Can supporters of populist radical right parties be persuaded to reconsider their views - and if so, how? It is among the more vexed questions in the study of contemporary democracy, and one to which the standard answer - better facts, fact-checking, firmer gatekeeping and so on - have proven largely inadequate, when they have not actively backfired. This paper pursues a different tack on two fronts at once. First, we put a genuinely new tool to work: a conversational AI that talks with voters one-on-one, at a scale and level of personalisation no canvasser or campaign could match. Second, and more fundamentally, we ask what that tool should say - and our wager is that the answer lies in voters' own moral values rather than in the facts we wish they would accept. We present a preregistered three-arm randomised experiment (N = 2,400) testing whether AI conversations can soften populist attitudes and support for Pauline Hanson's One Nation - now commanding between 20 and 30 per cent of primary vote intention in Australia. Participants receive either no intervention, a generic AI conversation, or one in which the model reframes populist appeals through the moral foundations a given participant already holds - whether care, loyalty, or purity. The generic arm is doing real work: by holding the AI constant and varying only the moral tailoring, we isolate whether value alignment itself is what moves people, rather than mere novelty or engagement. We further test whether effects survive at one- and four-week follow-up, and whether a brief "booster" conversation shores up their durability.
