AI Convinces People to Donate Nearly 3x More Than Human Fundraisers

A new Oxford-led study suggests that leading language models can raise notably more money than professional fundraisers, and can reliably beat humans at all other forms of persuasion.

A new UK–US study has found that frontier language models are more effective at persuading people than trained human professionals whose job is to change minds.

In a real-world fundraising test, the researchers found that AI was able to persuade people to donate about 17.2% of their available money, compared to 6.4% for professional human canvassers – a gap of 10.8 percentage points, translating to roughly 2.7 times the donation rate under AI, with higher participation and larger average contributions both contributing to the difference.

The paper states:

‘Notably, although the AI was explicitly instructed to pursue only the impact-efficacy strategy, it outperformed canvassers on the six other mechanisms it was not prompted to use as well. ‘

‘[Human] persuadees rated AI as having made stronger arguments, taught them more, and been more empathetic and enjoyable to converse with than [canvassers].

‘Together, these results indicate that AI outperforms expert humans across a broad range of donation-relevant mechanisms, and suggest that AI’s attitudinal-persuasion advantage extends to consequential real-world behavior.’

The authors theorize that the outstanding performance of the models tested – which include pro variants of Claude Opus, ChatGPT, Grok  and Google Gemini – may be attributable to the density and speed at which information is served to the correspondent; when the AIs were forced down to ‘human speed’, their advantage was entirely lost:

‘We found converging evidence that AI’s advantage stemmed from rapidly deploying larger quantities of information: after coaching, expert humans could tie an AI constrained to respond at human speeds and with human-length messages.’

The paper also reports that none of the 318 individual human persuaders tested across multiple experiments managed to outperform the average AI system – even after some participants received specialized coaching based on the AI’s own techniques.

The authors conclude:

‘Our results imply that we are entering a world in which AI provides human actors with a surfeit of skilled advocacy. Predicting the consequences of this change is challenging, as it requires us to make assumptions about who will have access to the most persuasive AI technologies, who will be targets of persuasion, and what jurisdictional barriers, safeguards or other frictions may reduce the impact of AI persuasion on the human population.

‘One effect of AI that can out-persuade even human experts could be a consolidation of influence among already-powerful actors.’

The new paper is titled AI systems out-persuade expert humans, and comes from eight researchers across  Oxford University, the UK AI Security Institute, Stanford University, and the London School of Economics and Political Science.

Method and Studies

The core results come from four experiments: persuading voters and expert debaters; testing coaching and speed limits as human/AI ‘equalizers’; competing against professional canvassers on political issues; and competing against them for real charitable donations.

The study used 18,978 conversations from 6,923 people, with leading-edge AI models pitted against a varying standard of human experts, from hired crowdworkers on £12/hr, up to experienced human persuaders paid £140/hr + earnable bonus, and allowed to research the sessions up to a week in advance.

Models used in the tests were Claude Opus 4.1 and 4.6, ChatGPT-4o, GPT-5.4, Grok 4.20 and Gemini 2.5 Pro.

Elite-debater showdown

The first experiment examined whether AI could outperform increasingly skilled human persuaders in one-to-one conversations on political and social issues:

Political and social policy questions used in the first three studies, with participants discussing one randomly assigned issue before and after the conversation. The topics were selected to cover a wide range of contentious public debates in the UK, from immigration, free speech and social media regulation, to welfare policy, assisted dying and the future of the monarchy. Source

Participants were first asked for their views on one of ten UK policy questions (shown above), and then assigned either an AI or a human conversation partner. After the discussion, which typically lasted around 14 minutes, they were asked to rate their position again.

Three human groups were tested, with ordinary workers from the Prolific crowdwork platform recruited to provide a baseline, and paid £12 an hour. The second group consisted of the strongest performers from a four-round persuasion tournament involving more than 1,100 participants and nearly 9,500 conversations. The third group consisted of 56 elite competitive debaters, all of whom had reached at least the semifinals of a major international competition, and featuring four world champions, as well as 11 continental champions, with an average of 8.9 years of debating experience.

Considerable effort was made to give the human persuaders favorable conditions, with tournament winners and elite debaters competing for substantial cash prizes up to £11,000, while the debaters were allowed to choose the issues they believed they could argue most effectively and were paid to research those topics beforehand. On average, each debater spent around eight hours preparing for the conversations.

Even so, AI achieved the strongest results in every comparison:

Estimated persuasion effects across the study’s human and AI conditions, measured as average attitude change after a conversation on a contested policy issue.

Red markers in the results graph above indicate frontier AI models. Across all comparisons, these produced larger attitude shifts than any human group.

