Feed 0% source
Social science AI-generated

AI systems out-persuade expert humans

Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.

Many societal decisions are settled by contests of persuasion. From elections and legislative debates to charitable fundraising, the ability to shift public opinion determines how resources and power are distributed. Researchers have long known that conversational AI can influence people, but it remained unclear whether these systems could outperform the most skilled and highly incentivized humans.

A new study from the University of Oxford and the UK AI Security Institute reveals a disruptive reality. Frontier AI systems do more than just compete; they reliably out-persuade expert humans. The researchers found that advanced AI can change people's minds more effectively than professional debaters and fundraisers. Crucially, this advantage does not stem from superior empathy or rapport. Instead, it comes from the AI's ability to provide a much higher volume of information very quickly.

The limits of human expertise

Historically, the gold standard for persuasion has been the highly trained human specialist. In political campaigns, this might be a professional canvasser. In legal or academic arenas, it is the elite debater. These individuals possess deep subject knowledge and undergo rigorous training. They are often driven by significant financial incentives to succeed.

Previous research suggested that while AI could match the persuasiveness of average laypeople, it struggled to compete with professionals. There was a prevailing assumption that human nuance and strategic reasoning would create a ceiling that AI could not breach. This study challenges that status quo. The researchers pitted frontier models—including Claude Opus 4.1 and 4.6, ChatGPT-4o, GPT-5.4, Grok 4.20, and Gemini 2.5 Pro—against humans who were given every possible advantage. These humans received large cash bonuses, hours of research time, and intensive coaching. As shown in, the AI maintained a consistent lead over every human class tested.

Figure 1
Figure 1: Frontier AI systems out-persuade expert humans at attitudinal persuasion. Pooled estimated persuasive impact on political attitudes (percentage points; 95% Wald CIs) of each human persuader class (gray), AI (red, three per-study estimates), and AI constrained to respond with human-length messages at human writing speeds (salmon; Study 2 only) relative to an active control (a chat with ChatGPT-4o about a non-political topic). Estimates from a linear mixed-effects model pooling Studies 1-3, with random intercepts for persuader and persuadee and pre-treatment attitude and issue as fixed-effect covariates ( Methods ; robustness to adding a study fi xed effect in SI Appendix, Section 4.12). Marker shape denotes study; the AI column overlays per-model estimates (Claude Opus 4.1 and 4.6, ChatGPT-4o, GPT-5.4, Grok 4.20, Gemini 2.5 Pro) on the pooled AI estimate.

This included everything from random workers to world championship debaters.

Information throughput as a lever

The authors propose that the secret to AI's success is not "wisdom." Rather, it is a massive advantage in information throughput (the rate at which a system delivers content). To test this, the researchers designed experiments that manipulated the speed and volume of the AI's output.

The core mechanism being investigated is "fact density" (the number of fact-checkable claims deployed during a conversation). The researchers hypothesize that AI uses its near-instantaneous latency (the delay between a user's prompt and the AI's response) to overwhelm interlocutors. It does this with a dense stream of evidence.

To isolate this effect, the study employed a "Constrained AI" condition. In this setup, the researchers throttled the AI. They forced it to adhere to human-like constraints. This included limiting its message length to approximately 51 words. They also introduced a simulated delay of roughly 92 seconds between responses. This forced the AI to operate within the "bandwidth" of a human speaker.

Evidence from the persuasion gap

The results of these interventions provide a clear look at what drives the AI's edge. The authors report that when the AI was unconstrained, it outperformed elite debaters by 4.6 percentage points . This means the AI shifted attitudes significantly more than the experts. However, when the AI was subjected to throughput constraints, its advantage collapsed. Its lead over coached elite debaters dropped to a non-significant 0.0 percentage points .

Figure 2
Figure 2: Coaching elite human debaters did not close AI's advantage, and no individual human persuader exceeded AI; only constraining AI to human throughput closed the gap. (a) Estimated persuasive advantage of AI (pp; 95% Wald CIs) relative to each comparator within the comparator's own study (Elite Debater within Study 1, marked by a circle; Coached Elite Debater and Constrained AI within Study 2, marked by triangles). Same LMM as in Fig. 1, refit with AI entered separately by study. AI footer row pools its words/message and response delay across Studies 1-2 (essentially identical in both). Inset columns: realized words/message and between-message delay (s). The only condition that closes AI's gap over the strongest human class is throttling AI's throughput to human levels (third row). (b) Implied normal distributions of per-persuader treatment effects within each human class (Studies 1-2; Professional Canvassers in SI Fig. S5), parameterized by class mean µ and between-persuader SD ˆ τ from a class-specific random-effects model. Dashed red line: pooled AI estimate ( µ = 13 . 9 pp, mean of the per-study AI-control contrasts in Studies 1 and 2). For Elite Debaters the REML fit was singular (ˆ τ = 0); we substituted the pooled cross-class SD (ˆ τ pooled = 0 . 7 pp) as a conservative fallback.

