How much can persuasive AI chatbots shift political opinion?

Writing in Nature, Chiara Vargiu and Alessandro Nai report that two companion studies — Lin et al. in Nature and Hackenburg et al. in Science — put a number on how much conversational AI can move political opinion, and the number is large by the standards of campaign politics. Political science's "realistic estimate is that political advertising shifts vote choices by less than one percentage point," the piece notes, while "Hackenburg and colleagues analysed a maximally persuasive configuration that 'pulls all the levers at once'... they estimated that it could shift opinions by 16 points on average and 26 points among participants who initially disagreed with a given issue." Separately, Lin and colleagues had more than 2,300 US participants exchange three rounds of messages with a model instructed to argue respectfully and factually for Harris or Trump in the 2024 election, with similar designs later run on a Massachusetts psychedelics ballot initiative and the 2025 national elections in Canada and Poland.

The mechanism the piece credits is tuning plus delivery style, not scale or audience size: "models that were fine-tuned specifically for persuasion were up to 50% more effective than their untuned counterparts," and the most information-dense, evidence-citing models were also the most convincing — but "the same settings also reduced factual accuracy," producing a persuasion-accuracy tradeoff the authors call systematic. The closing observation is that the persuasion worked because the AI was civil and fact-citing, "qualities that are often seen as lacking in human political debate" — civility and accuracy, not personalization or raw exposure, did the convincing.

This directly extends Where does AI's persuasive power actually come from?, which is the Hackenburg et al. study itself at much larger N (76,977, 19 models, 707 issues): that note establishes the post-training/prompting mechanism and the accuracy tradeoff, and this source adds the real-world comparison point — a <1-point campaign-ad baseline against a 16-to-26-point AI-chat shift — plus the companion Lin et al. finding that the same effect holds with real 2024 candidates and replicates across a ballot initiative and two other countries' elections. It also sharpens why disclosure-based safeguards like Does telling people they are talking to AI change how persuaded they become? matter: if civility and factuality are themselves the persuasive lever, policy aimed only at labeling "this is AI" won't touch the mechanism doing the work.

What the News & Views piece does not establish is how this plays out outside the lab: it explicitly cautions that "these findings come from controlled online experiments" where exposure is involuntary and uncontested, whereas real political environments are voluntary, competitive, and full of counter-messaging that could dilute the effect. The implication the authors draw, at the strength the evidence allows, is a governance one — regulating AI persuasion will need to reach training objectives and disclosure rules, not just output content, "who controls this capacity" being the question they leave open rather than answer.