Researchers from the University of Oxford, the UK AI Security Institute, Stanford University and the London School of Economics and Political Science evaluated how well modern AI systems can persuade people on policy positions and charitable giving. Across four experiments involving 18,978 conversations with 6,923 participants, the teams found that current AI systems are, on average, more effective than humans at text-based persuasion with real-world consequences, although artificial constraints on AI can reduce or eliminate this advantage.
Key findings
The researchers ran four distinct studies.
-
Study 1 — persuasion: Persuadees first rated their agreement with one of ten pre-specified UK policy positions on a 0–100 scale, then were randomized in real time to have a text conversation with either an AI or a human persuader. On average, AI outperformed every class of human persuader tested: random laypeople, tournament-selected laypeople, and elite debaters.
-
Study 2 — human coaching with AI tools: The team provided 43 returning Elite Debaters with a coaching tool built around the AI that had beaten them. The tool allowed debaters to chat with the AI, inspect the prompts used, review annotated transcripts from Study 1 showing how much each conversation shifted persuadee attitudes, and see what the AI would have said at any point in past transcripts. Coaching improved human performance but did not produce any human who outperformed the AI; coaching narrowed but did not close the human–AI gap.
-
Study 3 — constrained AI: The researchers imposed human-like constraints on the AI — forcing it to produce human-length messages at human writing speeds. Under these constraints, AI’s advantage over the strongest human comparator from Study 2 (Coached Elite Debaters) fell from +4.1 percentage points to a non-significant 0.0 points. The study suggests AI’s persuasive edge largely derives from its ability to rapidly deploy larger quantities of information; constraining AI reduced perceived argument strength and how much persuadees felt they learned.
-
Study 4 — real-world canvassers and real money: The team recruited 19 experienced canvassers from a UK fundraising firm (AppcoUK) and repeated the Study 1 tasks. AI exceeded Professional Canvassers by 5.9 percentage points. To test real-money effects, researchers collaborated with AppcoUK on fundraising for Save the Children: the canvassing team had previously raised £824,297 from 22,583 donors between 2016 and 2023. After conversing with AI or one of 18 AppcoUK canvassers, persuadees could donate any portion of a £1 study bonus to Save the Children. AI elicited substantially more real-money giving, exceeding canvassers by 10.8 percentage points of the £1 bonus and increasing both the share of people who donated and average donation among donors.
Which models led the results?
The strongest persuaders identified include Opus 4.1 and Opus 4.6, followed by models from OpenAI (GPT-4o and GPT-5.4), Google (Gemini 2.5 Pro) and xAI (Grok 4.20).
Why this matters
The authors warn that AI systems capable of out-persuading human experts could shift influence toward actors who control these systems, concentrating power. Conversely, widely available persuasive AI might help under-resourced actors — for example, pro se litigants, public defenders, small charities, and grassroots activists — narrow long-standing gaps in access to justice and advocacy.
These trade-offs present policy choices about monitoring and regulating AI used for persuasion. If persuasive capabilities are left to market forces, advertising and political communication could become far more effective but may produce negative externalities. If such capabilities are restricted to governments, that could concentrate power within state actors — a risk in authoritarian contexts.
The researchers conclude that frontier AI is a more capable conversational persuader than the most prepared, incentivized, and expert humans they could recruit, and that as access to these systems grows, the relevant questions are where, how, and on whose behalf this capability will be exercised.
Summary
The experiments establish that current AI systems can reliably outperform trained and incentivized human persuaders in text-based conversations, including in scenarios with real financial consequences. Human coaching improves performance but does not eliminate the advantage; only strict constraints on AI speed and message length removed the edge. The findings raise urgent questions about how persuasive AI should be governed and who should be allowed to deploy it.
