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

1. Problem Statement

While existing literature extensively documents the effectiveness of AI in changing political attitudes (persuasion outcomes), there is a significant gap in understanding how the public evaluates the legitimacy and acceptability of AI-mediated political outreach. Most research focuses on attitude change under enforced exposure, ignoring whether people are willing to engage with such systems or if they view the practice as legitimate.

The authors argue that as AI becomes normalized in campaigns, the societal consequences extend beyond mere opinion shifts. If citizens perceive AI-mediated outreach as illegitimate, threatening to autonomy, or damaging to trust, it could create "hidden costs" that undermine democratic communication, reduce engagement, and erode trust in political institutions, regardless of whether the AI successfully changes minds.

2. Methodology

The study employs a preregistered 2×2 between-subjects factorial experiment conducted in two distinct political environments: the United States (N=1,800) and the United Kingdom (N=1,800).

• Experimental Design:

  • Factor 1: Outreach Intent: Informational (sharing info/answering questions) vs. Persuasive (explicitly trying to change opinions using tailored arguments).
  • Factor 2: Interaction Partner: Human Campaigner vs. AI Chatbot.
  • Context: Participants were asked to identify a political issue personally important to them (high-salience topics). They were then presented with a vignette describing a campaign outreach from an organization holding an opposing view on that specific issue.
  • Disclosure: The study explicitly disclosed both the intent (persuasive vs. informational) and the agent (human vs. AI), reflecting emerging regulatory frameworks (e.g., EU AI Act, US state laws) requiring transparency.

• Measures:
  The study measured six outcome variables on 7-point scales:

  1. Willingness to participate.
  2. Perceived threat to personal freedom/autonomy.
  3. Acceptability of the outreach.
  4. Perceived positive impact on public discourse.
  5. Intention for future campaign avoidance.
  6. Penalty for the source (negative evaluation of the sponsoring organization).

• Moderators: The study tested three individual-level moderators:

  • Avoidance of political conversations.
  • Affective polarization (feelings toward opposing opinions).
  • General AI risk perception.

• Analysis: OLS regression models with effect coding were used to estimate main effects and interaction effects, with separate models for the US and UK.

3. Key Contributions

• Shift from Efficacy to Legitimacy: The paper moves the research focus from "Does AI persuade?" to "How do people judge AI persuasion?" highlighting that legitimacy and acceptability are critical constraints on the scalability of automated persuasion.
• Identification of Two Distinct Penalties: The authors empirically distinguish between a Persuasion Penalty (driven by psychological reactance to intent) and an AI Penalty (driven by normative concerns about the appropriateness of the agent).
• Interaction Dynamics: The study challenges the assumption that negative factors compound (amplification). Instead, it finds evidence of attenuation (substitution) in specific contexts, suggesting that when one negative factor (e.g., persuasive intent) already drives evaluations down, a second factor (AI) has less room to worsen them further.
• Cross-National Comparison: By comparing the US and UK, the paper demonstrates how political communication cultures (e.g., frequency of contact, regulation) moderate these effects.

4. Key Results

A. The Persuasion Penalty (H1)

Explicitly persuasive outreach was evaluated significantly more negatively than informational outreach across nearly all outcomes in both countries.

• Findings: Persuasive intent increased perceived threat to freedom, reduced acceptability, lowered perceived positive impact, increased future avoidance, and damaged trust in the source.
• Exception: In the US, the effect on willingness to participate was not statistically significant (likely due to a baseline floor effect where participants were already reluctant to engage with opposing views).
• Mechanism: Consistent with Reactance Theory; people perceive persuasive intent as a threat to their attitudinal freedom.

B. The AI Penalty (H2)

AI-mediated outreach produced less favorable evaluations compared to human-mediated outreach across five of six outcomes in both countries.

• Findings: AI outreach increased perceived threat to freedom, reduced acceptability, lowered positive impact, increased avoidance, and damaged organizational trust.
• Exception: Similar to the persuasion penalty, the effect on willingness to participate was generally non-significant (except in the UK under persuasive intent).
• Mechanism: Consistent with Normative Violation; people view machines as inappropriate agents for political persuasion, lacking social accountability and authenticity.

C. Interaction Effects (H3)

Contrary to the preregistered hypothesis of amplification (where AI + Persuasion would be worst), the results showed attenuation (substitution).

• US: No significant interaction effects; the two penalties operated independently and additively.
• UK: Significant interactions indicated that the AI penalty was smaller when the outreach was already framed as persuasive. The authors interpret this as a "floor effect": when persuasive intent already drives evaluations very low, the additional "normative shock" of AI has less marginal impact.
• Implication: The combination of AI and persuasion is not always the "worst-case scenario" in a multiplicative sense; in high-reactance environments, the penalties may substitute for one another.

D. Moderators

• AI Risk Perception: Consistently amplified the AI penalty across both countries, particularly for future avoidance and source penalty. Those who view AI as risky react more strongly to its use in politics.
• Political Orientation: In the UK, political orientation moderated the effect of persuasive outreach (left-leaning respondents reacted more negatively). In the US, it moderated the effect of AI-mediated outreach.

5. Significance and Implications

• Regulatory Relevance: The findings support the necessity of transparency regulations (disclosure of AI use). Even with full disclosure, AI-mediated outreach incurs a "normative penalty" that damages trust, suggesting that disclosure alone does not mitigate the negative public reaction.
• Strategic Implications for Campaigns:
  • Campaigns deploying AI for "informational" or "civic engagement" purposes (often viewed as safer) may actually suffer the largest relative penalty because they disrupt an otherwise open evaluation.
  • The "AI Penalty" is distinct from the "Persuasion Penalty." Organizations cannot assume that framing AI outreach as non-partisan or informational will shield them from negative evaluations.
• Democratic Health: The study suggests that the widespread adoption of automated persuasion may face a "legitimacy ceiling." If citizens systematically reject AI-mediated contact as illegitimate, the real-world scalability of these tools may be limited not by their technical ability to persuade, but by the public's refusal to engage with them.
• Theoretical Advancement: The paper provides a nuanced view of human-AI interaction in politics, showing that negative evaluations are driven by both the intent of the message (reactance) and the nature of the sender (normative appropriateness), and that these mechanisms interact differently depending on the political culture.