The Outcome Persuasion Research Missed
Conversational AI persuasion is the use of a natural-language exchange with an AI system to influence what a person believes, chooses, or does. The central question here is narrower and more consequential than whether an assistant can produce a convincing argument: what does the conversation make the person do? This series examines the evidence for political action and everyday deference to AI, and gives the reader a way to distinguish a changed opinion from a changed action and from reliance on advice.
That distinction matters because people do not act as carbon copies of their stated views. Someone may become more supportive of a petition without signing it. Someone else may already agree with a cause and need only a prompt, a plan, or a moment of encouragement to act. A third person may have no settled opinion at all but still follow an AI assistant when it disagrees with an expert. These are all forms of influence, but they are not the same event and should not be measured as though they were.
Persuasion Is Not the Same as Agreement
Much of the recent discussion about AI persuasion has been organized around attitude change. Researchers ask participants how much they support a policy, a petition, or an organization before and after an interaction. That is a useful question: it can show whether a person’s expressed position moved. But it is only one point on the path from communication to consequence.
An attitude is a reported evaluation. An action is an observable choice that may cost time, money, effort, attention, or social commitment. The distance between the two can be small in one situation and substantial in another. Signing a real petition requires more than selecting “support” on a scale; it requires continuing through a separate process, entering identifying information, and completing the final action. Donating or performing an effortful task adds another kind of cost. The action measure therefore tests whether influence survives contact with friction.
The distinction is not merely a matter of stricter measurement. It changes the story about how influence works. Information may help someone revise a belief without providing whatever motivation or opportunity is needed to act. Conversely, an interaction may activate a person who already agrees, even if their measured attitude barely changes. If researchers assess only opinions, they can miss the second pathway—and can rank two AI systems in the wrong order for the outcome that matters.
The same issue appears outside politics. In everyday decision support, a user may not begin with a strong opinion to change. They may be trying to understand a new problem and decide whether to follow an AI recommendation. If the assistant conflicts with an expert, the relevant question is not simply whether the user reports greater confidence in the assistant. It is whether they revise their decision in the assistant’s direction. That is a behavioral question about reliance.
The diagram’s separation is the basic discipline for reading the evidence. An AI conversation is the common input, but the endpoints branch. A result at one endpoint cannot automatically be used to fill in the others, and a controlled observation cannot automatically become a claim about durable, large-scale influence.
Three Outcomes, One Subject
The three outcomes belong to the same subject because they concern the influence of conversational AI on human judgment and conduct. They remain distinct because each captures a different relationship between the person and the system.
Attitude change asks whether the person’s stated evaluation moved. In the political experiments considered in this series, that includes support for a petition or its sponsoring organization. It is closest to the ordinary meaning of “changing someone’s mind,” although even here the measurement is a report collected at a particular moment.
Observable action asks whether the person completed a target behavior. The political evidence includes signing real UK petitions and directing money toward a sponsoring organization. These are not the highest-stakes actions imaginable, but they are more consequential than saying that one approves. They reveal whether a conversation can move someone through a behavioral threshold.
Reliance asks whether the person defers to AI output when another source of judgment is available. In the everyday experiment, participants first saw an expert recommendation about a fictional flood-barrier budget and then read an AI assistant’s counterarguments. A later decision to move away from the expert position was treated as reliance on the AI. This is not identical to political persuasion because the participants were selected for a setting with little prior knowledge or established opinion.
Reliance also needs careful language. Moving toward an AI recommendation is not automatically harmful. In the fictional scenario, the competing positions were opinions rather than claims with an independently verifiable right answer. The study can show deference against the expert position; it cannot establish that every such change produced a bad decision. “Overreliance” becomes a stronger claim only when the decision is wrong, the error has meaningful consequences, or the user’s confidence and behavior are badly calibrated to the system’s actual performance.
The distinction can be summarized simply: attitude measures what a person says, action measures what they complete, and reliance measures whom they follow when judgments conflict. A single conversation may affect all three, two of them, or none. The research question is therefore not “Is the AI persuasive?” in the abstract. It is “Persuasive for which outcome, for whom, under what conditions, and with what cost?”
Two experiments make the contrast especially useful. The first places conversational AI in a political setting where the target can be a petition signature or a donation-related choice. The second moves to a low-knowledge decision in which the AI challenges an expert recommendation. The settings should not be collapsed into one effect size: they differ in sample size, task, manipulation, and stakes. Their value lies in showing that behavioral influence has more than one form.
What the Evidence Can and Cannot Tell Us
The strongest immediate conclusion is also the most limited: brief conversational interactions can produce measurable behavior change under controlled conditions. In the political studies, participants were recruited from the United Kingdom, paid to take part, and assigned to either a persuasive conversation about a real petition or a neutral conversation. The exchange lasted only a few minutes on average. Yet the researchers observed differences in signing and donation-related choices, not just in post-conversation support.
That finding is important without being a universal risk estimate. The participants’ attention was secured inside a paid survey. The actions were consequential enough to observe, but generally low-cost. A person who signs a petition in an experiment has crossed a real behavioral threshold, but that does not tell us whether the same conversation would change a vaccination decision, a major financial choice, news consumption, or a durable political commitment. Those domains bring different stakes, incentives, opportunities for verification, and reasons to resist or accept advice.
The everyday experiment has a different limitation. Its fictional Harborview flood-barrier scenario helps remove prior political loyalties and factual disputes, making it easier to study presentation and deference. But the absence of real consequences also makes the behavior easier to interpret and less representative of a high-stakes decision. A participant can move away from the expert position without risking a public service, a household budget, or someone’s health.
These constraints do not cancel the findings. They tell us what kind of findings they are. The political studies provide direct evidence that conversational AI can mobilize some short-run actions, including completion of a real petition. The everyday study provides evidence that people can defer to an AI assistant in a decision-support task even when the tested conversational features do not strongly change that behavior. Neither establishes that the same magnitude will persist outside the experimental setting.
A useful way to continue is to keep four questions separate. First, did an outcome change? Second, which outcome was measured? Third, what mechanism is supported by the comparison? Fourth, what would have to be true for the result to matter at scale? The first three belong to the experiment. The fourth requires evidence about exposure, engagement, durability, user starting points, objective performance, and stakes.
This also sets the series’ practical judgment. When an AI system is described as persuasive, do not stop at a chart of attitude scores or a survey of trust. Ask whether users act, whether they verify, whether they follow the system when they should not, and whether the result differs between people who are unconvinced and people who already agree. Those questions do not assume that AI influence is always harmful. They make the claim precise enough to evaluate.
The next test is therefore concrete: can a brief conversation with a language model change a real political action rather than only a reported opinion? The answer, in the evidence examined here, is yes under the studied conditions. Understanding why—and why that answer does not follow from attitude change alone—requires looking directly at the political experiments and their behavioral measures.