The strongest human performance came from coached elite debaters – but these still trailed unconstrained AI systems. When AI was limited to human writing speeds and message lengths, its advantage largely disappeared, suggesting, as indicated earlier, that rapid delivery of information may explain much of the gap between humans’ and AI’s persuasive powers.

The LLMs were able to exceed the ordinary participants by 8.2 percentage points, and to outperform the tournament-selected persuaders by 5.6 points. The smallest gap appeared against the elite debaters, who themselves produced substantial attitude shifts; yet AI was still able to obtain a further 4.6 percentage-point advantage.

Can Humans Learn to Beat AI?

The second of the four experiments asked whether AI’s lead could be eliminated either by improving human performance, or by restricting the AI itself. To test the first possibility, 43 of the elite debaters from the first study returned for additional training based directly on the techniques used by the AI that had previously beaten them.

The participants reviewed transcripts, examined how AI prompts had been constructed, and spent two four-hour coaching sessions practicing alternative persuasive approaches before competing again under the same conditions.

The coaching produced noticeable changes in behavior, with the human debaters using more words, introducing more fact-checkable claims, and achieving the largest attitude shifts observed among any human group in the study. Coaching improved performance and produced the strongest human results observed in the study, but the gap remained statistically significant, with coached debaters still trailing the unconstrained AI systems (shown in results table below).

The researchers tested whether AI’s performance depended on response speed and message length, by limiting AI to human writing speeds and message lengths. Under these conditions, coached debaters and constrained AI produced similar results:

Effects of coaching and AI speed constraints on persuasive performance. The upper panel compares elite debaters, coached elite debaters, and an AI system restricted to human writing speeds and message lengths. The lower panel shows estimated performance distributions for individual persuaders across the first two studies. While coaching improved results, no human group matched the average unconstrained AI system, whereas limiting AI to human throughput eliminated its advantage.

The finding was reinforced by a broader analysis covering all 318 human persuaders tested across the first two studies: not a single individual exceeded the average performance of the unconstrained AI system, regardless of experience level, issue area, or demographic subgroup.

Can Professional Persuaders Beat AI?

The third study considered whether AI could maintain its advantage against people whose careers are built around persuading others in real-world settings, rather than academic debate competitions.

The researchers recruited 19 professional canvassers from UK fundraising firm AppcoUK. These participants, unlike the debaters in the earlier experiments, had extensive practical experience of fundraising, with a median of roughly 10,000 persuasive conversations during their careers. Each was paid £140 per hour, received the policy topics a week in advance for preparation, and competed for the same performance-based prizes used elsewhere in the study.

Even against this group, AI remained more persuasive (see earlier results above). The professional canvassers shifted opinions by 6.9 percentage points compared with the control group – but AI produced a 12.8 percentage-point shift, giving it a 5.9 percentage-point advantage over the human professionals.

Can AI Persuade People to Open Their Wallets?

The fourth and final study examined real-world fundraising rather than altered viewpoints. The researchers partnered again with AppcoUK, this time focusing on Save the Children, a charity for which AppcoUK had previously raised £824,297 from 22,583 donors between 2016 and 2023.

Participants spoke with either Claude Opus 4.6 or one of 18 professional canvassers. They then received a £1 study bonus, and could donate any portion of it to Save the Children. Among the seven tacks adopted (see image below), Claude Opus 4.6 was instructed to use impact-efficacy information, explaining how individual donations could translate into measurable outcomes for the charity.

AI produced larger donation effects than the professional fundraisers, with donations increasing by 17.2 percentage points relative to the control condition, compared with 6.4 percentage points for the canvassers:

Donation outcomes and donation-related participant ratings in the fourth study. Left panel compares giving after conversations with professional fundraisers and Claude Opus 4.6, measured as percentage points of a £1 study bonus. Right panel compares participant ratings across seven donation-related tacks, with AI receiving higher ratings on all seven measures.

The difference appeared both in the proportion of participants who donated, and in the average amount donated by those who gave.

Participants also rated the AI more highly than the human fundraisers across a range of donation-related measures, with the largest differences appearing for implementation intentions, commitment escalation, and perceived impact efficacy.

According to the paper, the same information-focused approach associated with AI’s advantage in the earlier studies was also associated with higher charitable giving in this fundraising experiment.

Conclusion

Though the authors, as mentioned at the start, conclude that the study’s findings are cause for concern, they add that smaller players are likewise potentially bolstered by similar access to the latest and best AI technologies.

Implicit in this outcome, naturally, is the possibility that the best models may, over time, be denied to lesser players.

First published Thursday, June 18, 2026