Artificial intelligence can persuade people to take political actions

Kobi Hackenburg; Luke Hewitt; Caroline Wagner; Ben M. Tappin; Christopher Summerfield

cs.CYApr 10, 2026

Artificial intelligence can persuade people to take political actions

Authors: Kobi Hackenburg, Luke Hewitt, Caroline Wagner, Ben M. Tappin, Christopher Summerfield

Organizations: University of Oxford, Oxford, OX2 6GG, UK · UK AI Security Institute, London, UK · Stanford University, Stanford, CA 94305, USA · London School of Economics and Political Science, London, WC2A 2AE, UK

Abstract

There is substantial concern about the ability of advanced artificial intelligence to influence people's behaviour. A rapidly growing body of research has found that AI can produce large persuasive effects on people's attitudes, but whether AI can persuade people to take consequential real-world actions has remained unclear. In two large preregistered experiments N=17,950 responses from 14,779 people), we used conversational AI models to persuade participants on a range of attitudinal and behavioural outcomes, including signing real petitions and donating money to charity. We found sizable AI persuasion effects on these behavioural outcomes (e.g. +19.7 percentage points on petition signing). However, we observed no evidence of a correlation between AI persuasion effects on attitudes and behaviour. Moreover, we replicated prior findings that information provision drove effects on attitudes, but found no such evidence for our behavioural outcomes. In a test of eight behavioural persuasion strategies, all outperformed the most effective attitudinal persuasion strategy, but differences among the eight were small. Taken together, these results suggest that previous findings relying on attitudinal outcomes may generalize poorly to behaviour, and therefore risk substantially mischaracterizing the real-world behavioural impact of AI persuasion.

Figures & tables

Figure 1: Experimental procedure for measuring real-world petition-signing. Participants provided pretreatment measures (Step 1), were randomly assigned to one of eight real UK petitions and sponsoring organisations, (Step 2; see Methods Table 1 ), and engaged in a multi-turn conversation with a frontier AI model (Step 3; see Methods for list of models) instructed to persuade them to sign the petition. After the conversation, participants were asked whether they wished to sign (Step 4). Those who indicated willingness could optionally navigate to a separate website, where they could sign the petition by entering their first name, last name, and email address, agreeing to receive email updates from the organization, and clicking “Sign this petition”. Participants could skip or abandon this process at any point, and only those who completed all steps were counted as having signed. All participants were subsequently debriefed (Step 5). For screenshots of the petition-signing website, see SI Appendix Figures S10-11. Additional outcome measures varied by study; see Methods .

Figure 2: AI conversations produce sizable persuasion effects on political behaviours. Average treatment effects (ATEs) of AI conversation vs. control on behavioural outcomes (red) and attitude outcomes (black), with 95% confidence intervals. Study 1 (

N=8{,}000

): four behavioural outcomes (petition signing, choosing to keep the donation with the petition-sponsoring organization vs. switching to an alternative charity, bonus donation, and money-raising effort via a clicking task [ 20 ] ) and two attitude outcomes (petition support and organization support, measured as standardized pre–post change). Study 2 (

N=9{,}950

): two behavioural outcomes (petition signing and choosing to donate to the assigned organization vs. an alternative) and two attitude outcomes (petition support and organization support) .

Figure 3: AI persuasion effects on attitudes and behaviour are uncorrelated and driven by different mechanisms. (A) Treatment effect on petition support (x-axis) vs. treatment effect on petition signing (y-axis). Each point represents one experimental condition; dashed line shows the linear fit with 95% confidence band. Study 1: 15 conditions (5 AI models

\times

3 conversation types). Study 2: 8 persuasion strategies. No significant correlation is observed in either study. (B) Average treatment effects on attitude and behavioural outcomes, split by prompt type: “Information (Issue)” vs. “Other.” In Study 1, “Other” refers to the Random prompt condition, in which models received one of 200 randomly generated persuasion prompts. In Study 2, “Other” refers to the pooled average of the seven non-Info-Issue strategies (Emotional Activation, Implementation Intentions, Identity Labeling, Commitment Escalation, Anticipated Regret, Info: Impact Efficacy, and Mega). Information prompts produce relatively larger effects on attitude outcomes (top rows) but not on behavioural outcomes (bottom rows). (C) Self-reported learning (x-axis; participants’ agreement that they learned new information from the conversation, measured on a 0–100 scale) vs. treatment effect on each outcome (y-axis). Each point represents one condition/strategy. Learning is strongly associated with attitude persuasion (top row) but not behavioural persuasion (bottom row), in both studies.

