Problem Addressed
AI persuasion has become an increasingly important area of research in political campaigning. A growing body of academic work suggests that large language models and conversational AI may be capable of influencing opinions, increasing engagement, and changing behaviour under certain conditions. At the same time, modern elections are characterised by high levels of voter volatility, weakening party attachment, and fragmented media environments, making persuasion more valuable than ever before.
Despite the growing interest in AI-assisted persuasion, most existing experiments have taken place in controlled academic environments rather than real-world campaigns. There remains relatively little evidence about how these systems function when deployed in live electoral settings, where campaigns must navigate practical constraints, ethical concerns, transparency requirements, and the unpredictability of genuine voter interaction.
Campaign Lab wanted to explore whether conversational AI could play a constructive role in democratic campaigning, and what practical lessons could be learned from deploying these tools in the field.
Approach & Implementation
During the May 2026 local elections, we partnered with a political candidate to conduct a small-scale field experiment using an AI-assisted persuasion chatbot.
We built a conversational persuasion model trained on publicly available information, campaign materials, local political context, and issue positions associated with the incumbent candidate. The system was designed not simply to broadcast campaign messaging, but to engage voters conversationally through Facebook Messenger in a more interactive and personalised way.
The chatbot was connected directly to the candidate’s Facebook Messenger page and operated as part of the live campaign. Particular attention was paid to transparency and consent throughout the experiment. Users were informed they were interacting with an AI system, and the campaign tested different approaches to disclosure, conversational tone, and escalation to human follow-up.
One particularly important element of the trial was experimentation with the use of a hybrid engagement model. In some cases rather than attempting to fully automate persuasion, the AI system engaged in conversations designed to answer basic questions and identify engaged users, after which the human candidate could step into conversations directly. This allowed us to explore whether AI could reduce friction and expand campaign capacity without entirely replacing authentic human interaction.
Evidence & Evaluation
Over the course of the campaign, the system engaged in approximately 75 voter conversations. While the scale of the experiment was relatively small and does not allow for causal claims, the results provide encouraging early indications that AI-assisted engagement may have practical value in campaign settings.
We observed suggestive persuasive effects in some interactions, alongside several broader operational findings. In particular, the experiment highlighted the importance of transparency, conversational style, and expectation-setting in reducing user resistance to AI systems.
The hybrid human-AI model also showed promise. In a number of cases, the chatbot successfully initiated conversations that later developed into more substantive exchanges with the candidate directly. This suggests that AI systems may be particularly effective as tools for scaling initial voter contact and engagement, while still preserving meaningful human interaction at key moments.
The experiment remains exploratory and observational in nature, and there are clear limitations to the data. However, the findings provide reasons for cautious optimism and suggest that further experimentation in real-world campaign environments is both possible and valuable.
Below you can read the full paper detailing the methodology, ethical considerations, implementation process, and findings from the project.