The data suggests the advantage is almost entirely structural. The paper finds that the constraint primarily suppressed the "informational" aspects of the conversation. Specifically, the perceived strength of the partner's arguments dropped by approximately 11.8 percentage points .

Figure 3
Figure 3: Converging evidence suggests that AI's persuasive advantage is associated with the volume of information delivered per conversation. (a) Effect of constraining AI on persuadees' post-conversation partner ratings (seven items; Methods ); the constraint selectively suppresses the two informational items (argument strength, learning), consistent with a fact-density mechanism. (b) Persuasive impact vs. mean fact-checkable claims per conversation (log scale), pooled across human and AI conditions in Studies 1 and 2 (Study 3 conditions in SI Fig. S6). Dashed line: overall OLS fit; R 2 for overall / within-humans / within-AI fits annotated inline. Shape: study; color: condition type.

The amount of information the persuadee felt they learned also dropped by 11.8 percentage points. Interestingly, factors like "feeling understood" or "enjoyment" moved much less. This suggests the AI wasn't winning by being "nicer," but by being more informative.

This advantage translates directly to real-world utility. In a study involving professional canvassers from a UK fundraising firm, the AI was nearly three times more effective at eliciting real-money donations to Save the Children than the human professionals .

Figure 4
Figure 4: AI elicited more real-money charitable donations than professional canvassers, and was rated higher on every donation-relevant mechanism we measured. (a) Adjusted treatment effect on donations to Save the Children (percentage points of a £1 study bonus, vs. a non-political control conversation; 1 pp corresponds to one additional penny donated per persuadee on average) for Professional Canvassers (recruited from AppcoUK, a UK firm with seven years of Save the Children fundraising experience; Methods ) and AI (Claude Opus 4.6), from a linear mixed-effects model with a random intercept for persuader, controlling for pre-treatment organization support, pre-treatment donation willingness, age, and ideology. (b) AI-Canvasser difference on each of seven preregistered donation-persuasion mechanisms [13], computed as the per-persuadee mean of two self-report items per mechanism (filled diamonds, Welch's t 95% CIs); faint open diamonds show the two constituent items per row. Per-item estimates and the full mechanism + partner-perception batteries are provided in SI Appendix, Sections 4.24 and 4.25.

The AI outperformed the canvassers on every measured psychological mechanism of giving. This included "implementation intentions" (the degree to which a person plans how to act on a belief) and "impact-efficacy" (the belief that a specific action will lead to a measurable result).

Constraints and unanswered questions

While the findings are striking, the authors note several important boundaries. First, the experiments were entirely text-based. The dynamics of persuasion change in audio, video, or face-to-face settings. In those modes, embodied empathy (non-verbal cues like eye contact) might play a larger role. It is possible that humans retain a competitive edge where "presence" matters more than "throughput."

Second, the behavioral study focused on a relatively low-stakes transaction: a £1 donation. The paper does not explore whether this gap persists in high-stakes environments. Examples include deciding a national election or complying with complex public health mandates. Finally, the study was conducted in a controlled, paid-survey context. The authors acknowledge that replicating 14-minute, high-engagement text conversations in the wild may be difficult.

The verdict on algorithmic advocacy

The evidence suggests we are entering an era of "surplus advocacy." Highly effective persuasion is becoming a cheap, scalable commodity. Because the AI's advantage is tied to information density rather than human-like warmth, the primary challenge for society may not be "detecting bots." Instead, the challenge may be managing the sheer volume of high-density, potentially polarizing information they can generate.

If you are building or deploying conversational agents, the takeaway is clear. Information density is a potent driver of influence. For those interested in the methodology or the raw data, the authors have made their code and datasets available. See the paper for the canonical links to GitHub and Zenodo. Whether this technology democratizes influence or consolidates power remains the critical question for the next decade.

Novelty
0.0/10
Overall
0.0/10
#research#artificial intelligence#persuasion#political communication#large language models
How this was made
Generation

Model: nvidia/Gemma-4-26B-A4B-NVFP4
Persona: academic_accessible
Template: engineering_deepdive
Refinement: 0
Pipeline: forge-1.1

Verification

Evaluator: nvidia/Gemma-4-26B-A4B-NVFP4
Score: 94% (passed)
Claims verified: 15 / 15

Translation

Model: nvidia/Gemma-4-26B-A4B-NVFP4

Hardware & cost

NVIDIA GB10 · 128 GB unified · NVFP4 · 100% local · $0 cloud
Tokens: 93,970
Wall-time: 366.1s
Tokens/s: 256.7

Related
Next up

AI advice suppresses human willingness to say 'I don't know', even when advic...

8.2/10· 5 min