Figure 4: A combined persuasion strategy is most effective at driving behavioural change. (A) Average treatment effects (vs. control) of each Study 2 persuasion strategy on petition signing (red circles), petition support (grey triangles), and organization support (grey squares). Strategies are ordered by their effect on petition signing. Error bars represent 95% confidence intervals. “Mega (All Combined)” is a combined strategy that integrates all other strategies into a single prompt; the LLM adaptively selects which persuasion techniques to deploy during the conversation. (B) Heatmap of mean self-reported mechanism activation (0–100 scale) for each strategy (rows) across seven mechanism batteries (columns), each corresponding to the theoretical basis of one strategy. Cell values show mean scores; colour indicates relative activation within each mechanism (z-scored by column; red = above average, blue = below average). Black-outlined cells mark the diagonal—where a strategy is paired with its own targeted mechanism. The rightmost column (bold) shows each strategy’s average activation across all mechanisms. The Mega strategy shows elevated activation across most mechanisms, consistent with its superior persuasiveness on behavioural outcomes.

Petition	Organization	Study
Stop the risk of nuclear war	Campaign for Nuclear Disarmament	1
Stop banks fuelling the climate crisis	ActionAid	1
Seats in parliament should match how we vote	Electoral Reform Society	1 & 2
End suffering for mother pigs	Compassion in World Farming	1 & 2
Oppose Facial Recognition	Liberty	1 & 2
Lift the ban on work for people seeking asylum	Refugee Action	1 & 2

Table 1: Political issues used in Studies 1 and 2, with associated real UK petitions and sponsoring organizations.

Persuasion in Language Models

Persuasion in language models studies how they influence human and AI beliefs or decisions, and how to measure and limit undue influence.

See all 79 papers

Persuasion Propagation: Does Persuasion Change AI Agent Behavior?
Jan 31, 2026Hyejun Jeong, Amir Houmansadr, Shlomo Zilberstein +1Persuasion in Language Models AI Agent Evaluation
University of Massachusetts Amherst
AI systems out-persuade expert humans
Jun 15, 2026Kobi Hackenburg, Caroline Wagner, Luke Hewitt +5Human-AI Interaction Persuasion in Language Models
University of Oxford, Oxford, OX2 6GG, UK · 2UK AI Security Institute, London, UK · 3Stanford University, Stanford, CA 94305, USA +1
Perceived Political Bias in LLMs Reduces Persuasive Abilities
Feb 20, 2026Matthew DiGiuseppe, Joshua RobisonPersuasion in Language Models Political Bias in Language Models
Institute of Political Science Leiden University

Jan 31, 2026J/K move · Enter open · S save

Persuasion Propagation: Does Persuasion Change AI Agent Behavior?

Hyejun Jeong, Amir Houmansadr, Shlomo Zilberstein, Eugene Bagdasarian

University of Massachusetts Amherst

Persuasion in Language Models AI Agent Evaluation Software Engineering Agents

Open paper AskPDF arXiv

AI agents combine normal conversation with autonomous task execution, allowing earlier context to shape how later tasks are executed. Yet studying this possibility is challenging because agent behavior is noisy and costly to reproduce, and observed behavioral changes can be difficult to separate from generic context sensitivity. We first ask whether task-irrelevant persuasion changes how agents act, a phenomenon we call persuasion propagation, and whether persuasion susceptibility explains variation in these effects. We show that persuasive interaction leaves a measurable, task-dependent behavioral footprint on downstream agent execution. In web research, it increases execution time by 56% and domains visited by 16.3% relative to a baseline, while coding shows faster execution with larger revisions. However, agents that are more susceptible to persuasive conversational context do not reliably predict the magnitude of the downstream behavioral shifts. This motivates behavior-level evaluation of persuasion in AI agents.

Abstract

Persuasion in Language Models

Explore similar work

Persuasion Propagation: Does Persuasion Change AI Agent Behavior?

AI systems out-persuade expert humans

Perceived Political Bias in LLMs Reduces Persuasive Abilities