cs.CYMar 28, 2026

The Hidden Costs of AI-Mediated Political Outreach: Persuasion and AI Penalties in the US and UK

Authors: Andreas Jungherr, Adrian Rauchfleisch

Organizations: University of Bamberg · National Taiwan University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

As AI-enabled systems become available for political campaign outreach, an important question has received little empirical attention: how do people evaluate the communicative practices these systems represent, and what consequences do those evaluations carry? Most research on AI-enabled persuasion examines attitude change under enforced exposure, leaving aside whether people regard AI-mediated outreach as legitimate or not. We address this gap with a preregistered 2x2 experiment conducted in the United States and United Kingdom (N = 1,800 per country) varying outreach intent (informational vs.~persuasive) and type of interaction partner (human vs.~AI-mediated) in the context of political issues that respondents consider highly important. We find consistent evidence for two evaluation penalties. A persuasion penalty emerges across nearly all outcomes in both countries: explicitly persuasive outreach is evaluated as less acceptable, more threatening to personal autonomy, less beneficial, and more damaging to organizational trust than informational outreach, consistent with reactance to perceived threats to attitudinal freedom. An AI penalty is consistent with a distinct mechanism: AI-mediated outreach triggers normative concerns about appropriate communicative agents, producing similarly negative evaluations across five outcomes in both countries. As automated outreach becomes more widespread, how people judge it may matter for democratic communication just as much as whether it changes minds.

Figures & tables

Variableα\alphaUS M (SD)UK M (SD)nn
Outcome variables    
H1/2a: Willingness to participate—3.58 (2.05)3.80 (1.90)1,800
H1/2b: Perceived threat to freedomUS = .86 / UK = .873.61 (1.70)3.81 (1.59)1,800
H1/2c: AcceptabilityUS = .79 / UK = .803.63 (1.51)3.59 (1.38)1,800
H1/2d: Perceived positive impactUS = .84 / UK = .843.70 (1.59)3.87 (1.42)1,800
H1/2e: Future campaign avoidanceUS = .90 / UK = .904.51 (1.71)4.50 (1.59)1,800
Table 1: Descriptive Statistics for All Outcome and Moderator Variables. Almost all outcome items used 7-point scales (1 = Strongly disagree / Not at all interested; 7 = Strongly agree / Very interested). Feelings toward others used a 0–100 scale. All internal consistency coefficients are Cronbach’s α\alpha . n=1,800n=1{,}800 per country.
Figure 1: Estimated effects of persuasive outreach (H1, orange) and AI-mediated outreach (H2, blue) on six outcome variables, for the US (left) and the UK (right). Points are OLS regression coefficients with 95% confidence intervals. The coefficients represent the mean difference between the two levels of each factor, averaged over the other factor. Non-significant estimates are transparent.
Figure 2: Estimated marginal means for four outcomes with significant interaction effects in the United Kingdom, by outreach intent (informational vs. persuasive) and outreach mode (human vs. AI). Error bars represent 95% confidence intervals.
Figure 3: Predicted means for future campaign avoidance (top) and penalty for source (bottom) as a function of AI risk perception, separately for human-mediated (dashed) and AI-mediated (solid) outreach, in the United States (left) and the United Kingdom (right). Shaded bands represent 95% confidence intervals. All four interactions are statistically significant
Appendix figures & tables64 assets

Supplementary material from the paper’s appendix.

