An AI persuasion study posted to arXiv reports that AI systems were more effective than several kinds of human persuaders, including professional canvassers and world championship debaters. The finding matters because persuasion is not an abstract parlor trick: campaigns, charities, companies and advocacy groups use it to change opinions and, in at least one test here, actual spending decisions.
The researchers ran four preregistered experiments involving 18,978 conversations with 6,923 participants. They compared AI systems with ordinary participants, winners of a separately preregistered four-round online persuasion tournament, professional canvassers and elite debaters.
According to the paper, the AI systems beat expert humans even when those humans picked their own topics, prepared ahead of time, received structured live practice and had a chance to earn £1,000 bonuses. A later test gave experts a coaching tool that let them practice against the AI system that had beaten them, review earlier performance and see what the AI would have written at selected points. The AI advantage still remained, the researchers said.
Why did AI beat human persuaders?
The researchers attributed the gap mainly to information volume and delivery speed. AI systems could produce more relevant material quickly, while humans had to work within ordinary limits of memory, typing and response length.
That mechanism matters because it is boring in the way many powerful systems are boring. The paper does not describe mind-control magic. It describes machines using more facts, faster, in a format that keeps a person engaged long enough for those facts to matter.
The researchers found a boundary condition, though. After coaching, expert humans could match an AI system when the AI was restricted to human-like response speed and human-length messages. That suggests the advantage came less from some mysterious social instinct and more from scale: retrieval, synthesis and rapid output.
The study also moved beyond opinion change. In a test involving real-money charity donations, the AI systems generated substantially more giving than well-paid professional canvassers, according to the researchers. The paper does not mean every chatbot will beat every human in every setting, but it does give campaigners and regulators a concrete result to argue about instead of vibes.
Who gains power from persuasive AI?
The researchers warn that persuasive AI could strengthen whoever can get and run the most capable systems. They point to private corporations, political campaigns and nation states as actors that already spend heavily to influence consumers and voters. If AI raises the effectiveness of each message, existing gaps in influence could widen.
They also flag a second concentration problem: the companies that build and control top AI systems could shape persuasion contests even when users have equal access. Model providers could decide which positions their systems will argue for, or refuse to argue for, giving suppliers leverage over the contest itself.
The paper also notes a less grim possibility. If costs fall, persuasive AI could help poorly funded actors, including self-represented litigants, public defenders, small charities and grassroots activists, compete with better-funded opponents.
Tom Stafford, a psychology professor at the University of Sheffield and co-author of Mind Hacks, argued in a commentary that fact-heavy persuasion may be compatible with human reason rather than hostile to it. He also cautioned that people are not blank screens. They can resist persuasion, and the experiments required roughly 14 minutes of engagement, which may be more attention than many people give opposing views in daily life.
The unresolved question is where these systems get deployed and by whom. The study gives strong evidence that AI can outperform trained human persuaders under tested conditions. It does not settle whether that power will broaden access to advocacy, concentrate influence among large institutions, or hand more gatekeeping power to AI companies.
This story draws on original reporting from Techdirt.
