Across four preregistered experiments and nearly 19,000 conversations, frontier AI systems reliably changed minds more often than expert humans who had chosen their topics, researched ahead, drilled for hours, and even chased cash bonuses for top performance.
AI systems out-persuade expert humans
"We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses."
"AI was nearly 3x more effective than professional canvassers from a UK fundraising firm at raising real-money donations to Save the Children."
The AI kept its lead until researchers forced it to reply at human speed and length; once those limits applied, coached people tied the model, showing the edge was volume and pace rather than superior rhetoric.
In the final test the same systems nearly tripled the money raised by professional canvassers for Save the Children, hinting at how quickly persuasion tech could move from lab benchmarks into campaigns and policy fights.
Positive users call the Oxford study on frontier AI outpersuading world-champ debaters fascinating research, while negative users find the results concerning.
No Digg Deeper questions have been answered for this story yet.
abs: https://arxiv.org/abs/2606.16475
AI systems out-persuade expert humans
"We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses."
"AI was nearly 3x more effective than professional canvassers from a UK fundraising firm at raising real-money donations to Save the Children."
AI's advantage also extended to high-stakes real-world action:
We pitted AI against a professional fundraising firm with 7 years' experience raising money for Save the Children - to raise real donations for Save the Children.
AI's effect on giving was nearly 3x the firm's.
More on this project below 🧵
New w/ @AISecurityInst & @UniofOxford:
Frontier AI can now out-persuade expert humans in conversation - incl. world-champ debaters and professional canvassers.
This held even when humans chose their topics, prepared in advance, and competed for £1,000 prizes 🧵
Could humans get better once they saw what AI was doing?
In a follow-up study, we built a coaching tool that let the elite debaters: - practice against the model that beat them - see exactly how it was prompted - revisit their own past conversations
Even after coaching, AI was still more persuasive.
Why does throughput increase persuasion? We find converging evidence that it does so by allowing AI to deliver more information in support of its arguments.
Unconstrained, AI made ~37 fact-checkable claims per conversation, vs ~12 when constrained - and across humans and AI, fact density predicted persuasion (R² = 0.89).
Our results suggest that AI can now reliably out-persuade expert humans.
Predicting the consequences of this for society is challenging and requires us to make assumptions about how, where, and on whose behalf this capability will be exercised.
@sama may have been correct
This was a big effort - huge thanks to my collaborators @acarolinewagner @lukebeehewitt @Ben_Tappin @hannahrosekirk @HelenMargetts @summerfieldlab, and to many at @AISecurityInst for making it possible 🙏
Comments and feedback welcome!
Paper: https://arxiv.org/abs/2606.16475
Our evidence suggests much of AI's advantage comes from its superior *throughput*:
when we constrained the AI to human-length messages and human writing speeds, its advantage over our strongest human group (coached debaters) was no longer statistically significant.
and since much human persuasion is itself misleading, it's possible that AI persuasion could actually inform people better than the status quo 🤷♂️
what matters is whether AI is better or worse than the persuasion it replaces (see great essay from @Ben_Tappin):
One plausible concern is that this could consolidate power: persuasion contests once decided by human skill could instead hinge on who can access (purchase?) the most capable AI.
Or, in cases where both sides can access the best AI, power would flow to those who build and controls the models themselves (and therefore can decide the issues on which the models will and will not advocate for).
Prior work comparing human & AI persuasion has used random crowd workers as the human comparator.
But IRL, consequential persuasion contests - e.g. elections, fundraising, public health debates - are won by trained, highly motivated experts.
Could AI out-persuade *those* people?
To find out, we recruited the strongest human persuaders we could find:
- top performers from a multi-week, 1,150-person online persuasion tournament - world-championship debaters (incl. 4 world & 11 continental champs) - professional canvassers with thousands of hours of experience
Were there conditions where humans came out on top? Not really.
Humans didn't out-persuade AI on any of 10 issues, or in any of 49 demographic, political & psychological persuadee subgroups (AI's edge significant in 46 of them).
Across 318 per-persuader estimates (275 humans), we estimate 0 beat AI.
It could also have democratizing effects. Cheap, widely available expert-level persuasion might help under-resourced actors (e.g., small charities, legal aid, grassroots) compete with richer rivals.
Very interesting. AI is better than human experts at persuasion. That's coupled to its ability to deliver a high volume of facts. Mastery of the subject, rather than a silver tongue, appears to be the driver. An optimistic finding for our polluted information environment.
New w/ @AISecurityInst & @UniofOxford:
Frontier AI can now out-persuade expert humans in conversation - incl. world-champ debaters and professional canvassers.
This held even when humans chose their topics, prepared in advance, and competed for £1,000 prizes 🧵
We then ran a series of large, pre-registered RCTs (~19k conversations, ~6.9k people), matching each persuadee in real time with an AI or a human persuader.
Despite giving the humans every advantage we could think of (e.g., research time, live practice, numerous cash bonuses), AI out-persuaded every human group.
In our study, participants were paid to engage with our models. But getting people's attention IRL might be a more important constraint on the impact of AI persuasion than actual per-message persuasiveness.
A thoughtful piece on this from @BrendanNyhan & collaborators:
Regardless of these uncertainties, we hope this work can be a starting place for discussion (and further research!) on the impacts these persuasive systems could have on society in the near future
AI is more persuasive than the best humans because it delivers more facts. Not because it's better with words.
Why does throughput increase persuasion? We find converging evidence that it does so by allowing AI to deliver more information in support of its arguments.
Unconstrained, AI made ~37 fact-checkable claims per conversation, vs ~12 when constrained - and across humans and AI, fact density predicted persuasion (R² = 0.89).
@ravi_mohan numerically, the coached humans did get a a bit better - and humans wrote longer, more fact dense messages after coaching (this was statistically significant) - but the overall improvement in their persuasiveness wasn’t statistically significant
