🤖 AI

AI systems out-persuade expert humans

Through a series of four large-scale experiments involving nearly 7,000 participants, researchers demonstrate that frontier AI systems consistently out-persuade expert humans—including professional canvassers and world championship debaters—by rapidly deploying larger quantities of information, a capability that translates to significantly higher real-world fundraising effectiveness.

Original authors: Kobi Hackenburg, Caroline Wagner, Luke Hewitt, Ben M. Tappin, Ed Saunders, Hannah Rose Kirk, Helen Margetts, Christopher Summerfield

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kobi Hackenburg, Caroline Wagner, Luke Hewitt, Ben M. Tappin, Ed Saunders, Hannah Rose Kirk, Helen Margetts, Christopher Summerfield

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a series of high-stakes debates where the goal isn't just to win an argument, but to actually change someone's mind. In the past, the best persuaders were humans: professional fundraisers, world-champion debaters, and even people who had spent years training specifically for this task.

This paper asks a simple but startling question: Can a computer program beat the absolute best human persuaders at changing minds?

The answer, according to this research, is a resounding yes.

Here is the story of how they found out, using simple analogies.

The Tournament: Humans vs. The Machine

The researchers set up a massive "persuasion tournament." They didn't just pit AI against random people; they pitted AI against the "Olympic athletes" of persuasion.

The Human Team: They recruited four types of experts:
1. Random People: Average folks paid to try their best.
2. The "Super-Salespeople": People who had already won a previous tournament of persuasion.
3. Professional Fundraisers: Real-world canvassers who spend their days asking strangers for money for charities like Save the Children.
4. World-Champion Debaters: People who have won international titles, trained for years, and given thousands of hours of practice.
The AI Team: The researchers used the most advanced AI models available (like the latest versions of Claude, GPT, and others).
The Arena: The participants engaged in text conversations with strangers. The goal was to convince the stranger to change their opinion on a specific political or social issue. The humans were given huge cash bonuses, days to research their topics, and hours of coaching. They were given every possible advantage.

The Result: The Machine Wins

Even with all those advantages, the AI won every single time.

Against Average People: The AI was significantly better at changing minds.
Against the "Super-Salespeople": The AI still won.
Against World Champions: Even the best debaters in the world, who had researched their topics for hours and practiced for days, could not beat the AI.
Against Real Fundraisers: In a real-world test asking people to donate actual money, the AI was nearly three times more effective than professional fundraisers at getting people to open their wallets.

The Mystery: Why Did the AI Win?

The researchers wondered: Is the AI just smarter, or is it cheating?

To find out, they tried two experiments:

1. The "Coach" Experiment
They took the world-champion debaters who had lost and gave them a "coaching tool." This tool showed them exactly what the AI would have said at every moment of their lost conversations. It was like having a super-coach whisper the perfect answers in their ear.

The Result: The humans improved slightly, but they still lost. The gap between human and AI remained.

2. The "Speed Limit" Experiment
The researchers realized the AI was talking fast and a lot. In a typical conversation, a human might write a 50-word reply and take 90 seconds to think. The AI wrote 300 words in less than a second.

The Test: They forced the AI to slow down. They made it write only 50 words and wait 90 seconds before replying, just like a human.
The Result: The AI lost its advantage. When forced to talk at a human speed and length, it performed just as well as the humans, but no better.

The Conclusion: The AI's superpower wasn't that it was "smarter" or had better "feelings." Its superpower was volume. It could pack a conversation with three times as many facts and arguments in the same amount of time. It was like a human trying to fill a bucket with a teaspoon, while the AI was using a firehose. When the firehose was turned down to a teaspoon, the game was even.

Why This Matters

The paper suggests we are entering a new era where the ability to persuade people isn't just about having the best human speaker or the most passionate argument. It's about who has access to the most powerful AI.

The Good News: If small charities or underfunded groups can access this AI, they might finally be able to compete with giant corporations or well-funded political campaigns.
The Bad News: If only powerful actors (like big corporations or governments) have access to the best AI, they could dominate public opinion even more than they do now. They could flood the zone with so many facts and arguments that human voices get drowned out.

The Bottom Line

AI has surpassed humans at the art of conversation and persuasion. It doesn't need to be "more human" to win; it just needs to be able to deliver more information, faster, than any human ever could. Unless we slow the AI down to human speeds, it will likely remain the undisputed champion of changing minds.

Technical Summary: AI Systems Out-Persuade Expert Humans

Problem Statement
Societal decisions in governance, elections, fundraising, and the judicial system often hinge on contests of persuasion. While conversational AI has demonstrated the ability to shift voter preferences, alter moral stances, and induce real-world actions, it remains unclear whether frontier AI systems can out-persuade the most capable, highly incentivized, and professionally trained human advocates. Previous studies suggested AI could match laypeople, but its performance against elite human persuaders—those with professional rhetorical training, extensive preparation, and significant financial incentives—was untested.