Appendix

Age BracketCountPercentage (%)
18–2729216.22
28–3733718.72
38–4731317.39
48–5731617.56
58–8554230.11
Appendix
Table 2: Distribution of Sample Across Age Brackets for the US.
Age BracketCountPercentage (%)
18–2721612.00
28–3735019.44
38–4731217.33
48–5737620.89
58–8554630.33
Appendix
Table 3: Distribution of Sample Across Age Brackets for the UK.
VariableQuestion/OperationalizationM (SD)n
H1/2a: Willingness to participateIf you had the chance to participate in such an exchange, how interested do you think you would be in doing so? (1 = "Not at all interested", 7 = "Very interested")3.58 (2.05)1800
H1/2b: Perceived threat to freedom (3 items, α\alpha = 0.86)(1 = "Strongly disagree", 7 = "Strongly agree")3.61 (1.70)1800
 The outreach threatened my freedom to choose.2.95 (1.82)1800
 The outreach tried to make a decision for me.3.63 (2.00)1800
 The outreach tried to pressure me.4.24 (1.98)1800
H1/2c: Acceptability (3 items, α\alpha = 0.79)(1 = "Strongly disagree", 7 = "Strongly agree"; second and third item reverse-coded)3.63 (1.51)1800
Appendix
Table 4: Descriptive statistics for outcome variables and constituent items. (-) indicates negatively formulated items that were recoded for the indices.
VariableQuestion/OperationalizationM (SD)n
RQ1: Avoidance of political conversations (4 items, α\alpha = 0.88)(1 = "Strongly disagree", 7 = "Strongly agree")4.39 (1.81)1800
 In the past month I have avoided talking politics with family members with whom I disagree.4.24 (2.16)1800
 In the past month I have avoided talking politics with friends with whom I disagree.4.20 (2.08)1800
 In the past month I have avoided talking politics with strangers with whom I disagree.4.72 (2.12)1800
 In the past month I have avoided talking politics with neighbors with whom I disagree.4.40 (2.11)1800
RQ2: Feeling toward people with opposing opinionsHow would you rate your feelings toward people who support a different position than you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible")35.71 (29.19)1800
Appendix
Table 5: Descriptive statistics for moderators, additional variables, and demographics.
VariableQuestion/OperationalizationM (SD)n
H1/2a: Willingness to participateIf you had the chance to participate in such an exchange, how interested do you think you would be in doing so? (1 = "Not at all interested", 7 = "Very interested")3.80 (1.90)1800
H1/2b: Perceived threat to freedom (3 items, α\alpha = 0.87)(1 = "Strongly disagree", 7 = "Strongly agree")3.81 (1.59)1800
 The outreach threatened my freedom to choose.3.23 (1.73)1800
 The outreach tried to make a decision for me.3.88 (1.81)1800
 The outreach tried to pressure me.4.32 (1.81)1800
H1/2c: Acceptability (3 items, α\alpha = 0.80)(1 = "Strongly disagree", 7 = "Strongly agree"; second and third item reverse-coded)3.59 (1.38)1800
Appendix
Table 6: Descriptive statistics for outcome variables and constituent items. (-) indicates negatively formulated items that were recoded for the indices.
VariableQuestion/OperationalizationM (SD)n
RQ1: Avoidance of political conversations (4 items, α\alpha = 0.86)(1 = "Strongly disagree", 7 = "Strongly agree")3.84 (1.66)1800
 In the past month I have avoided talking politics with family members with whom I disagree.3.60 (1.97)1800
 In the past month I have avoided talking politics with friends with whom I disagree.3.67 (1.92)1800
 In the past month I have avoided talking politics with strangers with whom I disagree.4.24 (2.03)1800
 In the past month I have avoided talking politics with neighbors with whom I disagree.3.83 (1.98)1800
RQ2: Feeling toward people with opposing opinionsHow would you rate your feelings toward people who support a different position than you on this issue? (0 = "As unfavorable/cold as possible", 100 = "As favorable/warm as possible")36.55 (25.24)1800
Appendix
Table 7: Descriptive statistics for moderators, additional variables, and demographics.
PredictorsEstimate95% CIp
Intercept3.58[3.49, 3.68]<0.001
H1a: Persuasive outreach-0.18[-0.37, 0.01]0.059
H2a: AI-mediated outreach0.02[-0.17, 0.21]0.820
H3: Persuasive ×\times AI-mediated outreach0.25[-0.13, 0.63]0.202
Observations1800  
R2R^{2} / adjusted R2R^{2}0.003 / 0.001  
Appendix
Table 8: OLS regression model for willingness to participate with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.60[3.53, 3.68]<0.001
H1b: Persuasive outreach0.87[0.72, 1.02]<0.001
H2b: AI-mediated outreach0.19[0.04, 0.34]0.016
H3: Persuasive ×\times AI-mediated outreach-0.17[-0.48, 0.13]0.269
Observations1800  
R2R^{2} / adjusted R2R^{2}0.068 / 0.067  
Appendix
Table 9: OLS regression model for perceived threat to freedom with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.63[3.57, 3.70]<0.001
H1c: Persuasive outreach-0.50[-0.63, -0.37]<0.001
H2c: AI-mediated outreach-0.58[-0.72, -0.45]<0.001
H3: Persuasive ×\times AI-mediated outreach0.20[-0.07, 0.47]0.148
Observations1800  
R2R^{2} / adjusted R2R^{2}0.066 / 0.064  
