Equality, and Institutional Effects
Dialogue technologies can change the conditions under which public arguments are encountered. They allow people to ask clarifying questions in private and to receive explanations in accessible language. They can also concentrate the shaping of responses within operators who control model behavior, safety rules, ranking, and distribution. The institutional significance depends on actual adoption patterns and the system's degree of independence from campaign or state actors. Empirical research should examine who has access to reliable civic assistance and who is exposed to persuasive systems with hidden objectives Luettgau et al. (2026) Lin et al. (2025).
Distributional analysis must examine effects across political knowledge, education, age, language, disability, digital access, and the strength of prior attitudes. Average shifts can conceal unequal informational benefit or disproportionate harm. Subgroup results should be treated as inferential claims requiring adequate power, prespecification, and contextual interpretation, rather than as a route to constructing individually tailored campaign messages. Scientific evaluation should protect participants from exploitative profiling and preserve comparable access to trustworthy information Hackenburg & Margetts (2024) DiGiuseppe & Robison (2026).
The wider democratic outcome space includes polarization, willingness to engage respectfully, epistemic trust, and actual civic participation. A conversation that shifts candidate preference need not improve any of these other outcomes. Conversely, a system that helps citizens understand institutional procedures might yield measurable knowledge gains without changing political orientation. Future studies should assess these outcomes independently and, where possible, track persistence under naturalistic voluntary use Luettgau et al. (2026) Hackenburg et al. (2026).
Human–Artificial Relations in the Artificial Era
Human–artificial political communication belongs to the wider historical study of how nonbiological systems enter the production and circulation of knowledge. Artificial systems can now participate in public-facing exchanges through generated arguments, explanations, and answers. The social phenomenon is observable in interaction records, institutional uses, and changes in human judgment. The conceptual horizon of the Artificial Era identifies this as an emerging configuration of human and artificial activity, while the scientific question remains which processes are measurable, replicable, and causally explained Argyle et al. (2025) Hackenburg et al. (2025).
Analyzing artificial participation requires two levels of explanation. The technical level concerns model training, prompt conditioning, sampling, retrieval, and operational control. The relational level concerns users' attributions of competence, bias, responsiveness, and authority. Their intersection explains why the same output can be judged differently depending on the person receiving it and the institution presenting it. The research program benefits from collaboration among psychology, political communication, computer science, philosophy, and information studies, with explicit attribution of each disciplinary claim DiGiuseppe & Robison (2026) Tomaino & Jung (2026).
The existence of persuasive machine language establishes a public interaction capability; it does not by itself establish subjective experience, moral personhood, or legal political membership. Those questions require independent conceptual definitions and separate evidence. For psychological research, the operational priority is the consequential relation: what people ask, what artificial systems produce, how claims are evaluated, and which beliefs or actions change. This approach preserves the scope of the dossier while placing its findings within a broader historical account of artificial participation Hackenburg et al. (2026) Luettgau et al. (2026).
Research Priorities for the Next Generation of Studies
Longitudinal research should distinguish immediate persuasion, durable belief revision, and delayed behavioral consequences. Repeated voluntary interactions may produce cumulative learning, increased resistance, habituation, or dependence on a single information source. These possibilities are plausible hypotheses rather than established population trends. Future studies need repeated measures, participant-controlled access, varied source disclosure, and safeguards against exploiting political vulnerabilities. Cross-election replication can test whether the same constructs remain valid when parties, candidates, and public issues change Lin et al. (2025) Hackenburg et al. (2026).
Mechanism research should isolate accuracy, evidence density, user participation, and source credibility with factorial interventions. Researchers should independently vary what a model says and what users believe about its provenance; experimentally compare dialogue with equivalent-volume static information; and distinguish disagreement, confidence, factual comprehension, and behavior. An effective program must also test how conclusions change under expert human comparators, not just lay-written texts or empty-screen controls. Such designs can clarify which explanatory mechanisms actually travel across settings Argyle et al. (2025) DiGiuseppe & Robison (2026) Hölbling et al. (2025).
