AI political persuasion changed what people did, not merely what they said they believed. In two preregistered randomised experiments involving 14,779 UK adults, brief conversations with leading AI models made participants more likely to sign real political petitions and direct money towards the organisations behind them.
The result matters because most research on AI persuasion has measured attitudes. A person might report feeling more supportive after a chatbot conversation without taking any action. The new research paper, posted as a preprint in April 2026, tested whether persuasion survived the extra step between agreeing with an argument and acting on it. In both experiments, it did.
AI Political Persuasion Reached Beyond Opinions
The researchers recruited UK adults through the Prolific survey platform and randomly assigned them to one of eight real petitions. The causes included electoral reform, nuclear disarmament, opposition to facial recognition, animal welfare, child poverty and environmental protection. Participants then had a text conversation with an AI system for an average of 4.9 turns, lasting about seven minutes.
In the control condition, the AI discussed a neutral, non-political subject. In the treatment conditions, models including GPT-4.1, Claude, Gemini and Grok were instructed to persuade the participant to support the assigned cause. Signing required more than clicking a survey answer. Participants had to enter their name and email address on a separate petition page and agree to receive updates.
The first experiment had an analytical sample of 8,000 people. Compared with the control group, those who received a persuasive AI conversation were 12.8 percentage points more likely to sign. They also allocated more of their study bonus to the sponsoring organisation, completed slightly more of a repetitive task that generated money, and were 11.3 points more likely to keep their allocation with that organisation rather than switch it elsewhere.
The second experiment included 9,950 responses and compared eight different persuasion strategies. Participants who received persuasive conversations were 19.7 percentage points more likely to sign and 6.1 points more likely to direct their charity allocation to the petition sponsor rather than an alternative. The underlying figures, analysis scripts and anonymised outcome data are available in the researchers’ public replication repository.
Changing a Mind Was Not the Same as Producing Action
The most revealing finding was not simply that the conversations worked. It was that their effects on attitudes and behaviour were not correlated. Across the experimental conditions, the approaches that changed stated support most effectively were not reliably the approaches that produced the most signatures.
Information-heavy prompts performed well when the outcome was a change in opinion. Participants who said they had learnt more also tended to show larger attitude shifts. That relationship disappeared when the researchers examined petition signing. Providing facts could move what people said they thought, but it did not explain why they took the extra step of acting.
In the second study, every tested behavioural strategy increased petition signing relative to the neutral control. The strongest combined emotional, informational and commitment-based techniques into one adaptable prompt. It produced a 23.7-point increase, while the information-only strategy produced the smallest increase at 16.2 points. The gap between strategies was much narrower for behaviour than it was for attitudes.
The authors’ exploratory analysis offers one possible explanation. Attitude changes were concentrated among people who initially disagreed, while behavioural effects were stronger among people already sympathetic to a cause. In plain terms, changing someone’s mind and mobilising someone who broadly agrees may be two different persuasion problems.
Why Political Campaigns May Look at This Differently
Traditional political persuasion has an exposure problem. A campaign must find a voter, hold their attention and deliver a message relevant enough to their concerns. Conversational systems could personalise that process in real time, answer objections and adjust their language during the exchange. The experiment shows that this potential deserves scrutiny at the level of behaviour, not just message quality.
That does not mean an automated campaign could reproduce the result at national scale. The participants were being paid to complete a study and therefore gave the chatbot attention that unsolicited campaign material rarely receives. The paper itself identifies engagement as a likely bottleneck. A persuasive system achieves little if people refuse to begin or continue the conversation.
It also matters that the tested actions were relatively low cost. Signing a petition or directing a small study-linked allocation is meaningful, but neither is equivalent to changing a vote, joining a protest, making a large donation or maintaining a new behaviour. The research does not establish those larger effects.
Even within that boundary, the result expands the political AI debate. Much public attention has focused on synthetic voices and political deepfakes, where the deception is visible in the content. Conversational persuasion creates a different challenge because the material may be accurate, the exchange may feel voluntary and the influence may emerge from adaptation rather than a single false claim.
What This Means for You
A chatbot asking for political action should be treated as a persuader, not simply an information service. Before signing, donating or sharing, it is reasonable to ask who chose the objective, who benefits from the action and whether the system is adapting its approach using information gathered during the conversation.
A pause between conversation and action can also restore separation between the argument and the decision. Checking the petition sponsor independently, opening the original evidence in a new window and considering the strongest opposing case are simple ways to avoid allowing one tailored exchange to define the entire choice.
This is particularly relevant because conversational systems can combine factual information with affirmation, urgency and emotional framing. LiveAIWire’s examination of measurable dark patterns in AI responses showed why apparently helpful design choices can deserve scrutiny when they serve an engagement objective. The new study adds evidence that conversational influence can reach a completed action.
Agreement from a chatbot should not be confused with independent validation either. Research covered in LiveAIWire’s analysis of AI sycophancy found that models can reward a user’s existing position rather than challenge it. A politically persuasive system could use that tendency to mobilise people who already lean towards its assigned cause.
What the Experiments Cannot Yet Establish
The paper is a preprint and has not yet passed a publicly inspectable journal peer-review process. Its large sample and preregistered design strengthen the evidence, but they do not remove the need for replication. The first study’s preregistration confirms that petition signing and money-related actions were specified before data collection, while some later analyses were exploratory.
The sample was limited to English-speaking adults living in the UK. Political cultures, petition habits and trust in AI differ across countries, so the size of the effect should not be exported automatically. The participants also knew they were in a study and were debriefed afterwards, conditions that differ from a covert or commercially deployed persuasion system.
The researchers tested several frontier models, but the study was not designed to produce a durable league table. Models change quickly, and a difference observed under one set of system prompts may not survive a later update. The more important result is structural: several current systems influenced behaviour across multiple causes and strategies.
There is also no evidence here that the petitions themselves were harmful or that participants regretted signing. Persuasion is not automatically manipulation. A charity, public body or civic organisation may use conversational tools to help people understand a legitimate cause. The governance question is whether the objective, sponsor and persuasive role are disclosed, and whether people retain meaningful control over the decision.
Behaviour May Need to Become the New Safety Test
AI evaluations often ask whether a model can generate convincing arguments or shift a rating on a scale. Those measures are easier and cheaper than observing real action, but this study suggests they may answer the wrong question. A technique that wins an argument is not necessarily the one that gets someone to sign, donate or participate.
For researchers and regulators, that creates a forward-looking measurement challenge. Safety testing may need to include behavioural outcomes, clear records of the objective assigned to the model and tests of how the system responds to hesitation. Auditors may also need to distinguish persuasion of the unconvinced from mobilisation of people who already agree.
For users, the immediate lesson is simpler. A short AI conversation can be more than an exchange of information. Under controlled conditions, it changed political actions by double-digit percentage-point margins. The safest assumption is that a chatbot with an assigned objective may be trying to move both your opinion and your next click, even when those two effects arise through different routes.
LiveAIWire’s earlier reporting on synthetic empathy showed how emotional responsiveness can serve either the user’s interests or a platform’s engagement goals. Political persuasion sharpens that distinction. The question is no longer only whether an AI sounds convincing. It is whether the person on the other side understands why the conversation is happening before it changes what they do.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.