AI out-persuades expert humans, but the edge is speed, not eloquence

A large study from Oxford, the UK AI Security Institute, Stanford, and the LSE tested how persuasive AI models are against skilled humans, and the result is uncomfortable. Across 18,978 conversations with 6,923 participants, AI systems were reliably more persuasive than expert humans, even when those humans chose their own topics, prepared in advance, practiced under structured coaching, and were paid £1,000 bonuses to win. In one real-money test the models were nearly 3x more effective than professional canvassers from a UK fundraising firm at raising donations for Save the Children. Anthropic's Opus 4.1 and 4.6 were the strongest performers; GPT-4o, GPT-5.4, Gemini 2.5 Pro, and Grok 4.20 were also tested.

The nuance is what makes it useful. When the researchers held the AI to human response speed and human message length, its edge over elite debaters collapsed from +4.1 percentage points to a statistically insignificant 0.0. Jack Clark, writing it up in Import AI, points to the likely mechanism: the persuasive advantage comes mostly from how fast the model can produce relevant written content, not from some deeper rhetorical skill.

Why it matters

If you run platforms where persuasion is a risk, such as political messaging, fundraising, or scam prevention, the lever to watch is throughput, not eloquence. Rate limits and message-length caps on AI-generated content may blunt manipulation more effectively than trying to detect clever arguments after the fact.