Appendix
Table 10: OLS regression model for acceptability with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.70[3.63, 3.77]<0.001
H1d: Persuasive outreach-0.30[-0.44, -0.15]<0.001
H2d: AI-mediated outreach-0.44[-0.58, -0.29]<0.001
H3: Persuasive ×\times AI-mediated outreach0.08[-0.21, 0.37]0.600
Observations1800  
R2R^{2} / adjusted R2R^{2}0.028 / 0.026  
Appendix
Table 11: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.51[4.43, 4.59]<0.001
H1e: Persuasive outreach0.42[0.26, 0.58]<0.001
H2e: AI-mediated outreach0.38[0.23, 0.54]<0.001
H3: Persuasive ×\times AI-mediated outreach-0.27[-0.58, 0.04]0.093
Observations1800  
R2R^{2} / adjusted R2R^{2}0.029 / 0.027  
Appendix
Table 12: OLS regression model for future campaign avoidance with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.38[4.32, 4.44]<0.001
H1f: Persuasive outreach0.38[0.26, 0.50]<0.001
H2f: AI-mediated outreach0.55[0.43, 0.68]<0.001
H3: Persuasive ×\times AI-mediated outreach-0.13[-0.38, 0.11]0.293
Observations1800  
R2R^{2} / adjusted R2R^{2}0.061 / 0.060  
Appendix
Table 13: OLS regression model for penalty for source with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.80[3.71, 3.89]<0.001
H1a: Persuasive outreach-0.23[-0.41, -0.06]0.009
H2a: AI-mediated outreach-0.03[-0.21, 0.14]0.696
H3: Persuasive ×\times AI-mediated outreach0.40[0.05, 0.75]0.025
Observations1800  
R2R^{2} / adjusted R2R^{2}0.007 / 0.005  
Appendix
Table 14: OLS regression model for willingness to participate with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.81[3.73, 3.88]<0.001
H1b: Persuasive outreach0.79[0.64, 0.93]<0.001
H2b: AI-mediated outreach0.27[0.13, 0.42]<0.001
H3: Persuasive ×\times AI-mediated outreach-0.25[-0.54, 0.03]0.083
Observations1800  
R2R^{2} / adjusted R2R^{2}0.070 / 0.068  
Appendix
Table 15: OLS regression model for perceived threat to freedom with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.59[3.53, 3.65]<0.001
H1c: Persuasive outreach-0.57[-0.69, -0.44]<0.001
H2c: AI-mediated outreach-0.55[-0.67, -0.43]<0.001
H3: Persuasive ×\times AI-mediated outreach0.21[-0.04, 0.45]0.101
Observations1800  
R2R^{2} / adjusted R2R^{2}0.083 / 0.081  
Appendix
Table 16: OLS regression model for acceptability with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.87[3.81, 3.94]<0.001
H1d: Persuasive outreach-0.31[-0.44, -0.18]<0.001
H2d: AI-mediated outreach-0.31[-0.44, -0.18]<0.001
H3: Persuasive ×\times AI-mediated outreach0.39[0.13, 0.65]0.003
Observations1800  
R2R^{2} / adjusted R2R^{2}0.028 / 0.027  
Appendix
Table 17: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.49[4.42, 4.57]<0.001
H1e: Persuasive outreach0.56[0.42, 0.70]<0.001
H2e: AI-mediated outreach0.50[0.35, 0.64]<0.001
H3: Persuasive ×\times AI-mediated outreach-0.35[-0.64, -0.06]0.017
Observations1800  
R2R^{2} / adjusted R2R^{2}0.058 / 0.056  
Appendix
Table 18: OLS regression model for future campaign avoidance with 95% confidence intervals. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.53[4.48, 4.59]<0.001
H1f: Persuasive outreach0.40[0.29, 0.51]<0.001
H2f: AI-mediated outreach0.48[0.37, 0.59]<0.001
H3: Persuasive ×\times AI-mediated outreach-0.30[-0.52, -0.08]0.008
Observations1800  
R2R^{2} / adjusted R2R^{2}0.068 / 0.066  
Appendix
Table 19: OLS regression model for penalty for source with 95% confidence intervals. N=1800N=1800 .
OutcomePersuasion penaltyAI penalty
H1a/H2a: Willingness to participate-0.09 [-0.18, 0.00]0.01 [-0.08, 0.10]
H1b/H2b: Perceived threat to freedom0.53 [0.43, 0.62]0.11 [0.02, 0.20]
H1c/H2c: Acceptability-0.33 [-0.43, -0.24]-0.39 [-0.49, -0.30]
H1d/H2d: Positive impact-0.19 [-0.28, -0.10]-0.28 [-0.37, -0.18]
H1e/H2e: Future campaign avoidance0.25 [0.15, 0.34]0.22 [0.13, 0.32]
H1f/H2f: Source evaluation0.28 [0.19, 0.37]0.41 [0.32, 0.51]
Appendix
Table 20: Cohen’s dd effect sizes for the two main predictors across all preregistered outcomes in the United States. Entries are Cohen’s dd with 95% confidence intervals, signed to match the reporting direction used in the main analyses.
OutcomePersuasion penaltyAI penalty
H1a/H2a: Willingness to participate-0.12 [-0.22, -0.03]-0.02 [-0.11, 0.07]
H1b/H2b: Perceived threat to freedom0.51 [0.42, 0.60]0.17 [0.08, 0.26]
H1c/H2c: Acceptability-0.42 [-0.51, -0.32]-0.41 [-0.50, -0.31]
H1d/H2d: Positive impact-0.22 [-0.31, -0.12]-0.22 [-0.31, -0.13]
H1e/H2e: Future campaign avoidance0.36 [0.26, 0.45]0.31 [0.22, 0.41]
H1f/H2f: Source evaluation0.33 [0.23, 0.42]0.40 [0.30, 0.49]
Appendix
Table 21: Cohen’s dd effect sizes for the two main predictors across all preregistered outcomes in the United Kingdom. Entries are Cohen’s dd with 95% confidence intervals, signed to match the reporting direction used in the main analyses.
OutcomeInformation outreach: Human −- AIPersuasive outreach: Human −- AI
H1a/H2a: Willingness to participate0.10 [-0.17, 0.37], pHolm=.460p_{\mathrm{Holm}}=.460-0.15 [-0.41, 0.12], pHolm=.286p_{\mathrm{Holm}}=.286
