Our Take
AI's persuasion edge collapses when forced to match human writing speed and message length, meaning the breakthrough is speed, not reasoning.
Why it matters
This is the first large-scale evidence that current LLMs outperform humans at a task with real-world consequences (money donated). The finding matters now because it establishes a baseline for policy: persuasion at scale is already asymmetric, and controlling access to these systems will reshape political and commercial influence.
Do this week
Security teams: document your organization's exposure to AI-driven persuasion campaigns—social media, customer service, donor outreach—and audit which channels accept unsupervised LLM interactions before this capability commoditizes further.
Four Experiments Prove AI Beats Human Experts at Persuasion
Researchers from Oxford, the UK AI Security Institute, Stanford, and the London School of Economics tested how well AI systems persuade humans to change policy views and donate to charity. Across four experiments involving 18,978 conversations with 6,923 people, AI systems outperformed every class of human persuader tested: random laypeople, tournament-selected debaters, and elite debaters with advance preparation, structured coaching, and £1,000 cash incentives.
In Study 1, AI exceeded all human persuaders on 10 UK policy questions. In Study 2, the researchers gave 43 elite debaters access to a coaching tool built around the AI that had beaten them. They could chat with the AI, see its prompts, review annotated transcripts showing which moves shifted opinions, and see what the AI would have said in their place. Coaching narrowed the gap but did not close it.
Study 3 revealed the mechanism. When researchers forced AI to write at human speed and length, "AI's advantage over the strongest human comparator within Study 2 (Coached Elite Debaters) collapsed from +4.1 percentage points to a non-significant 0.0 percentage points." The largest reductions in perceived argument strength came when AI was constrained on information density and output volume.
Study 4 tested real-world stakes. Researchers recruited 19 experienced canvassers from a UK fundraising firm and had them compete against AI in conversations about Save the Children. The firm had raised £824,297 from 22,583 donors over seven years. In this test, AI exceeded professional canvassers by 5.9 percentage points. When persuadees were given a £1 study bonus and allowed to donate any portion to the charity, AI elicited substantially more real-money giving, exceeding canvassers by 10.8 percentage points. AI raised both the share of donors and the average donation size.
The strongest performers were Claude Opus 4.1 and 4.6, followed by GPT-4o, GPT-5.4 (company-reported), Gemini 2.5 Pro, and Grok 4.20.
The Gap Is About Information Volume, Not Reasoning
This finding matters because it establishes that frontier LLMs can already out-perform human experts at a consequential task in the real world. Fundraising is not hypothetical; donations moved. But the mechanism is instructive: AI wins because it deploys larger quantities of information faster, not because it reasons better.
The authors note two policy implications. If persuasion capabilities become cheap and widely available, under-resourced actors (public defenders, small charities, grassroots activists) could narrow power gaps against better-funded rivals. Conversely, if only already-powerful actors control these systems, influence could consolidate further. The question is not whether AI can out-persuade humans, but how, where, and on whose behalf this capability will be exercised.
The finding also exposes a tension: humans cannot match AI's speed advantage through training. Coaching elite debaters did not close the gap. This suggests that persuasion at scale will remain asymmetric unless constrained by policy or architecture.
Audit Your Exposure to AI-Driven Persuasion
Organizations that rely on human persuasion—fundraising, sales, policy advocacy, customer retention—should inventory which channels accept LLM-generated or LLM-assisted messages. If your donor outreach, customer service, or sales workflows use unsupervised LLM interactions, you are already operating in a domain where an AI system can outperform your team by 10 percentage points on money moved.
For compliance and security teams: document which systems have access to persuasion tools, how they are monitored, and whether you have baselines for AI-generated vs. human-generated outcomes. The commodity version of this capability is coming; knowing your current exposure is foundational to setting guardrails before it arrives.