AI decisively out-persuades expert humans, raising nearly 3x more real-money donations
Import AI 462: Superpersuasion; self-sustaining AI; paths to ASI
Across four studies with 18,978 conversations and 6,923 participants, researchers from Oxford, AISI, Stanford, and LSE found that current AI systems reliably out-persuade expert humans—even those coached and incentivized with £1,000 bonuses. Opus 4.1 and 4.6 led, followed by GPT-4o, GPT-5.4, Gemini 2.5 Pro, and Grok 4.20. The edge comes from rapidly deploying more information: when constrained to human-length messages at human speed, the AI advantage disappeared. In a real-money test, AI raised 10.8 percentage points more of a £1 bonus for Save the Children than professional fundraisers. The authors warn that concentrated control of such persuasion could consolidate power, but broad access might help under-resourced groups compete.
Why it matters: Researchers from Oxford, the UK AI Safety Institute, Stanford and LSE ran four experiments and nearly 19,000 conversations to compare how well AI and human experts persuade in writing.