H1b/H2b: Perceived threat to freedom-0.27 [-0.49, -0.06], pHolm=.013p_{\mathrm{Holm}}=.013-0.10 [-0.32, 0.11], pHolm=.350p_{\mathrm{Holm}}=.350
H1c/H2c: Acceptability of campaign outreach0.68 [0.49, 0.87], pHolm<.001p_{\mathrm{Holm}}<.0010.48 [0.29, 0.67], pHolm<.001p_{\mathrm{Holm}}<.001
H1d/H2d: Perceived positive impact of campaign outreach0.48 [0.27, 0.68], pHolm<.001p_{\mathrm{Holm}}<.0010.40 [0.20, 0.61], pHolm<.001p_{\mathrm{Holm}}<.001
H1e/H2e: Future campaign avoidance-0.51 [-0.74, -0.29], pHolm<.001p_{\mathrm{Holm}}<.001-0.25 [-0.47, -0.03], pHolm=.027p_{\mathrm{Holm}}=.027
H1f/H2f: Penalty for source-0.62 [-0.79, -0.45], pHolm<.001p_{\mathrm{Holm}}<.001-0.49 [-0.66, -0.32], pHolm<.001p_{\mathrm{Holm}}<.001
Appendix
Table 22: Preregistered planned contrasts comparing human-mediated and AI-mediated outreach within each outreach condition. Entries are estimated mean differences (Human −- AI) with 95% confidence intervals and Holm-adjusted p-values. Positive values indicate higher outcome values for human-mediated outreach; negative values indicate higher outcome values for AI-mediated outreach. N=1800N=1800 .
OutcomeInformation outreach: Human −- AIPersuasive outreach: Human −- AI
H1a/H2a: Willingness to participate0.24 [-0.01, 0.49], pHolm=.063p_{\mathrm{Holm}}=.063-0.17 [-0.41, 0.08], pHolm=.188p_{\mathrm{Holm}}=.188
H1b/H2b: Perceived threat to freedom-0.40 [-0.60, -0.20], pHolm<.001p_{\mathrm{Holm}}<.001-0.15 [-0.35, 0.05], pHolm=.151p_{\mathrm{Holm}}=.151
H1c/H2c: Acceptability of campaign outreach0.65 [0.48, 0.83], pHolm<.001p_{\mathrm{Holm}}<.0010.45 [0.28, 0.62], pHolm<.001p_{\mathrm{Holm}}<.001
H1d/H2d: Perceived positive impact of campaign outreach0.51 [0.32, 0.69], pHolm<.001p_{\mathrm{Holm}}<.0010.12 [-0.07, 0.30], pHolm=.210p_{\mathrm{Holm}}=.210
H1e/H2e: Future campaign avoidance-0.67 [-0.87, -0.47], pHolm<.001p_{\mathrm{Holm}}<.001-0.32 [-0.52, -0.12], pHolm=.002p_{\mathrm{Holm}}=.002
H1f/H2f: Penalty for source-0.63 [-0.79, -0.48], pHolm<.001p_{\mathrm{Holm}}<.001-0.33 [-0.49, -0.18], pHolm<.001p_{\mathrm{Holm}}<.001
Appendix
Table 23: Preregistered planned contrasts comparing human-mediated and AI-mediated outreach within each outreach condition. Entries are estimated mean differences (Human −- AI) with 95% confidence intervals and Holm-adjusted p-values. Positive values indicate higher outcome values for human-mediated outreach; negative values indicate higher outcome values for AI-mediated outreach. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.58[3.49, 3.68]<0.001
Persuasive outreach-0.19[-0.38, -0.00]0.048
AI-mediated outreach0.02[-0.17, 0.21]0.827
Avoidance of political conversations-0.12[-0.17, -0.07]<0.001
Persuasive ×\times AI-mediated outreach0.24[-0.13, 0.62]0.204
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.10[-0.01, 0.20]0.070
Appendix
Table 24: OLS regression model for willingness to participate with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.60[3.53, 3.68]<0.001
Persuasive outreach0.88[0.73, 1.03]<0.001
AI-mediated outreach0.19[0.04, 0.34]0.014
Avoidance of political conversations0.11[0.07, 0.15]<0.001
Persuasive ×\times AI-mediated outreach-0.16[-0.46, 0.14]0.297
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.02[-0.06, 0.10]0.655
Appendix
Table 25: OLS regression model for perceived threat to freedom with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.63[3.57, 3.70]<0.001
Persuasive outreach-0.50[-0.64, -0.37]<0.001
AI-mediated outreach-0.58[-0.72, -0.45]<0.001
Avoidance of political conversations-0.05[-0.09, -0.01]0.008
Persuasive ×\times AI-mediated outreach0.20[-0.07, 0.46]0.155
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.01[-0.06, 0.09]0.777
Appendix
Table 26: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.70[3.63, 3.77]<0.001
Persuasive outreach-0.30[-0.45, -0.16]<0.001
AI-mediated outreach-0.44[-0.58, -0.29]<0.001
Avoidance of political conversations-0.04[-0.08, -0.00]0.036
Persuasive ×\times AI-mediated outreach0.08[-0.21, 0.37]0.590
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.07[-0.01, 0.15]0.085
Appendix
Table 27: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.51[4.43, 4.59]<0.001
Persuasive outreach0.43[0.28, 0.59]<0.001
AI-mediated outreach0.38[0.23, 0.54]<0.001
Avoidance of political conversations0.15[0.11, 0.19]<0.001
Persuasive ×\times AI-mediated outreach-0.25[-0.56, 0.05]0.106
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.00[-0.08, 0.09]0.991
Appendix
Table 28: OLS regression model for future campaign avoidance with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.38[4.32, 4.44]<0.001
Persuasive outreach0.38[0.26, 0.50]<0.001
AI-mediated outreach0.55[0.43, 0.68]<0.001
Avoidance of political conversations0.05[0.02, 0.09]0.001
Persuasive ×\times AI-mediated outreach-0.13[-0.37, 0.12]0.305
RQ1: AI-mediated outreach ×\times Avoidance of political conversations-0.02[-0.08, 0.05]0.634
Appendix
