AI is more persuasive than trained, paid human experts at changing people's opinions on political topics, and several times more effective than professional canvassers at driving real donations — but the advantage largely disappears when the AI is forced to operate at human speed and message length. The finding comes from a large multi-institution study across tens of thousands of conversations and over 70,000 participants.
Key facts
- What: Across nearly 19,000 conversations, AI outargued incentivized human experts and raised real donations far more effectively, but its edge collapsed when slowed to human speed.
- When: 2026-06-26
- Primary source: read the source (arXiv 2507.13919)
The study, titled The Levers of Political Persuasion with Conversational AI, appears in the journal Science with a preprint at arXiv:2507.13919, and the AI Security Institute summarized it. It led Jack Clark's widely read Import AI newsletter. The work involved researchers at Oxford, the UK's AI Security Institute, Stanford, and LSE.
Persuasion research has traditionally been small-scale: a few hundred people in a lab, one argument, a measured shift. The concern with AI is that a system can hold tailored, patient, well-sourced conversations with millions of people simultaneously, making influence campaigns vastly cheaper and more scalable. This study brought the scale needed to test that concern: tens of thousands of conversations, tens of thousands of participants, and hundreds of political topics, with direct comparisons against humans who were trained and paid to be convincing.
The AI was more effective at shifting opinions than the expert humans, and several times more effective than professional canvassers at getting people to make real donations — not just say they would. That matters because actual behavior is a far harder test than a survey answer.
But the researchers also identified the mechanism behind the advantage, and that mechanism is reassuring in a specific way. The AI's edge largely collapsed when it was forced to operate at human speed and human message length. The model was not winning through superhuman insight or deeper empathy. It was winning because it could deploy a large volume of relevant, on-point information very quickly. Slow it down to the pace and length a person can manage, and the gap mostly closes. Think of a debate where one side can instantly pull up a relevant fact, a tailored example, and a crisp rebuttal for anything said, while the other works at normal human speed. The fast side wins on sheer throughput, not wisdom — like a chess player who gets ten moves to your one. Remove the speed advantage and it becomes a fair fight. The AI's superpower here is volume and velocity of relevant information, not a magical grasp of what moves people. For the underlying concept, see the explainer on AI persuasion.
This reframes the policy conversation. If the danger were that AI understands us better than we understand ourselves, there would be little to do but despair. If the danger is instead rapid, scaled information deployment, the countermeasures are more concrete: rate limits, transparency about when you are talking to a machine, and friction on automated mass outreach. It also raises the stakes for the persuasion arriving in everyday life — from chatbots to customer service to political messaging — where the speed advantage is fully unleashed.
The team reported nearly 19,000 conversations and over 70,000 participants, but the full paper sits behind journal and bot-detection walls that blocked direct extraction of every figure during review, so the precise numbers are drawn from the authors' own abstract, the AI Security Institute's summary, and trusted reporting rather than a line-by-line read of the final text. The headline finding — AI out-persuading paid experts with an edge that depends on speed — is well supported across those sources. Treat exact magnitudes as approximate until the full text is openly readable. There is also a generalization limit: this measured political and donation persuasion in a controlled setting, and real-world influence is messier, with trust, identity, and repeated contact all playing roles a single conversation cannot capture.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)