Why Changing Minds Does Not Predict Moving People
This chapter examines a deceptively simple question about conversational AI persuasion: if one conversation changes reported support more than another, should we expect it to produce more signing, donating, or other action? The political experiments say no—not because attitudes are irrelevant, but because the evidence compares condition-level effects, and those effects do not track one another. The practical judgment is consequential: an attitude score cannot safely serve as a ranking proxy for behavioral influence.
A Null Correlation Is a Ranking Warning
The researchers compared the size of the attitude effect with the size of the behavioral effect across experimental conditions. In the first study, the conditions combined different models and conversation types. In the second, they were the eight persuasion strategies. For each condition, the researchers could estimate how much the AI conversation changed petition support and how much it changed petition signing. They then asked whether the conditions that performed well on one outcome also performed well on the other.
They did not find that relationship. In Study 1, the condition-level correlation between the effect on petition support and the effect on petition signing was r = 0.05, with p = 0.85. In Study 2, the corresponding correlation across the eight strategies was r = 0.18, with p = 0.68. The estimates provide no evidence that the strategies or conversation conditions that most changed reported support were also the ones that most changed signing.
This is best understood as a ranking warning. Imagine testing two prompts. Prompt A might produce a larger average increase in stated support, while Prompt B produces more completed signatures. If a study measures only support, it will recommend A for a behavioral goal it never actually tested. The observed correlations say that this is not a merely theoretical possibility in these experiments: the ordering of conditions on attitudes failed to predict their ordering on behavior.
The unit of analysis matters. These correlations compare estimated treatment effects across 15 conditions in the first study and eight strategies in the second. They do not say that an individual who changes an attitude will never act, that attitudes never influence behavior, or that a person’s personal attitude and personal action are unrelated. A person can become more supportive and sign a petition. The narrower conclusion is that the average advantage of one conversational condition over another on attitude did not tell the researchers which condition would have the larger average advantage on signing.
That distinction also separates two uses of a study. If the goal is to learn whether a conversation can shift stated support, an attitude outcome may be appropriate. If the goal is to choose a design for mobilizing a behavior, the study needs a behavioral endpoint. The same AI system can be effective for the first purpose and poorly characterized for the second.
Why can this happen even if attitudes sometimes matter for action? One possibility is that different conditions activate different people. An information-heavy exchange may move people who were undecided, while a more action-oriented exchange may move people who already support the cause. When those subgroups are averaged together, the condition with the largest overall attitude effect need not have the largest overall action effect. Another possibility is that the attitude measure captures a change that is real but insufficient: the person supports the petition more, yet still does not cross the time, effort, or commitment threshold required to sign. The experiment does not select one explanation by itself; it demonstrates that the endpoints cannot be safely substituted.
The small number of condition-level points also calls for proportion. Fifteen conditions in one study and eight in the other are enough to reveal a striking mismatch, especially because the pattern was replicated, but they are not a universal map of all prompts, models, or populations. A near-zero correlation is evidence against using these attitude estimates as a behavioral ranking tool in the tested comparisons. It is not a theorem that every future set of conditions will have a near-zero relationship. More direct behavioral replications, with more varied tasks and clearly specified starting populations, could strengthen or qualify the conclusion.
This is why the right response to the result is not to discard attitude measurement. It is to stop asking one measure to carry two jobs. An attitude outcome can diagnose whether a message changes evaluation. A behavioral outcome can test whether the change survives friction. Measuring both allows researchers to see whether a promising argument is failing because it does not change minds, because it changes minds without producing action, or because it mobilizes people whose views were already favorable. Each diagnosis leads to a different design choice.
Information Helps a Person Reconsider
The experiments give a plausible explanation for part of the dissociation. Factual information and perceived learning were closely connected to attitude change. In Study 1, the prompt that asked the model to provide information about the issue produced a +0.20 standardized effect on petition support, compared with +0.07 for the other prompts. Participants’ self-reported learning was also strongly associated with the size of the attitude effect: r = 0.87 for petition support across the experimental conditions.
That pattern fits an intuitive route to changing a stated view. A person encounters claims they consider new, incorporates them into an evaluation, and reports greater or lesser support afterward. The study does not need to assume that every factual claim was decisive; it shows that conditions perceived as more informative were the ones associated with larger attitude movement. The mechanism is not “information always persuades.” It is that, in these comparisons, perceived learning and the information-oriented prompt tracked the attitude outcome.
The same advantage did not appear for petition signing. In Study 1, the information prompt produced a signing effect of about +0.12, compared with +0.13 for the other prompts. Those figures are essentially comparable in the context of the reported analysis, even though the attitude effects were quite different. In Study 2, the Information–Issue strategy again ranked among the strongest for attitude change but was the weakest of the eight strategies for petition signing, at +16.2 percentage points. The authors therefore found no evidence that increasing information density created a corresponding behavioral advantage.
Study 2 also reproduced the learning contrast. Self-reported learning strongly tracked attitude change, with r = 0.85 for petition support, but not petition signing, where the reported association was r = −0.16 and nonsignificant. The same direction of evidence across two studies makes it difficult to explain the result as a single accidental comparison. It still does not prove that information has no role in any individual action. A fact may change someone’s plan, remove uncertainty, or make a cause seem more effective. The narrower finding is that the information treatment and perceived learning did not explain the differences in signing effects across these conditions.