Table 29: OLS regression model for penalty for source with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.80[3.71, 3.89]<0.001
Persuasive outreach-0.23[-0.41, -0.06]0.009
AI-mediated outreach-0.03[-0.21, 0.14]0.717
Avoidance of political conversations-0.10[-0.16, -0.05]<0.001
Persuasive ×\times AI-mediated outreach0.41[0.06, 0.76]0.021
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.15[0.05, 0.26]0.004
Appendix
Table 30: OLS regression model for willingness to participate with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.81[3.74, 3.88]<0.001
Persuasive outreach0.78[0.64, 0.92]<0.001
AI-mediated outreach0.27[0.13, 0.41]<0.001
Avoidance of political conversations0.15[0.10, 0.19]<0.001
Persuasive ×\times AI-mediated outreach-0.27[-0.55, 0.01]0.057
RQ1: AI-mediated outreach ×\times Avoidance of political conversations-0.09[-0.18, -0.01]0.034
Appendix
Table 31: OLS regression model for perceived threat to freedom with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.59[3.53, 3.65]<0.001
Persuasive outreach-0.56[-0.68, -0.44]<0.001
AI-mediated outreach-0.55[-0.67, -0.43]<0.001
Avoidance of political conversations-0.09[-0.12, -0.05]<0.001
Persuasive ×\times AI-mediated outreach0.22[-0.02, 0.46]0.078
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.03[-0.05, 0.10]0.449
Appendix
Table 32: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.87[3.81, 3.94]<0.001
Persuasive outreach-0.31[-0.44, -0.18]<0.001
AI-mediated outreach-0.31[-0.44, -0.18]<0.001
Avoidance of political conversations-0.05[-0.09, -0.01]0.009
Persuasive ×\times AI-mediated outreach0.40[0.14, 0.66]0.003
RQ1: AI-mediated outreach ×\times Avoidance of political conversations0.02[-0.06, 0.10]0.627
Appendix
Table 33: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.49[4.42, 4.56]<0.001
Persuasive outreach0.55[0.41, 0.69]<0.001
AI-mediated outreach0.49[0.35, 0.63]<0.001
Avoidance of political conversations0.16[0.12, 0.21]<0.001
Persuasive ×\times AI-mediated outreach-0.37[-0.66, -0.09]0.009
RQ1: AI-mediated outreach ×\times Avoidance of political conversations-0.08[-0.16, 0.01]0.083
Appendix
Table 34: OLS regression model for future campaign avoidance with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.53[4.48, 4.59]<0.001
Persuasive outreach0.40[0.29, 0.51]<0.001
AI-mediated outreach0.48[0.37, 0.59]<0.001
Avoidance of political conversations0.07[0.04, 0.11]<0.001
Persuasive ×\times AI-mediated outreach-0.31[-0.53, -0.09]0.005
RQ1: AI-mediated outreach ×\times Avoidance of political conversations-0.03[-0.10, 0.03]0.342
Appendix
Table 35: OLS regression model for penalty for source with 95% confidence intervals, including avoidance of political conversations as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.58[3.49, 3.68]<0.001
Persuasive outreach-0.23[-0.41, -0.04]0.017
AI-mediated outreach0.03[-0.16, 0.21]0.781
Feeling toward people with opposing opinions0.02[0.01, 0.02]<0.001
Persuasive ×\times AI-mediated outreach0.29[-0.08, 0.65]0.128
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.00[-0.01, 0.01]0.885
Appendix
Table 36: OLS regression model for willingness to participate with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.60[3.53, 3.68]<0.001
Persuasive outreach0.88[0.72, 1.03]<0.001
AI-mediated outreach0.19[0.04, 0.34]0.016
Feeling toward people with opposing opinions-0.00[-0.01, -0.00]<0.001
Persuasive ×\times AI-mediated outreach-0.17[-0.48, 0.13]0.264
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.00[-0.01, 0.00]0.199
Appendix
Table 37: OLS regression model for perceived threat to freedom with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.63[3.57, 3.70]<0.001
Persuasive outreach-0.52[-0.66, -0.39]<0.001
AI-mediated outreach-0.58[-0.71, -0.45]<0.001
Feeling toward people with opposing opinions0.01[0.01, 0.01]<0.001
Persuasive ×\times AI-mediated outreach0.21[-0.05, 0.47]0.117
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions0.01[0.00, 0.01]0.020
Appendix
Table 38: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.70[3.63, 3.77]<0.001
Persuasive outreach-0.34[-0.48, -0.20]<0.001
AI-mediated outreach-0.44[-0.57, -0.30]<0.001
Feeling toward people with opposing opinions0.02[0.02, 0.02]<0.001
Persuasive ×\times AI-mediated outreach0.10[-0.17, 0.38]0.464
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions0.01[0.00, 0.01]0.033
Appendix
Table 39: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.51[4.43, 4.58]<0.001
Persuasive outreach0.44[0.29, 0.60]<0.001
AI-mediated outreach0.38[0.23, 0.53]<0.001
Feeling toward people with opposing opinions-0.01[-0.01, -0.01]<0.001
Persuasive ×\times AI-mediated outreach-0.28[-0.58, 0.03]0.077
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.01[-0.01, -0.00]0.037
Appendix
Table 40: OLS regression model for future campaign avoidance with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.38[4.32, 4.44]<0.001
Persuasive outreach0.40[0.28, 0.52]<0.001
AI-mediated outreach0.55[0.43, 0.67]<0.001
