Conversations With AI Moved People's Actions and Their Opinions Separately
A conversation with an AI model ends, and the screen asks for something more than an opinion. It asks for a signature on a real petition, or for money to go to a charity. A study published Oct. 16, 2026, in the Proceedings of the National Academy of Sciences was built around that moment, and most research on AI persuasion never reaches it.
The work comes from Kobi Hackenburg at the Oxford Internet Institute and the UK AI Security Institute, a UK government body. Ben M. Tappin at the London School of Economics is the senior author, and three colleagues sit at the security institute and at Stanford University. They report two large preregistered experiments covering 17,950 responses from 14,779 people, in which conversational AI models tried to move participants on both kinds of outcome: what they said they believed, and what they did. The behavioral effects were sizable. Signing a real petition rose by 19.7 percentage points.
The petition was offered on the spot and cost nothing to sign. It came directly after a conversation designed to persuade, inside online panels of people who had signed up to take surveys. It is a real action rather than a stated intention, which is the point of the design. It is also the cheapest kind of real action: one click, taken immediately.
The result the authors put the most weight on is a negative one. Across the conditions they tested, they observed no evidence of a correlation between a model's effect on attitudes and its effect on behavior. That is an absence of evidence in these two experiments, not a demonstration that no link exists. Both kinds of outcome moved. They did not move together. For anyone who evaluates these systems, that is the practical sting: a score built on how much a model shifts opinions carries no promise about how much it shifts conduct. The authors state the implication as a hedge. Earlier findings that rest on attitudes may generalize poorly to behavior, and so risk mischaracterizing what AI persuasion does in the world.
Understanding why that matters requires a step back. The modern evidence that AI can persuade is built mostly on attitudes. The best known example came from Thomas H. Costello, Gordon Pennycook and David G. Rand, who reported in Science in September 2024 that personalized dialogues with GPT-4 Turbo reduced conspiracy believers' belief in their chosen theory by about 20%. The effect was still present two months later. That paper also reported a shift in conspiracy-related behavioral intentions, which is a statement about what people expected of themselves.
The gap the new paper measures is an old question in psychology. In 2006, Thomas L. Webb and Paschal Sheeran gave a meta-analysis in Psychological Bulletin a title that asked whether changing behavioral intentions engenders behavior change. What the PNAS experiments add is that the gap turns up inside a single study, with both kinds of outcome measured on the same people at the same time.
One more piece of the pattern points the same way. Hackenburg and colleagues replicated an established finding that information provision drove the effects on attitudes, and found no such evidence for the behavioral outcomes. In a test of eight behavioral persuasion strategies, the one built on information provision produced the smallest behavioral effect, though the differences among the eight were small. The approach that does most to change what people say appears to do least to change what they do.
The authors say it remained unclear whether AI could persuade people to take consequential real-world actions. For one immediate, costless action, they now have an answer, and it is a large one. The more durable result is the question it hands to everyone else working on this: when a study reports that an AI model persuaded somebody, which outcome did it measure?