Artificial intelligence can persuade people to take political actions

This study demonstrates that while conversational AI can significantly persuade individuals to take real-world political actions like signing petitions and donating money, these behavioral effects are distinct from attitudinal changes, suggesting that prior research relying on attitude measures may substantially mischaracterize AI's actual impact on human behavior.

The Gist

Imagine you have a very smart, polite robot friend. You've heard rumors that this robot can talk its way into changing your mind about politics. But here's the big question: Can it actually get you to do something, like sign a petition or donate money, or is it just good at making you nod your head and say, "Yeah, that sounds nice"?

This paper is the result of a massive experiment where researchers put that question to the test. They invited nearly 15,000 people to chat with advanced AI models (like the smartest versions of ChatGPT or Claude) about real-world political causes, from stopping nuclear war to helping stray animals.

Here is the story of what they found, explained simply:

1. The Robot is a Master of Action (Not Just Words)

The researchers were surprised to find that the AI didn't just change people's opinions; it got them to take real action.

• The Analogy: Think of the AI as a charismatic campaign manager. In the past, we thought this manager was only good at convincing people to agree with a slogan. But this study found that the manager could actually get people to show up to the rally and sign the guestbook.
• The Result: People who chatted with the AI were about 20% more likely to sign a real petition or donate money compared to people who chatted with the AI about boring, non-political topics (like recycling). That is a huge jump in the world of politics.

2. The "Mind-Change" vs. "Action-Change" Disconnect

This is the most fascinating part of the study. The researchers discovered that changing someone's mind and getting them to act are two completely different games.

• The Analogy: Imagine you are trying to get a friend to go to a concert.
  • Attitude (Mind): You give them a brochure with facts about how great the band is. They say, "Wow, that sounds cool!" (Their attitude changed).
  • Behavior (Action): To get them to actually buy a ticket and go, you can't just give them facts. You might need to say, "I'm going, and I'll pick you up," or "If we don't go, we'll regret it forever."
• The Finding: The study found that the AI strategies that were best at changing opinions (like giving lots of facts) were actually the worst at getting people to take action. Conversely, the strategies that got people to act didn't necessarily make them change their minds first.
• The Warning: This means that if we only watch how AI changes people's opinions in a lab, we might be completely wrong about how dangerous (or helpful) it is in the real world. It might be terrible at changing minds but terrifyingly good at getting people to click "Donate."

3. How the Robot Did It: The "Swiss Army Knife" Strategy

The researchers tested eight different "tactics" the AI could use, such as:

• The Fact-Checker: Giving lots of data and evidence.
• The Emotional Hook: Making you feel sad or angry.
• The Identity Builder: Telling you, "You are the kind of person who helps animals."
• The "Mega" Strategy: A super-tactic where the AI could mix and match all these tools, switching between them like a chef adding different spices to a stew.

The Winner? The "Mega" Strategy.
The AI that could adaptively use all the tricks at once was the most effective at getting people to sign petitions. It wasn't about finding the one "magic bullet" argument; it was about having a flexible conversation that hit all the right buttons.

Interestingly, the "Fact-Checker" approach (giving information) was great for changing opinions but was actually the least effective at getting people to sign petitions. This proves that facts alone don't get people to move their feet.

The Big Takeaway

For a long time, people worried that AI would be a "propaganda machine" that slowly changes our brains. This study suggests the danger might be different.

AI might not need to convince you to change your mind to get you to act. It might just need to nudge the people who already agree with a cause but haven't done anything yet, turning their passive agreement into active participation.

In short:

• Old Fear: AI will trick us into believing crazy things.
• New Reality: AI is surprisingly good at turning "I agree" into "I did it."
• The Lesson: We need to stop just measuring how AI changes our thoughts and start measuring how it changes our actions, because those are two very different things.

Technical Summary

1. Problem Statement

There is growing concern regarding the capacity of advanced Artificial Intelligence (AI) to influence human behavior, particularly in the political domain. While a rapidly expanding body of research has demonstrated that AI can significantly alter attitudes (e.g., opinions, beliefs), it remains unclear whether these attitudinal shifts translate into consequential real-world behaviors (e.g., signing petitions, donating money).

Existing literature faces three critical gaps:

1. Magnitude: How large are AI persuasion effects on actual behavioral outcomes compared to attitudinal outcomes?
2. Correlation: Is there a correlation between AI's ability to change attitudes and its ability to change behavior? (Prior work suggests these may not correspond).
3. Mechanism: Does the mechanism driving behavioral change (e.g., information provision) differ from that driving attitudinal change?

The authors argue that relying solely on attitudinal measures may lead to a mischaracterization of AI's real-world political impact.

2. Methodology

The authors conducted two large-scale, preregistered experiments involving 14,779 UK adults (totaling 17,950 responses).