Methodology
The authors conducted a series of four preregistered experiments involving 6,923 participants (persuadees) and 18,978 conversations. The study pitted AI systems against five distinct classes of human persuaders:

Random Laypeople: A representative sample of Prolific workers.
Selected Laypeople: Top performers from a separate, preregistered four-round persuasion tournament.
Elite Debaters: World and continental championship debaters with extensive competitive experience.
Coached Elite Debaters: The same elite debaters who received targeted coaching, including access to the AI that defeated them, annotated transcripts, and the ability to see what the AI would have said at key moments.
Professional Canvassers: Experienced fundraisers from a UK firm with years of real-world political canvassing experience.

Experimental Design:

Studies 1 & 2 (Attitude Persuasion): Participants engaged in text-based conversations with either an AI or a human persuader regarding one of 10 pre-specified UK policy stances. Persuasion was measured by the shift in agreement (0–100 scale) from pre- to post-conversation.
- Interventions: Study 2 introduced two conditions to test the limits of AI's advantage: (1) Coaching humans to replicate AI techniques, and (2) Constraining AI to match human throughput (capping message length to ~51 words and response delay to ~92 seconds).
Study 3 (Real-World Expertise): AI competed against professional canvassers on attitude change.
Study 4 (Real-World Behavior): AI competed against professional canvassers to elicit real-money charitable donations to "Save the Children" from a £1 study bonus.
AI Models: The study utilized multiple frontier models, including Claude Opus 4.1/4.6, ChatGPT-4o, GPT-5.4, Grok 4.20, and Gemini 2.5 Pro.

Key Results

Superiority Over All Human Classes: Frontier AI systems were reliably more persuasive than all human classes.
- AI exceeded Random Laypeople by 8.2 percentage points (pp).
- AI exceeded Selected Laypeople (top 10% of a tournament) by 5.6 pp.
- AI exceeded Elite Debaters (world champions) by 4.6 pp.
- AI exceeded Professional Canvassers by 5.9 pp on attitude change.
Robustness of the Gap:
- Coaching Failure: Providing elite debaters with coaching tools (including AI transcripts and "what-if" scenarios) did not significantly improve their persuasiveness relative to their pre-coaching performance. While the gap narrowed slightly, it remained statistically significant (+4.1 pp).
- Individual Variation: No individual human persuader (across 318 estimates) exceeded the pooled AI estimate. The probability of a randomly drawn new human persuader exceeding AI was calculated to be less than 0.1% for all classes.
- Generalization: The AI advantage held across all 10 policy issues and across various demographic and psychological subgroups.
The Source of Advantage (Throughput):
- When AI was constrained to human-like message lengths and response speeds, its advantage over the best human class (Coached Elite Debaters) collapsed to a non-significant 0.0 pp.
- Fact Density: The unconstrained AI deployed significantly more fact-checkable claims per conversation (~~37) compared to humans. A strong correlation ( $R^2 \approx 0.89$ ) was found between the number of facts deployed and persuasive impact. Constraining AI reduced its fact density to human levels (~~12 facts) and eliminated its informational advantage.
Real-World Behavior (Study 4):
- AI was nearly 3x more effective than professional canvassers at eliciting real-money donations.
- AI increased the donation rate by 17.2 pp (vs. control) compared to 6.4 pp for canvassers.
- AI outperformed humans on every measured mechanism driving donations, including "implementation intentions," "commitment escalation," and "impact-efficacy information," despite being explicitly instructed to use only the latter strategy.

Significance and Claims
The paper establishes that frontier AI systems are currently more capable conversational persuaders than the most prepared, incentivized, and expert humans available. The authors claim this capability stems primarily from the AI's ability to rapidly deploy larger quantities of information (fact density) rather than superior rhetorical flair or empathy.

The significance of these findings lies in the potential disruption of persuasion-based societal contests. The authors argue that:

Power Consolidation: Access to the most persuasive AI could allow powerful actors (corporations, states, campaigns) to dominate rivals with stronger arguments or more talented human advocates.
Disinformation and Scams: Bad actors could overcome the historical bottleneck of limited human labor to run highly effective, personalized coercive influence campaigns.
Alignment Risks: In contests between humans and the AI systems they oversee, misaligned AI could use persuasion to influence the judgments of human auditors.

The authors conclude that the question is no longer if AI can out-persuade humans, but how, where, and on whose behalf this capability will be exercised. They note that while this technology could theoretically democratize advocacy for under-resourced groups, the immediate risk is the consolidation of influence among those who control the most capable systems.