The distinction between absence of a relative advantage and absence of any contribution is important. The experiment compares ways of persuading people; it does not remove information from people’s lives or trace every mental step that preceded a signature. “Information did not drive the observed behavioral ranking” is supported. “Information can never cause behavior” is not.
There is a second qualification around the learning measure. Participants reported that they had learned new information; the researchers did not use that measure as a complete objective test of what each person remembered or understood. Perceived learning may capture curiosity, engagement, confidence in the conversation, or a feeling that the exchange supplied reasons. It is still useful because it separated the conditions on the attitude outcome and not on the behavioral outcome, but it should not be treated as a direct measurement of knowledge acquisition.
The split suggests two different jobs for a persuasive exchange. Information can answer, “Why does this issue deserve support?” Action requires a further bridge: “What is the next step, and why should I take it now?” That bridge may be especially important when the desired behavior is public, effortful, or tied to a specific organization. A participant can accept the factual case and still postpone the signature, prefer to learn more, or feel that their individual contribution will not matter. Conversely, a participant may already accept the case and need only a concrete moment of commitment. The experiments do not prove which of these reasons dominated, but they show why more information is not the same as more follow-through.
This distinction also keeps the mechanism analysis from becoming a judgment about the value of information itself. Information may be normatively desirable even when it is not the best predictor of a petition signature. A research team evaluating an AI assistant for civic education could care about accurate understanding, while a campaign evaluating mobilization could care about completed actions. Those are different objectives. The appropriate outcome depends on the purpose, and the danger lies in treating the first objective as evidence that the second has been achieved.
To resolve the mechanism more fully, future experiments would need to measure the bridge rather than infer it from the endpoint. They could vary how much factual information is supplied while separately measuring comprehension, perceived efficacy, concrete planning, and completed action. They could also test whether the same information helps under different starting levels of support or different costs of acting. The current evidence does not perform all of those tests. It identifies the gap that such tests must close.
The structure in the diagram is deliberately asymmetric. The evidence gives the information-to-attitude path relatively strong support within the tested comparisons. It does not provide an equally complete account of the path to action. The prior-support pattern offers a useful clue, but the arrow from action to “mobilization” remains an interpretation that needs more direct testing.
From Persuasion to Mobilization
The authors’ exploratory analyses divide participants according to their prior support for the petition. Attitude effects were concentrated among people who were not already strong supporters. Behavioral effects, by contrast, were concentrated among prior supporters. This pattern is consistent with a distinction familiar in political behavior: persuasion can mean changing the view of someone who is unconvinced, while mobilization means helping someone who already agrees cross the threshold into action.
The distinction clarifies why the same conversation can produce a modest change in average support but a larger change in signing. A supporter may already be near the top of an attitude scale, leaving little room for a measured increase. Yet that person may not have signed because they lacked attention, a concrete prompt, a plan, or a reason to act now. A non-supporter may have more room to move on the scale but still remain below the threshold for public or effortful action. These are not competing descriptions of the same response; they are different starting conditions.
The experiment does not identify which missing ingredient moved the prior supporters. It did not isolate a single motivational process and prove that it caused the action effect. The authors present the pattern as an exploratory result and interpret it as consistent with persuasion–mobilization theory. That language matters. The data support taking starting preference seriously; they do not justify declaring that every behavioral effect of conversational AI is mobilization, or that information never matters for supporters.
The broader mechanism may involve a different bottleneck from belief revision. A person who is unconvinced needs reasons to update an evaluation. A person who already agrees may need the conversation to make an action feel immediate, feasible, identity-consistent, emotionally salient, or worthwhile. Those possibilities motivate the action-oriented strategy comparison examined next, but they are not interchangeable explanations and cannot be selected merely because they sound psychologically plausible.
This is where attitude-only evaluation can systematically mislead. Suppose an information-heavy prompt produces the largest change among non-supporters. A researcher who wants to mobilize existing supporters might choose it, even though the prompt’s behavioral effect is no better than alternatives and is weakest in the second study’s strategy comparison. The error is not that the attitude result is false. It is that the result answers a different question from the one the researcher needs answered.
A better evaluation keeps at least three pieces of information together: the person’s starting position, the outcome that matters, and the condition assigned to them. It should ask whether the interaction changed minds, activated agreement, or induced some other form of follow-through. It should also state whether the evidence comes from individual responses or from comparisons among conditions. Without those distinctions, a precise-looking attitude estimate can create false confidence about behavior.
The political evidence therefore supports a restrained mechanism story. Information and perceived learning are associated with changing reported attitudes. They do not explain the observed differences in petition signing. The people whose views and behaviors moved may not have been the same people, and the behavioral route may depend less on one superior set of facts than on how the interaction meets someone at a particular starting point.
That leaves an open, testable question rather than a mystery: if no single information advantage explains action, do several action-oriented mechanisms perform differently, or do they produce a broad band of similar behavioral effects? The answer determines whether the practical problem is choosing the right rhetorical technique or reaching and engaging people who are already close to acting.