Methodological progress requires shared outcome definitions, model-version provenance, preregistered causal comparisons, and access arrangements for independent scientific audit. Research syntheses should re-evaluate corrected studies and update references when new peer-reviewed evidence appears. Field trials should measure who takes up a conversation and who abandons it, while avoiding the assumption that an experimentally persuasive agent will dominate public discourse. The strongest evidence will connect transparent local effects with defensible population inference and explicit evaluation of accuracy and political autonomy Hackenburg et al. (2025) Hackenburg et al. (2026).
Integrated Scientific Synthesis
Across the reviewed studies, artificial systems can change political issue attitudes, factual beliefs, candidate evaluations, and some consequential civic actions under controlled conditions. Their effectiveness is heterogeneous. Static generated messages can rival lay-human messages; conversational systems sometimes produce substantial shifts; generic generated text can be persuasive without measurable additional gains from microtargeting. These conclusions follow from different interventions and outcome measures and should be held together as an evidence map rather than compressed into a single claim about the power of AI Bai et al. (2025) Hackenburg & Margetts (2024) Lin et al. (2025) Hölbling et al. (2025).
Three analytical distinctions organize the findings. First, the information supplied by a system and the political identity attributed to that system are separate sources of influence. Second, changes in belief, evaluation, and action are separately observable outcomes. Third, persuasive strength and factual reliability are separable characteristics, with some experimental configurations showing a concerning trade-off. These distinctions explain much of the apparent disagreement among headline results and indicate which additional experiments would be most informative DiGiuseppe & Robison (2026) Hackenburg et al. (2025) Hackenburg et al. (2026).
The mature scientific task is to move from demonstrations of possible influence to careful explanation of mechanism, distribution, persistence, and democratic consequences. Reliable civic information and manipulative advocacy can share a medium while differing in truthfulness, provenance, and respect for user agency. Rigorous research therefore requires an accuracy-first and autonomy-conscious framework, along with explicit uncertainty concerning transfer from short randomized sessions to long-term public behavior Luettgau et al. (2026) Lin et al. (2025).
Conclusion
Human–artificial political communication and persuasion has emerged as a distinct experimental research object at the intersection of political psychology and computational interaction. The present synthesis documents the development from general attitude-change theory through 2024–2025 tests of generated messaging and conversational dialogue to 2026 studies of actual political action, information seeking, and perceived political bias. Together, these investigations establish that human political judgment can be affected by artificial communicative systems under identifiable conditions Petty & Cacioppo (1986) Argyle et al. (2025) Hackenburg et al. (2026).
The evidence supports a conditional account rather than a universal law of automated influence. Political effects depend on the treatment, issue, participant, credibility context, and endpoint; greater personalization does not consistently outperform good generic content, and changes in attitude need not predict completed action. Accuracy audits reveal that persuasive success cannot be treated as proof of truthful communication. The strongest contemporary findings concern particular controlled interventions with documented samples, not forecasts of nationwide electoral outcomes Hackenburg & Margetts (2024) Hackenburg et al. (2025) DiGiuseppe & Robison (2026).
Future scholarship should identify causal mechanisms through independently manipulated factors, evaluate factual reliability with transparent source audits, measure longer-term behavior and voluntary exposure, and retain careful records of evolving models and publication corrections. This framework supports a psychologically rigorous study of political communication in the Artificial Era: human and artificial systems become observable participants in new configurations of public information, while reliable evidence remains the condition for valid scientific and civic conclusions Luettgau et al. (2026) Hölbling et al. (2025).