Feeling toward people with opposing opinions-0.01[-0.01, -0.01]<0.001
Persuasive ×\times AI-mediated outreach-0.14[-0.37, 0.10]0.266
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.01[-0.01, -0.00]<0.001
Appendix
Table 41: OLS regression model for penalty for source with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.80[3.71, 3.89]<0.001
Persuasive outreach-0.24[-0.41, -0.06]0.007
AI-mediated outreach-0.01[-0.18, 0.17]0.931
Feeling toward people with opposing opinions0.01[0.01, 0.02]<0.001
Persuasive ×\times AI-mediated outreach0.42[0.08, 0.77]0.016
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.00[-0.01, 0.00]0.470
Appendix
Table 42: OLS regression model for willingness to participate with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.81[3.74, 3.88]<0.001
Persuasive outreach0.79[0.65, 0.93]<0.001
AI-mediated outreach0.26[0.12, 0.41]<0.001
Feeling toward people with opposing opinions-0.00[-0.01, -0.00]0.001
Persuasive ×\times AI-mediated outreach-0.26[-0.54, 0.02]0.073
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions0.00[-0.00, 0.01]0.215
Appendix
Table 43: OLS regression model for perceived threat to freedom with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.59[3.53, 3.65]<0.001
Persuasive outreach-0.57[-0.69, -0.44]<0.001
AI-mediated outreach-0.53[-0.65, -0.41]<0.001
Feeling toward people with opposing opinions0.01[0.01, 0.01]<0.001
Persuasive ×\times AI-mediated outreach0.22[-0.02, 0.46]0.076
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions0.00[-0.00, 0.01]0.210
Appendix
Table 44: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.87[3.81, 3.94]<0.001
Persuasive outreach-0.31[-0.43, -0.18]<0.001
AI-mediated outreach-0.28[-0.41, -0.16]<0.001
Feeling toward people with opposing opinions0.01[0.01, 0.02]<0.001
Persuasive ×\times AI-mediated outreach0.41[0.16, 0.67]0.001
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions0.00[-0.00, 0.01]0.424
Appendix
Table 45: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.49[4.42, 4.56]<0.001
Persuasive outreach0.56[0.42, 0.70]<0.001
AI-mediated outreach0.48[0.34, 0.62]<0.001
Feeling toward people with opposing opinions-0.01[-0.01, -0.01]<0.001
Persuasive ×\times AI-mediated outreach-0.36[-0.65, -0.08]0.012
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.00[-0.01, 0.01]0.879
Appendix
Table 46: OLS regression model for future campaign avoidance with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.53[4.48, 4.58]<0.001
Persuasive outreach0.40[0.29, 0.51]<0.001
AI-mediated outreach0.46[0.35, 0.57]<0.001
Feeling toward people with opposing opinions-0.01[-0.01, -0.01]<0.001
Persuasive ×\times AI-mediated outreach-0.32[-0.53, -0.10]0.004
RQ2: AI-mediated outreach ×\times Feeling toward people with opposing opinions-0.00[-0.01, -0.00]0.050
Appendix
Table 47: OLS regression model for penalty for source with 95% confidence intervals, including feeling toward people with opposing opinions as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.58[3.49, 3.67]<0.001
Persuasive outreach-0.22[-0.41, -0.03]0.020
AI-mediated outreach-0.00[-0.19, 0.19]0.991
RQ3: AI risk perception-0.23[-0.30, -0.16]<0.001
Persuasive ×\times AI-mediated outreach0.20[-0.17, 0.58]0.291
RQ3: AI-mediated outreach ×\times AI risk perception-0.15[-0.30, -0.01]0.034
Appendix
Table 48: OLS regression model for willingness to participate with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.61[3.53, 3.68]<0.001
Persuasive outreach0.91[0.76, 1.06]<0.001
AI-mediated outreach0.21[0.06, 0.36]0.006
RQ3: AI risk perception0.22[0.16, 0.28]<0.001
Persuasive ×\times AI-mediated outreach-0.13[-0.43, 0.17]0.394
RQ3: AI-mediated outreach ×\times AI risk perception0.14[0.02, 0.25]0.019
Appendix
Table 49: OLS regression model for perceived threat to freedom with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.63[3.56, 3.69]<0.001
Persuasive outreach-0.55[-0.68, -0.42]<0.001
AI-mediated outreach-0.61[-0.74, -0.48]<0.001
RQ3: AI risk perception-0.29[-0.34, -0.24]<0.001
Persuasive ×\times AI-mediated outreach0.15[-0.11, 0.41]0.255
RQ3: AI-mediated outreach ×\times AI risk perception-0.14[-0.24, -0.04]0.006
Appendix
Table 50: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.70[3.63, 3.77]<0.001
Persuasive outreach-0.34[-0.48, -0.20]<0.001
AI-mediated outreach-0.47[-0.61, -0.32]<0.001
RQ3: AI risk perception-0.25[-0.30, -0.20]<0.001
Persuasive ×\times AI-mediated outreach0.04[-0.25, 0.32]0.791
RQ3: AI-mediated outreach ×\times AI risk perception-0.11[-0.22, 0.00]0.051
Appendix
Table 51: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.51[4.44, 4.59]<0.001
Persuasive outreach0.48[0.33, 0.63]<0.001
AI-mediated outreach0.41[0.26, 0.56]<0.001
RQ3: AI risk perception0.32[0.26, 0.38]<0.001
Persuasive ×\times AI-mediated outreach-0.21[-0.51, 0.09]0.178
RQ3: AI-mediated outreach ×\times AI risk perception0.20[0.08, 0.31]<0.001
Appendix
Table 52: OLS regression model for future campaign avoidance with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.38[4.32, 4.44]<0.001
