Overview
This series asks what conversational AI makes people do, not only what they say. It separates attitude change, completed action, and reliance; examines real petition signing and donation-related choices, action-oriented strategies, and deference to AI counterarguments; and explains why information, warmth, contextualization, trust, and self-reported literacy do not form one simple causal story. The evidence supports short-run influence among exposed, engaged participants, while leaving reach, durability, decision quality, and high-stakes generalization unresolved. It ends with an outcome-first way to test influence by endpoint, starting point, exposure, quality, stakes, and durability.
Learning outcomes
- Persuasive conversational AI increased real petition signing and donation-related choices in controlled UK experiments.
- Attitude effects did not predict condition-level signing effects, so opinion scores cannot rank behavioral influence.
- Several action-oriented strategies worked; the combined strategy led by a modest margin, not a universal best.
- In Harborview, AI shifted decisions against an expert for some users, but the fictional task cannot establish harmful error.
- Warmth and explicit contextualization interacted on persuasion but did not reliably change reliance in the tested task.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Can conversational AI change actions, not just opinions?
Yes, under the studied conditions. Brief persuasive conversations increased real petition signing and donation-related choices in controlled UK experiments, but the effects were short-run and conditional on supplied exposure and engagement.
Why doesn't attitude change reliably predict action?
Attitude change can fail to produce action because a revised evaluation may not cross the time, effort, or commitment threshold for behavior. In these studies, condition-level attitude effects did not predict signing effects, so opinion scores were not a safe behavioral ranking proxy.
Do warmth and personalization increase AI reliance?
Not reliably in the Harborview task. Explicit warmth and contextualization affected the confidence-shift outcome conditionally, but neither reliably changed reliance after the participant's initial decision was accounted for.
Does following AI advice prove harmful overreliance?
No. Following an AI recommendation shows reliance; overreliance requires evidence that deference was inappropriate, poorly calibrated, or led to an error. Harborview had fictional stakes and no objective ground truth, so it showed deference without proving harm.
How should conversational AI influence be evaluated?
Evaluate the endpoint directly—attitude, completed action, or reliance—then record the starting point, exposure and engagement, mechanism, action quality, stakes, verification, and durability. A result should not be generalized beyond the population and task that produced it.