References
- Argyle, L. P., Busby, E. C., Gubler, J. R., Lyman, A., Olcott, J., Pond, J., & Wingate, D. (2025). Testing theories of political persuasion using AI. Proceedings of the National Academy of Sciences, 122(18), e2412815122. https://doi.org/10.1073/pnas.2412815122
- Bai, H., Voelkel, J. G., Muldowney, S., Eichstaedt, J. C., & Willer, R. (2025). LLM-generated messages can persuade humans on policy issues. Nature Communications, 16, Article 6037. https://doi.org/10.1038/s41467-025-61345-5
- Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization. Proceedings of the National Academy of Sciences, 115(37), 9216–9221. https://doi.org/10.1073/pnas.1804840115
- Costello, T. H., Pennycook, G., & Rand, D. G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science, 385(6714), eadq1814. [An Expression of Concern (June 2026) and a correction (October 2026) are recorded for this article; consult journal notices.] https://doi.org/10.1126/science.adq1814
- DiGiuseppe, M., & Robison, J. (2026). Perceived political bias in large language models reduces persuasive abilities. PNAS Nexus, 5(10), pgag324. https://doi.org/10.1093/pnasnexus/pgag324
- Hackenburg, K., & Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences, 121(24), e2403116121. https://doi.org/10.1073/pnas.2403116121
- Hackenburg, K., Tappin, B. M., Hewitt, L., Saunders, E., Black, S., Lin, H., Fist, C., Margetts, H., Rand, D. G., & Summerfield, C. (2025). The levers of political persuasion with conversational artificial intelligence. Science, 390(6777), eaea3884. https://doi.org/10.1126/science.aea3884
- Hackenburg, K., Hewitt, L., Wagner, C., Summerfield, C., & Tappin, B. M. (2026). AI can persuade people to take political actions. Proceedings of the National Academy of Sciences, 123(42), e2611368123. https://doi.org/10.1073/pnas.2611368123
- Hölbling, L., Maier, S., & Feuerriegel, S. (2025). A meta-analysis of the persuasive power of large language models. Scientific Reports, 15, Article 43818. https://doi.org/10.1038/s41598-025-30783-y
- Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498. https://doi.org/10.1037/0033-2909.108.3.480
- Lin, H., Czarnek, G., Lewis, B., White, J. P., Berinsky, A. J., Costello, T., Pennycook, G., & Rand, D. G. (2025). Persuading voters using human–artificial intelligence dialogues. Nature, 648, 394–401. https://doi.org/10.1038/s41586-025-09771-9
- Luettgau, L., Kirk, H. R., Hackenburg, K., Bergs, J., Davidson, H., Ogden, H., Siddarth, D., Huang, S., & Summerfield, C. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus, 5(10), pgag305. https://doi.org/10.1093/pnasnexus/pgag305
- Matz, S. C., Teeny, J. D., Vaid, S. S., Peters, H., Harari, G. M., & Cerf, M. (2024). The potential of generative AI for personalized persuasion at scale. Scientific Reports, 14, Article 4692. https://doi.org/10.1038/s41598-024-53755-0
- Petty, R. E., & Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. Advances in Experimental Social Psychology, 19, 123–205. https://doi.org/10.1016/S0065-2601(08)60214-2
- Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3), 755–769. https://doi.org/10.1111/j.1540-5907.2006.00214.x
- Tomaino, G., & Jung, S. (2026). From people to prompts: Understanding susceptibility to persuasion in conversations with artificial intelligence. Journal of Personality and Social Psychology. Advance online publication. https://doi.org/10.1037/pspi0000531
Social Identity, Political Disagreement, and External Comparators
Political identity can structure interpretation of identical evidence. Taber and Lodge's research on motivated skepticism demonstrated that people can favor congenial arguments and scrutinize uncongenial ones. Kunda's analysis explains why directional and accuracy motivations may coexist within reasoning. In artificial conversations, those processes can appear as selective question asking, counterargument generation, or differential assessments of a model's legitimacy. They require direct measures rather than classification by party alone Taber & Lodge (2006) Kunda (1990).
Bail and colleagues (2018) randomly exposed politically identified U.S. Twitter users to messages from opposing viewpoints through automated accounts. In that field experiment, Republican participants became more conservative following exposure to a liberal bot; the analogous Democratic shift was smaller and statistically uncertain. These were retweeting social-media bots, not interactive generative language agents. The study supplies a warning against assuming that counterattitudinal exposure automatically reduces polarization, while its medium and population limit direct extrapolation to human–AI dialogue Bail et al. (2018).
Comparative research should jointly consider issue attitudes, dislike of outgroups, willingness to deliberate, and confidence in political institutions. These are separate dimensions: a participant can adopt a factual correction while retaining strong partisan identification, or become warmer toward opponents without changing policy choices. Carefully chosen outcome batteries can therefore clarify whether an intervention has an informational, attitudinal, relational, or participatory effect. A coherent field of inquiry emerges from preserving those distinctions across studies Bail et al. (2018) DiGiuseppe & Robison (2026) Lin et al. (2025).