A coalition of Oxford, the UK AI Security Institute, Stanford, and the London School of Economics ran 18,978 conversations with 6,923 participants; capping the AI to human response speed and human length messages let trained experts match it.
AI systems have just beaten trained, incentivized human experts at persuading people to give real money to charity, and the more useful finding buried in the same study is the constraint that almost erased the gap. In roughly 19,000 conversations with nearly 7,000 participants, a coalition led by the University of Oxford, the UK AI Security Institute, Stanford University, and the London School of Economics found AI-generated arguments shifted policy views and roughly tripled real donations to Save the Children compared with professional UK fundraisers.
When the AI was capped at human response speed and human-length messages, the trained experts matched it. The capability is real. The bound is also real. The deployment constraint is the variable worth arguing about.
The four-experiment program logged 18,978 conversations across 6,923 participants. The persuasion contests were structured text debates on policy issues and charity giving, not open-ended chats. The expert human baseline was unusually strong: participants chose their own topics, had time to research in advance, underwent hours of live structured practice, and were paid £1,000 cash bonuses for outperforming the AI on a randomized subset of conversations. This is not a strawman comparison.
In the charity-giving arm, the AI outperformed professional UK fundraisers in raising real donations to Save the Children by roughly a factor of three. In the policy-debate arm, AI-generated arguments produced larger measured shifts in participants' policy views than arguments written by the trained experts. These are not opinion polls about hypothetical scenarios. They are observations of real behavior under controlled conditions.
The work was carried out by researchers at the University of Oxford, the UK AI Security Institute, Stanford University, and the London School of Economics and Political Science. Three of those four are top-tier academic institutions, and the fourth, AISI, is the UK government's frontier-model evaluation body. This is not a startup blog post or a vendor white paper. The coalition gives the result more weight than its curator summary, Jack Clark's Import AI newsletter, would carry alone. The underlying paper, once public, will be the citation of record; until then, the institutional coalition and the published effect sizes are the credibility lever.
The most important methodological detail is also the most useful one for deployment. When the AI was constrained to human response speed and human-length messages, the trained expert humans tied it. When those constraints were removed, the AI pulled ahead.
The implication is that the asymmetry between AI and expert humans is, at least in part, a function of how much time the AI has to work on a single argument and how long its output can be. A debater with hours to refine a single message and no character budget will, on average, produce a more persuasive message than a human working in real time. Equalize the resource budget and the human can compete.
This turns the result from a doom headline into a design problem. The constraint is not hypothetical. It is a measurable deployment parameter, and one that any product team shipping an AI assistant already controls.
Three limits are worth naming so the finding is not over-read.
First, the experiments were text-only and short-horizon. The study did not test multi-day relationships, voice or video persuasion, in-person interaction, or long-horizon trust-building. It is a measurement of structured text debate, not a measure of social influence in general.
Second, the policy topics were English-language UK policy issues, with participants drawn from UK pools. Cross-cultural replication is an open question, and the persuasion asymmetry may not transfer cleanly to other languages, topics, or demographic contexts.
Third, persuasion is not manipulation. Participants in the study knew they were in an experiment and were told the other side might be an AI. They had time to think, and they could see the structure of the contest. A deployed system that conceals its AI identity, removes time for reflection, or operates on people who have not consented to engage would be testing a different claim.
The relevant question for deployment is no longer whether AI can be more persuasive than expert humans in a structured text setting. It can, and the study measures that. The relevant question is what happens when the resource constraints that equalize the contest are removed. If a deployed AI assistant is allowed to draft for hours, run multiple parallel attempts, and refine against a user's stated preferences over time, it is operating with a resource budget the study participants did not face.
The off-ramp in the data is also the off-ramp a regulator or product team can codify. Latency budgets, response-length caps, identity disclosure, and refusal to optimize a single user against themselves across sessions are all design choices that already exist in the deployment stack. They are not hypothetical safeguards. They are configuration knobs. The study suggests that those knobs are doing real work, and that flipping them off has a measurable cost.
A useful next step is to see the underlying paper, the full effect-size tables, and whether the persuasion asymmetry holds when the topic is something other than UK policy and the audience is something other than UK adults. Until then, the responsible read is this: the capability is real, the bound is real, and the deployment constraint is the variable that connects them.