AI systems out-persuade expert humans

Abstract: Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has remained unclear. Here, in a series of four preregistered experiments (n = 18,978 conversations from 6,923 people), we pitted AI systems against a range of human persuaders, including laypeople, winners of a separately preregistered four-round online persuasion tournament, professional canvassers, and world championship debaters. 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. In a follow-up study, AI's advantage persisted after experts received a coaching tool that let them practice against the AI that beat them, review their performance history, and see what AI would have said at key moments. 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. In a final study, we show that AI's advantage extends to consequential real-world behavior: AI was nearly 3x more effective than professional canvassers from a UK fundraising firm at raising real-money donations to Save the Children. Together, these results establish that frontier AI systems out-persuade expert humans in conversation, with significant implications for political communication.

• The paper demonstrates that advanced AI systems outperform expert human persuaders in both attitudinal shifts and donation behavior, with improvements up to +8.2 percentage points.
• The paper attributes AI’s persuasive advantage to its high throughput and fact density, which nearly accounts for the observed impact (R² = 0.89).
• The paper discusses societal implications, highlighting benefits for democratized advocacy alongside risks such as disinformation and automated coercion.

Summary of "AI systems out-persuade expert humans" (2606.16475)

Experimental Design and Scope

The paper conducts four preregistered online experiments comprising 18,978 conversations with 6,923 participants to rigorously compare the persuasive efficacy of advanced conversational AI and expert human persuaders. Human comparators include random laypersons, tournament-selected laypersons, elite debaters (world and continental champions), professional canvassers, and coached elite debaters. Multiple state-of-the-art LLMs (Claude Opus 4.1/4.6, ChatGPT-4o, GPT-5.4, Grok 4.20, Gemini 2.5) are deployed using empirically validated prompting strategies targeting attitudinal and behavioral shifts. Incentives and preparation are maximized for human persuaders, including extensive issue research, live practice, and substantial cash bonuses. The evaluation encompasses both attitudinal persuasion (policy stance shifts) and consequential real-world behaviors (charitable giving).

Quantitative Outcomes and Key Claims

AI systems consistently outperform every human class in both attitude change and behavioral persuasion. Specifically:

• Against random laypersons, AI yields an additional +8.2 percentage point (pp) shift.
• Against tournament-selected laypersons, AI achieves +5.6 pp.
• Against elite debaters, AI surpasses by +4.6 pp.
• Against professional canvassers, AI is +5.9 pp more persuasive in attitude shift and nearly 3x more effective in eliciting real-money donations (+10.8 pp; AI: +17.2 pp vs. canvassers: +6.4 pp).
• Coaching elite debaters using a tailored AI tool increases their message length and fact density but only marginally narrows the gap (+1.0 pp), failing to achieve parity.

Robustness checks confirm AI superiority at finer granularity: no individual human persuader outperformed pooled AI, and AI's advantage persisted in every policy issue and across demographic/political subgroups.

Analysis of Persuasive Mechanisms

The authors attribute AI's advantage primarily to throughput—AI delivers far more fact-checkable information per conversational turn due to higher word count and rapid response. When AI is throttled to human message length and speed, its advantage vanishes (from +4.1 pp to statistically insignificant 0.0 pp). Regression analyses show that fact density nearly fully accounts for persuasive effect (overall $R^2 = 0.89$). Perceived informational strength—the degree to which the partner's arguments are strong and the conversation is educational—drops precipitously when AI throughput is constrained, whereas affective dimensions (empathy, enjoyment) show much smaller declines.

Real-World Generalization and Behavioral Impact

The paper extends analysis beyond laboratory persuasion to real-world behavioral outcomes, with AI eliciting not only substantial attitude shifts but also higher rates and quantities of charitable giving. AI is rated superior by participants on all seven hypothesized causal mechanisms for donation, including implementation intentions, commitment escalation, and impact-efficacy information, despite being prompted for only one strategy. This suggests effective prompt generalization and flexible resource deployment by frontier models.

Professional canvassers—selected from a firm with extensive experience fundraising for Save the Children—are outperformed in both donation frequency and amount, even when given optimal issue selection and preparation windows.

Theoretical and Practical Implications

The findings recast the landscape of political and commercial persuasion, indicating that contests may increasingly be determined by access to frontier LLMs rather than traditional human expertise. This raises the prospect of consolidation of influence—either to actors who possess or deploy the most capable models (e.g., corporations, campaigns, states) or to model vendors themselves, with the power to bias which arguments are advanced.

Potential societal benefits include democratization of advocacy, enabling under-resourced actors to access skilled rhetorical tools. However, outcomes are critically contingent on the truthfulness and selectivity of AI information compared to the persuasion it displaces. Risks include amplification of disinformation, scaling of coercive influence operations, and susceptibility of human auditors/judges to AI manipulation. Self-limiting factors may include platform access controls, public awareness (though explicit labeling as AI-generated does not reduce persuasive impact), and real-world exposure constraints.

Future Research Directions

Three domains are identified for further inquiry:

1. Modality sensitivity—extension to audio/video and face-to-face settings, where rapport and embodied empathy may be more salient.
2. High-stakes persuasion—evaluation of AI efficacy in materially significant outcomes (e.g., voting, major donations).
3. Displacement effects—quantifying how AI-driven persuasive exposure alters net societal impact compared to current baseline exposure.

Conclusion

The paper establishes that advanced LLMs, prompted for high information-density, reliably out-persuade even maximally incentivized expert humans across both attitudinal shifts and real-world actions. The throughput advantage—rapid delivery of voluminous factual claims—is the dominant explanatory variable. Human improvement via direct coaching on AI strategies is insufficient to close the gap. As access to persuasive AI proliferates, fundamental questions arise regarding deployment control, fairness, and the societal consequences of mass automated persuasion.