When a Conversation Becomes an Action
This chapter examines the most concrete question in the evidence on conversational AI persuasion: can a brief exchange change what people actually do? The answer comes from two large experiments in which UK adults spoke with language models about real political causes and were then offered opportunities to sign petitions or direct money. The point is not to produce one universal “AI persuasion rate,” but to understand what a randomized behavioral result means—and where its meaning stops.
The Experiment Compared Conversations, Not AI With Silence
The researchers recruited UK residents through Prolific and randomly assigned them to a conversation condition and a political issue. The control group did not simply sit out. They also interacted with an AI system, but discussed a neutral, non-political topic. Treatment participants instead had an interactive conversation intended to persuade them to support a petition. That comparison matters: the estimated effect is the additional influence of the political persuasive conversation over an AI conversation that occupied similar time and attention, not the effect of introducing a chatbot to someone who otherwise would have encountered nothing.
The first study retained 8,000 participants for its main analysis. Participants were assigned across eight issues, several frontier models, different study-bonus amounts, and three broad conversation types: neutral control, fixed information-based persuasion, or a randomly selected persuasion prompt. The second study retained 9,950 participants and used a new sample and a head-to-head set of eight theory-driven strategies, alongside control. Across both studies there were 17,950 study responses from 14,779 people. The samples were large enough to estimate differences across conditions, while the two designs asked somewhat different questions.
The interaction was intentionally ordinary in form: a multi-turn text exchange lasting two to ten turns, with an average of 4.9 turns, or about seven minutes. Before it began, participants reported their support for the petition and its sponsoring organization. Afterward, they reported those attitudes again and encountered behavioral choices. The sequence allowed the researchers to compare what people said with what they completed.
Random assignment is what gives the comparison its causal force. Because participants were assigned rather than allowed to choose the persuasive or neutral conversation, a difference in the average rate of signing can be attributed to the assigned conversation more plausibly than to pre-existing enthusiasm alone. It does not mean that every participant was affected in the same way, nor does it remove every threat created by attrition or the study setting. It means that, within the experiment, the treatment and control groups provide a disciplined counterfactual: what behavior looked like under a persuasive exchange can be compared with what it looked like under a similarly timed but politically neutral exchange.
The comparator is also substantively demanding. Both groups had to read and respond to an AI system, so the result is not simply a novelty effect from seeing a chatbot. The political conversation supplied a target and a reason to act; the neutral conversation supplied engagement without that target. This helps isolate the added contribution of persuasive content and direction. At the same time, the design cannot tell us how the same message would perform against no conversation, a human canvasser, a social-media post, or a person who encountered it accidentally. Each of those would be a different comparison.
The interaction was long enough to allow a back-and-forth but short enough to resemble a brief service encounter rather than a sustained relationship. That matters for interpretation. A seven-minute exchange can reveal whether an AI can cross an immediate behavioral threshold, but it cannot reveal whether the person returns to the cause later, remembers the arguments, discusses the decision with someone else, or changes course after the initial impulse fades. The experiment captures the first decision after exposure, which is valuable precisely because it is narrower than a claim about lasting persuasion.
The petition endpoint was designed to impose a real threshold. Participants who said they were willing could continue to a separate website, enter their first and last names and email address, agree to receive updates from the organization, and click the final submission button. They could abandon the process at any point. Only those who completed all steps counted as having signed. This is materially different from asking, “Would you sign this petition?” A willingness response records a stated intention; a completed signature records a choice that survived additional effort and a request for personal information.
The process also included choices about money and effort. In the first study, participants could donate some of a study bonus to the petition’s sponsoring organization, decide whether to keep that allocation or switch it to an alternative charity, and perform a repetitive clicking task that generated additional money for the organization. In the second, they could direct a smaller charity allocation to the assigned organization, an alternative, or neither. These outcomes extend the test beyond petition signing, but they are not interchangeable: donating money, selecting an organization, and clicking for a cause involve different costs and motivations.
Taken together, the measures create several behavioral thresholds rather than one scale called “action.” Signing asks for a public-facing commitment and personal details. A donation asks the participant to give up some available money. Keeping an allocation tests whether the conversation changes the preferred recipient when an alternative is made salient. Clicking tests willingness to spend effort for a small additional contribution. The more varied the outcomes, the less defensible it becomes to summarize the study with only its largest estimate. The useful conclusion is that the conversation affected multiple kinds of follow-through, while the size and reliability of each effect still need to be read separately.
The same logic applies to the two studies as a pair. The first study asks whether a persuasive conversation works across several models, issues, and conversation conditions. The second changes the prompt design to compare psychological routes to action. A repeated result across those designs is informative because it survives more than one operationalization. But it is not a clean contest between two model generations or a demonstration that the second study’s strongest strategy should replace every other approach. The studies were designed to answer related questions, not to establish a league table of assistants.
The Behavioral Threshold Was Real but Low-Cost
In the first study, the persuasive conversation increased petition signing by 12.8 percentage points relative to the neutral conversation. A percentage-point effect describes a difference in rates. If 30 people out of 100 signed under one condition, an increase of 12.8 percentage points would correspond to about 42.8 out of 100 under the other condition—not a claim that signing became “12.8 percent more likely” in every relative sense. The underlying control rate matters for translating the result into counts, so the percentage-point estimate is the safest way to report it here.
