Study finds AI can out-persuade human champion debaters in real time

A paper by researchers including Kobi Hackenburg and Christopher Summerfield, “AI systems out-persuade expert humans”, reported that frontier AI models were more persuasive in live text conversation than expert human persuaders, across four preregistered experiments totalling nearly 19,000 conversations with almost 7,000 participants. The comparison was deliberately weighted toward the humans: expert opponents chose their own topics, researched in advance, practised for hours in structured sessions against the AI, and were paid cash bonuses of up to £1,000 for strong performance. The AI’s advantage held regardless, including against world-championship debaters.

The researchers traced the effect to throughput rather than superior argument quality: AI systems could produce and deploy a much larger volume of relevant information within the same conversation than a human could type or recall in real time. When both sides were constrained to match human speed and message length, the human experts’ performance caught up to the AI’s. In one real-world test, the researchers found AI persuaders raised close to three times as much in donations to Save the Children as professional canvassers from a UK fundraising firm working the same conversations.

The finding fed directly into ongoing arguments about AI and information environments: if persuasive advantage comes chiefly from speed and volume of argument rather than novel rhetorical skill, it suggests the effect scales with compute and access rather than requiring any conceptual breakthrough — a point cited on both sides of the debate over AI-enabled disinformation and political manipulation.