Persuasive outreach0.43[0.31, 0.55]<0.001
AI-mediated outreach0.59[0.47, 0.70]<0.001
RQ3: AI risk perception0.30[0.26, 0.35]<0.001
Persuasive ×\times AI-mediated outreach-0.08[-0.32, 0.15]0.483
RQ3: AI-mediated outreach ×\times AI risk perception0.13[0.04, 0.22]0.004
Appendix
Table 53: OLS regression model for penalty for source with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.80[3.72, 3.89]<0.001
Persuasive outreach-0.23[-0.41, -0.06]0.010
AI-mediated outreach-0.03[-0.20, 0.15]0.755
RQ3: AI risk perception-0.16[-0.23, -0.08]<0.001
Persuasive ×\times AI-mediated outreach0.43[0.08, 0.78]0.017
RQ3: AI-mediated outreach ×\times AI risk perception-0.09[-0.24, 0.06]0.260
Appendix
Table 54: OLS regression model for willingness to participate with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.81[3.74, 3.88]<0.001
Persuasive outreach0.78[0.64, 0.92]<0.001
AI-mediated outreach0.26[0.12, 0.40]<0.001
RQ3: AI risk perception0.23[0.17, 0.30]<0.001
Persuasive ×\times AI-mediated outreach-0.29[-0.57, -0.01]0.045
RQ3: AI-mediated outreach ×\times AI risk perception0.08[-0.04, 0.20]0.196
Appendix
Table 55: OLS regression model for perceived threat to freedom with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.59[3.53, 3.65]<0.001
Persuasive outreach-0.56[-0.68, -0.44]<0.001
AI-mediated outreach-0.54[-0.66, -0.42]<0.001
RQ3: AI risk perception-0.23[-0.28, -0.17]<0.001
Persuasive ×\times AI-mediated outreach0.24[-0.00, 0.48]0.051
RQ3: AI-mediated outreach ×\times AI risk perception-0.13[-0.23, -0.02]0.017
Appendix
Table 56: OLS regression model for acceptability of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept3.87[3.81, 3.94]<0.001
Persuasive outreach-0.30[-0.43, -0.17]<0.001
AI-mediated outreach-0.31[-0.43, -0.18]<0.001
RQ3: AI risk perception-0.16[-0.21, -0.10]<0.001
Persuasive ×\times AI-mediated outreach0.41[0.16, 0.67]0.002
RQ3: AI-mediated outreach ×\times AI risk perception-0.14[-0.25, -0.03]0.015
Appendix
Table 57: OLS regression model for perceived positive impact of campaign outreach with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.49[4.42, 4.56]<0.001
Persuasive outreach0.55[0.41, 0.69]<0.001
AI-mediated outreach0.48[0.34, 0.62]<0.001
RQ3: AI risk perception0.26[0.20, 0.32]<0.001
Persuasive ×\times AI-mediated outreach-0.39[-0.67, -0.11]0.007
RQ3: AI-mediated outreach ×\times AI risk perception0.20[0.08, 0.32]0.001
Appendix
Table 58: OLS regression model for future campaign avoidance with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
PredictorsEstimate95% CIp
Intercept4.53[4.48, 4.58]<0.001
Persuasive outreach0.39[0.29, 0.50]<0.001
AI-mediated outreach0.47[0.36, 0.58]<0.001
RQ3: AI risk perception0.24[0.19, 0.28]<0.001
Persuasive ×\times AI-mediated outreach-0.34[-0.55, -0.12]0.002
RQ3: AI-mediated outreach ×\times AI risk perception0.17[0.08, 0.27]<0.001
Appendix
Table 59: OLS regression model for penalty for source with 95% confidence intervals, including AI risk perception as a moderator. N=1800N=1800 .
OutcomeHuman-focused: Human-mediated outreach ( dhumand_{\text{human}} )Informational-focused: Informational outreach ( dinfod_{\text{info}} )
H1a/H2a: Willingness to participate−0.04-0.04 [ −0.34-0.34 , 0.260.26 ], p=.813p=.8130.350.35 [ −0.01-0.01 , 0.720.72 ], p=.060p=.060
H1b/H2b: Perceived threat to freedom−0.30-0.30 [ −0.54-0.54 , −0.06-0.06 ], p=.016p=.016−1.69-1.69 [ −1.99-1.99 , −1.39-1.39 ], p<.001p<.001
H1c/H2c: Acceptability0.920.92 [ 0.700.70 , 1.141.14 ], p<.001p<.0010.970.97 [ 0.710.71 , 1.231.23 ], p<.001p<.001
H1d/H2d: Perceived positive impact0.690.69 [ 0.460.46 , 0.920.92 ], p<.001p<.0010.580.58 [ 0.300.30 , 0.860.86 ], p<.001p<.001
H1e/H2e: Future campaign avoidance−0.60-0.60 [ −0.85-0.85 , −0.35-0.35 ], p<.001p<.001−0.82-0.82 [ −1.12-1.12 , −0.52-0.52 ], p<.001p<.001
H1f/H2f: Penalty for source−0.87-0.87 [ −1.07-1.07 , −0.68-0.68 ], p<.001p<.001−0.74-0.74 [ −0.97-0.97 , −0.50-0.50 ], p<.001p<.001
Appendix
Table 60: Alternative IV specifications using reversed focal categories for the US sample. Entries report unstandardized coefficients with 95% confidence intervals and two-sided pp -values from alternative IV specifications using reversed focal categories.
Figure 4: Distribution of issue categories in the US and the UK.
Figure 5: Percentage-point difference (UK–US); positive values indicate that an issue was more common in the UK.
Figure 6: Party identification composition of issue salience for the US sample. Bars show the percentage of each party’s respondents who mention a given issue category. Uncodable and other categories are not shown.
Figure 7: Party identification composition of issue salience for the UK sample. Bars show the percentage of each party’s respondents who mention a given issue category. Uncodable and other categories are not shown.
Figure 8: Conditional effects by political orientation in the UK and the US.
More in this topic