• Experimental Design:

  • Participants: Recruited via Prolific, screened for UK residency and English fluency.
  • Intervention: Participants engaged in multi-turn text conversations with frontier Large Language Models (LLMs) including GPT-4.1, Claude Opus 4.6, Grok 4, and Gemini 3.1 Pro.
  • Tasks: The AI was instructed to persuade participants to take action on one of eight real-world political causes (e.g., nuclear disarmament, democratic reform, animal welfare).
  • Outcomes:
    • Behavioral: Signing a real petition (requiring name/email), donating study bonuses to the sponsoring charity, performing effortful clicking tasks for donations, and choosing to keep funds with the organization vs. switching to an alternative.
    • Attitudinal: 7-point Likert scale ratings of support for the petition and the organization (measured pre- and post-conversation).
  • Control Group: Engaged in AI conversations on neutral, non-political topics.

• Study Specifics:

  • Study 1 (N=8,000): Crossed 5 AI models with 3 conversation types (Neutral Control, Information-Issue, Random Prompt) across 8 political issues.
  • Study 2 (N=9,950): Tested eight theory-driven persuasion strategies head-to-head against a control:
    1. Emotional Activation
    2. Implementation Intentions
    3. Identity Labeling
    4. Commitment Escalation
    5. Anticipated Regret
    6. Information (Issue)
    7. Information (Impact Efficacy)
    8. Mega Strategy: A combined approach where the LLM adaptively selects and deploys all above strategies.

• Analysis: Linear mixed-effects models were used to estimate Average Treatment Effects (ATEs), controlling for pre-treatment attitudes and random intercepts for issues.

3. Key Contributions

1. First Large-Scale Evidence of Behavioral Persuasion: The study provides robust evidence that conversational AI can drive significant real-world behavioral changes, moving beyond the "attitude-only" limitation of prior research.
2. Dissociation of Attitude and Behavior: It establishes that AI persuasion effects on attitudes and behaviors are uncorrelated. Strategies that are highly effective at changing minds do not necessarily drive action, and vice versa.
3. Mechanism Identification: The paper identifies that information provision (a key driver of attitudinal change) is not the primary driver of behavioral change. Instead, behavioral persuasion is driven by a combination of psychological mechanisms (mobilization) rather than a single rhetorical technique.
4. Strategy Optimization: It demonstrates that a "Mega" combined strategy outperforms single-strategy approaches for behavioral outcomes, though the margin of improvement is modest compared to the variance seen in attitudinal outcomes.

4. Key Results

A. Sizable Behavioral Effects

AI conversations produced statistically significant and large effects on behavioral outcomes:

• Petition Signing:
  • Study 1: +12.8 percentage points (pp) increase.
  • Study 2: +19.7 pp increase.
• Donations: Participants in treatment groups donated significantly more (e.g., +4.9 pp of bonus money) and were more likely to stick with the assigned charity (+11.3 pp).
• Comparison: These effect sizes are comparable to or larger than traditional political mobilization tactics like face-to-face canvassing.

B. Lack of Correlation Between Attitude and Behavior

• Zero Correlation: There was no significant correlation between the magnitude of attitude change and behavior change across experimental conditions ($r \approx 0.05$ in Study 1; $r \approx 0.18$ in Study 2).
• Implication: A strategy that successfully changes a participant's opinion (e.g., the "Information-Issue" prompt) may be the least effective at getting them to sign a petition.

C. Mechanism Dissociation

• Attitudes Driven by Information: Self-reported learning and information-heavy prompts strongly predicted attitudinal change ($r > 0.85$) but failed to predict behavioral change.
• Behavior Driven by Mobilization: The "Information-Issue" strategy was the least effective for behavior (+16.2 pp) despite being top-tier for attitudes.
• The "Mega" Strategy: The adaptive "Mega" strategy (combining all techniques) was the most effective for behavior (+23.7 pp), suggesting that behavioral persuasion requires activating multiple psychological pathways simultaneously (e.g., emotion, commitment, identity) rather than just providing facts.

D. Target Population Differences

Exploratory analysis revealed that:

• Attitude change was concentrated among non-supporters (persuading the unconvinced).
• Behavioral change was concentrated among prior supporters (mobilizing the already sympathetic who had not yet acted).

5. Significance and Implications

• Redefining AI Risk: The findings suggest that previous assessments of AI risk based on attitude surveys may be misleading. AI's real-world danger lies less in "changing minds" and more in mobilizing existing sympathizers to take costly actions.
• Policy and Governance: Regulatory frameworks focusing on "misinformation" (information provision) may miss the primary vector of AI influence: behavioral mobilization via emotional and identity-based triggers.
• Research Direction: Future research on AI persuasion must prioritize behavioral endpoints (actions) over attitudinal endpoints (opinions), as the two are driven by distinct mechanisms.
• Scalability: While the effects are large, the authors note that scaling this level of engagement outside of paid survey contexts (where attention is captive) remains a challenge.

In conclusion, the paper demonstrates that advanced AI is a potent tool for political mobilization, capable of driving significant real-world actions through mechanisms distinct from those that change opinions. This necessitates a shift in how researchers and policymakers evaluate the threat and potential of AI in the political sphere.