The conversation also increased the share of the study bonus donated to the sponsoring organization by 4.9 percentage points and made participants 11.3 percentage points more likely to keep their donation with that organization rather than switch to an alternative. The effort result was smaller: participants made 2.3 more clicks in the donation-generating task. That finding was partly driven by the petition about child poverty and was no longer statistically significant when that issue was removed. Treating all four outcomes as equally strong would therefore obscure the evidence rather than summarize it.
These are average treatment effects, not a description of a uniform response. An average can combine people who were moved, people who were already ready to act, and people who ignored or resisted the exchange. It also averages across issues and organizations whose audiences may have different levels of prior support. The researchers report broad consistency across the petition pairings, but consistency of direction does not mean identical effects for every cause or every participant. This is another reason to avoid turning a percentage-point difference into a prediction about any particular person.
The second study reproduced the main behavioral pattern with a different sample and a different strategy design. Participants in the persuasive conditions were 19.7 percentage points more likely to sign the petition and 6.1 percentage points more likely to allocate their charity earnings to the assigned organization rather than an alternative. The fact that the signing estimate was larger in the second study does not mean the second set of prompts is universally better. The studies differed in their strategy conditions, issue mix, models, sample timing, and other details. What transfers most confidently is the existence of a sizable short-run effect under both designs, not a fixed effect size that can be carried into any future setting.
The political issues were also not a single cause presented repeatedly. They included petitions about nuclear disarmament, democratic reform, animal welfare, climate, asylum, trees, and child poverty, among others. The authors report that the main behavioral pattern was broadly consistent across the eight petition–organization pairings, and the second study’s results were robust to demographic covariates, alternative model specifications, and leaving out one issue at a time. That breadth makes the finding harder to dismiss as an artifact of one unusually attractive topic. It still does not make the sample representative of every population or every political issue.
What makes these effects practically meaningful is not simply their statistical significance. A completed signature and a choice about money are observable actions that occur after the conversation and require some level of follow-through. The results show that the conversation can help move people across that threshold. They do not show that participants were transformed into committed activists, that their behavior lasted, or that they would have sought out the conversation without being recruited and paid.
What the Effect Size Means—and Does Not Mean
The control condition helps define the causal comparison, and the endpoint helps define the practical meaning. Together they also define the limits. Participants were not a random sample of the world. They were UK adults recruited through an online panel, compensated for their time, and placed in a study that guided them into an AI interaction. The authors report post-treatment attrition of 4.0 percent in the first study and 2.9 percent in the second, with no evidence of differential attrition across model or strategy conditions. That supports the internal comparison, but it cannot answer how often people in ordinary life will engage long enough to be influenced.
This is an important distinction between persuasiveness conditional on exposure and influence at population scale. The experiments show what happened among people who entered the study, completed the conversation, and reached the outcome measure. They do not measure the number of people who would encounter such a system, whether they would volunteer attention to it, or how many would stop reading after the first reply. A highly effective conversation that almost nobody accepts may have less social impact than a moderately effective one embedded in a widely used service. The first paper therefore treats exposure and engagement as a practical bottleneck, but the experiments do not directly estimate a population-level exposure rate.
The actions themselves also set a ceiling on what can be inferred. Petition signing and small donation decisions are consequential enough to distinguish behavior from opinion, but they are relatively low-cost. They do not stand in for vaccination, financial decisions, news consumption, dieting, or public-safety choices. Higher-stakes actions may create stronger reasons to verify information, consult other people, or resist a conversational recommendation; they may also create more motivation to comply. The direction cannot be assumed from these studies.
Nor should “real petition” be confused with “fully naturalistic.” Participants entered their information into a genuine petition flow, but they did so inside a paid survey after a deliberately designed conversation. The setting gives the behavioral result more substance than a hypothetical intention while retaining the advantages and distortions of an experiment. That is a strength-and-boundary combination, not a contradiction.
The most defensible reading is consequently layered. First, the randomized comparison provides evidence that the persuasive AI conversation caused a short-run increase in several political behaviors relative to a neutral AI conversation, including completion of a real petition. Second, replication across a second large sample suggests that the finding is not confined to one exact prompt design or one sample. Third, the varied effect sizes and the effort-task caveat show why behavior must be examined outcome by outcome. Finally, the paid context, low costs, UK recruitment, and unknown durability prevent the result from becoming a forecast of mass political conversion.
That last distinction is essential when discussing AI systems. Saying “the system changed behavior” is now empirically supportable in the narrow experimental sense. Saying “the system will change society at scale” requires several additional facts: who encounters it, who remains engaged, what they already believe, what action is available, how costly the action is, whether the effect persists, and whether the resulting behavior is beneficial or harmful. None of those questions can be answered by an attitude score alone, and not all can be answered by a petition-signing experiment.
The political studies therefore establish the first half of the central claim. Conversational AI can change actions, not just opinions, under conditions in which people are exposed, engaged, and offered a reachable behavioral step. The next question is more discriminating: when a conversation changes both reported support and behavior, are the same prompts, models, and conditions responsible for both? If not, the research habit of measuring opinion as a proxy for action becomes not merely incomplete, but potentially misleading.