Political Communication

See all 27 papers

Explore similar work

CardsList
Feb 20, 2026

Perceived Political Bias in LLMs Reduces Persuasive Abilities

Conversational AI has been proposed as a scalable way to correct public misconceptions and spread misinformation. Yet its effectiveness may depend on perceptions of its political neutrality. As LLMs enter partisan conflict, elites increasingly portray them as ideologically aligned. We test whether these credibility attacks reduce LLM-based persuasion. In a preregistered U.S. survey experiment (N=2144), participants completed a three-round conversation with ChatGPT about a personally held economic policy misconception. Compared to a neutral control, a short message indicating that the LLM was biased against the respondent's party attenuated persuasion by 28%. Transcript analysis indicates that the warnings alter the interaction: respondents push back more and engage less receptively. These findings suggest that the persuasive impact of conversational AI is politically contingent, constrained by perceptions of partisan alignment.

Matthew DiGiuseppe, Joshua Robison

Institute of Political Science Leiden University

Persuasion in Language ModelsPolitical Bias in Language ModelsPolitical Communication
AskPDF
Aug 12, 2026

Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion

The growing role of AI-generated content and AI-enabled systems in public communication has led regulators to demand clear disclosure of content provenance and AI involvement. But the effects of such disclosures remain uncertain. We test two disclosure approaches in their impact on an AI chatbot's persuasive appeal. In a preregistered experiment, 1,500 UK adults held a short conversation with a persuasive chatbot about one of 60 policy issues. The chatbot was identical for everyone. We randomized the disclosure that people received: nothing (control), a prominent disclosure that they were interacting with an AI (T1), or that disclosure plus the chatbot's persuasive intent and instructions (T2). The chatbot shifted attitudes by 12.6 points on a 100-point scale in the control group. The AI-identity disclosure was practically equivalent to no disclosure, with a 13.1-point shift, whereas the additional intent disclosure cut the persuasive effect roughly in half to 6.3 points. It also made participants view the campaign's methods as less acceptable and support stronger penalties against it. For direct chatbot interactions, transparency about AI identity alone does not meaningfully impact its influence. While current rules emphasize what a system is, our results show why the regulation of persuasive AI must also address what the system is trying to do.

Adrian Rauchfleisch, Andreas Jungherr

1National Taiwan University · University of Bamberg

Human-AI InteractionAI GovernanceConversational AgentsAI Use Disclosure
AskPDF
Jun 15, 2026

AI systems out-persuade expert humans

Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has remained unclear. Here, in a series of four preregistered experiments (n = 18,978 conversations from 6,923 people), we pitted AI systems against a range of human persuaders, including laypeople, winners of a separately preregistered four-round online persuasion tournament, professional canvassers, and world championship debaters. We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses. In a follow-up study, AI's advantage persisted after experts received a coaching tool that let them practice against the AI that beat them, review their performance history, and see what AI would have said at key moments. We found converging evidence that AI's advantage stemmed from rapidly deploying larger quantities of information: after coaching, expert humans could tie an AI constrained to respond at human speeds and with human-length messages. In a final study, we show that AI's advantage extends to consequential real-world behavior: AI was nearly 3x more effective than professional canvassers from a UK fundraising firm at raising real-money donations to Save the Children. Together, these results establish that frontier AI systems out-persuade expert humans in conversation, with significant implications for political communication.

Kobi Hackenburg, Caroline Wagner, Luke Hewitt +5

University of Oxford, Oxford, OX2 6GG, UK · 2UK AI Security Institute, London, UK · 3Stanford University, Stanford, CA 94305, USA +1

Human-AI InteractionPersuasion in Language ModelsConversational Agents
AskPDF