Measure the Action, Calibrate the Influence
The evidence on conversational AI persuasion supports a clear answer to the question that opened this series, but not a simple one. AI conversations can change actions, not just opinions. They can increase signing and donation-related behavior in political experiments, and they can move people toward an AI recommendation when it conflicts with an expert. Yet those behavioral effects do not follow reliably from the same conditions that change attitudes, and a persuasive interface is not the same thing as a calibrated or harmful one. The final judgment is therefore outcome-first: evaluate what changed, who was exposed, what the person was already inclined to do, and whether the resulting behavior was accurate, consequential, and durable.
One Influence Model, Three Endpoints
A useful way to connect the evidence is to treat conversational AI influence as a set of paths that share an interaction but diverge at the outcome. The first path is attitude change: a person reports more or less support, confidence, or agreement after the exchange. The second is observable action: the person signs, donates, keeps an allocation, or spends effort. The third is reliance: the person follows the AI when another judgment, such as an expert recommendation, is available.
These endpoints are related without being interchangeable. A person can learn something and revise a view without acting. A person can already agree and act after a prompt that adds little to their measured attitude. A person can distrust an AI in the abstract and still follow its advice in a specific decision. The experiments do not treat these as contradictions to be resolved by choosing one “true” measure. They show that each outcome has its own threshold and its own evidence requirements.
The political experiments establish the action path most directly. Participants in a persuasive AI conversation were more likely than participants in a neutral AI conversation to complete a real petition-signing flow. In the first study, signing rose by 12.8 percentage points; in the second, by 19.7 percentage points. Donation-related choices moved as well. The result is stronger than a shift in willingness because participants could abandon the separate signing process, and only completed submissions counted. It is still a short-run effect in a paid survey, not evidence of durable political conversion.
The same experiments establish a separation between attitude and action. Across conditions, the effect on petition support did not predict the effect on petition signing: the reported correlations were r = 0.05 in the first study and r = 0.18 in the second. Those are correlations among condition-level treatment estimates, not individual-level laws. Their value is as a warning about ranking. A prompt that looks best on an attitude chart may not be the prompt that produces the most action.
The everyday decision experiment adds reliance as a distinct behavioral path. Participants in the fictional Harborview scenario first saw an expert recommendation, then encountered an AI assistant’s counterarguments, and finally reconsidered both confidence and the approval decision. The confidence measure captured an attitude-like shift away from the expert position. Reliance captured movement in the decision toward the AI’s anti-expert position. A person could become less confident without changing the decision, or change the decision without that change being evidence of an objectively wrong choice.
The shared model is therefore not “AI says something persuasive, then people comply.” It is closer to this: an interaction presents arguments and social cues to a person with a particular starting point; the person processes or accepts some of them; then an outcome-specific threshold determines whether a reported evaluation, an action, or a reliance decision changes. The thresholds may be related, but the experiments show that they can be crossed independently.
The differences between the two settings matter just as much as their common structure. Political petition signing has a real external target and a reachable but low-cost action. Harborview has a judgment conflict but no real consequence and no objective ground truth. The first study can say more about completed action; the second can say more about deference under controlled disagreement. Neither supplies a universal measurement for the other.
What the Mechanisms Actually Support
The evidence supports a partial mechanism model rather than one master explanation. Information and perceived learning tracked attitude change. In the political experiments, an information-about-the-issue prompt produced a much larger attitude effect than other prompts—about +0.20 standardized units versus +0.07 in one comparison—and self-reported learning was strongly associated with attitude effects. But the signing effects were similar for information and other prompts, about +0.12 versus +0.13, and the Information–Issue strategy was the least effective of the eight tested strategies for signing in the second study.
The correct conclusion is not that information is unimportant. It is that information did not supply the behavioral advantage observed for attitudes. A message can improve a person’s understanding or evaluation without supplying a plan, urgency, identity connection, perceived efficacy, or immediate opportunity to act. The prior-support analysis points in a similar direction: attitude effects were concentrated among non-supporters, while behavioral effects were concentrated among prior supporters.