Personalised persuasion at scale

The defensible concern is not the viral deepfakeDeepfakeA fake video, audio clip or image made with AI, showing a real person saying or doing something that never happened.For exampleA voice message from your boss asking for an urgent transfer, which your boss never recorded. but the conversation: personalised, sustained and optimised to persuade. The effect exists and has been measured. Salvi and co-authors, in Nature Human Behaviour, with a 2×2×3 design and N = 900, find that in the debate pairs where AI and humans were not equally persuasive, GPT-4 with personalisation was more persuasive 64.4% of the time ^[930]On the conversational persuasiveness of GPT-4Salvi, Francesco; Horta Ribeiro, Manoel; Gallotti, Riccardo et al. · 2025 · paperView source ↗Accessed on 9 September 2026. The mechanism works in both directions: Costello, Pennycook and Rand reduced belief in conspiracy theories by around 20% with three rounds of dialogue, with the effect persisting for two months ^[284]Durably reducing conspiracy beliefs through dialogues with AICostello, Thomas H.; Pennycook, Gordon; Rand, David G. · 2024 · paperView source ↗verified through CrossrefAccessed on 9 September 2026. The capacity to persuade is neutral with respect to content. Detection does not protect either: in the unauthorised University of Zurich experiment on a subreddit of nearly four million users, nobody ever suggested that the comments might be AI-generated ^[1146]Can AI Change Your View? Evidence from a Large-Scale Online Field Experiment (extended abstract)Investigadores no identificados de la Universidad de Zúrich · 2025 · preprintView source ↗Accessed on 9 September 2026.

What this does not demonstrate. This is the worst-supported part of the epistemic axis, and what fails is precisely the word “personalised”. Hackenburg and Margetts, in PNAS, with n = 8,587 and live microtargetingMicrotargetingSending each person, or very small groups, a different political message, tailored to what is known about their tastes, fears and data.For exampleYou and your neighbour seeing ads from the same candidate with opposite promises, each the one you wanted to hear., find that the persuasive impact of microtargeted messages was not statistically different from that of untargeted ones (4.83 against 6.20 percentage points, P = 0.226) ^[492]Evaluating the persuasive influence of political microtargeting with large language modelsHackenburg, Kobi; Margetts, Helen · 2024 · paperView source ↗verified through CrossrefAccessed on 9 September 2026. And the study that puts the field in order —N = 76,977, 19 models, 707 topics, 466,769 verified claims— concludes that persuasion comes from post-training and promptingPromptThe instruction or question written to an AI. Prompting is drafting it to get a particular answer.For exampleWhat you type into the chat: “Summarise this contract in five points” is a prompt., with smaller effects from personalisation and model size, and adds the finding that defines the shape of the risk: where those methods increased persuasion, they systematically lowered factual accuracy ^[493]The levers of political persuasion with conversational artificial intelligenceHackenburg, Kobi; Tappin, Ben M.; Hewitt, Luke et al. · 2025 · paperView source ↗verified through CrossrefAccessed on 9 September 2026. A note on method: on 3 September 2026 an author correction to Salvi’s article was published whose content could not be verified ^[931]Author Correction: On the conversational persuasiveness of GPT-4Salvi, Francesco; Horta Ribeiro, Manoel; Gallotti, Riccardo et al. · 2026 · paperView source ↗Accessed on 9 September 2026.