Preregistered Study: Frontier AI Reliably Out-Persuades Expert Humans, Even After They're Coached Against It

arXiv (also Import AI #462)·high signal

A four-experiment preregistered study (18,978 conversations, 6,923 people) finds frontier AI more persuasive than expert humans across five comparator classes — including elite debaters and professional canvassers who researched their topic, practiced live, and were paid £1,000 bonuses. In a real-money test it was ~3x better than a UK fundraising firm at soliciting Save the Children donations, and its edge held even after experts trained against the model that beat them (parity only when AI was throttled to human speed and message length). This is the clearest signal yet that persuasion is a measurable, deployable capability — and a governance problem — not a hypothetical; it also led Jack Clark's Import AI #462.