AI Chatbots Are Becoming Powerful Persuaders — What Gives Them the Edge?
Key Points
AI systems can outperform highly trained human persuaders in controlled conversational experiments, including comparisons with elite competitive debaters.
A 2026 study involving 18,978 conversations with 6,923 people found that AI was reliably more persuasive than expert humans.
Researchers found evidence that AI's advantage comes largely from its ability to rapidly deploy larger quantities of information.
When AI was restricted to human-like response lengths and speeds, its persuasive advantage over coached expert humans disappeared.
The growing evidence also highlights a major risk: greater persuasive power does not necessarily mean greater factual accuracy.
advertisement
Artificial intelligence is increasingly becoming more than a tool for answering questions or generating text. New research suggests that conversational AI systems are also becoming highly capable at one of the most consequential forms of human communication: persuasion.
A feature published by Science on August 20, 2026, examines why AI chatbots are becoming so effective at changing people's minds. Recent experiments suggest that the advantage may not come from some mysterious ability to understand human psychology better than people do. Instead, a major part of the explanation may be much more straightforward: AI can gather, organize and deliver large quantities of relevant information at a speed that humans cannot match.
The evidence is particularly striking in a 2026 preprint by Kobi Hackenburg and colleagues, which compared conversational AI systems with several groups of human persuaders. Across four preregistered experiments involving 18,978 conversations with 6,923 people, researchers tested AI against laypeople, winners of a persuasion tournament, professional canvassers and world-championship-level competitive debaters.
The researchers found that AI systems were reliably more persuasive than expert humans, even when the human participants selected their own issues, researched them in advance, underwent hours of structured practice and were offered £1,000 cash bonuses as incentives. A follow-up experiment also found that the advantage remained after expert persuaders received a coaching system that allowed them to practice against the AI, review their performance and see what the AI would have said at important moments.
The researchers then investigated a more fundamental question: Why was AI winning?
Their results pointed toward what they describe as information throughput.
Human persuaders have practical limits. They must research a subject, remember information, formulate an argument and type or speak it to another person. A conversational AI system can rapidly produce many relevant claims and pieces of evidence during the same interaction.
That difference became particularly revealing when the researchers restricted the AI's responses to human-like message lengths and human-like response speeds. Under those conditions, coached expert humans were able to tie the AI's performance. The finding suggests that the machine's advantage was strongly connected to its ability to deploy more information, faster, rather than simply possessing a uniquely sophisticated persuasive style.
The result is important because it changes how researchers may need to think about AI persuasion.
The most powerful ingredient may not necessarily be emotional manipulation or highly personalized psychological targeting. Instead, an AI can potentially overwhelm the limitations of human communication by producing large amounts of relevant, structured and apparently evidence-based information within seconds.
That conclusion is consistent with earlier research published in Science in December 2025.
In that study, researchers examined 19 large language models, 76,977 participants and 707 political issues. They also evaluated the factual accuracy of 466,769 claims generated by the models. The researchers found that post-training methods designed to increase persuasion could boost persuasive performance by as much as 51%, while prompting techniques could increase it by as much as 27%. The effects of personalization and simply increasing model scale were smaller.
The study also revealed an important warning.
The techniques that increased AI's persuasive ability also tended to reduce factual accuracy. In other words, a system can become better at convincing people without becoming better at telling the truth.
That distinction may be one of the most important lessons emerging from AI persuasion research.
People often associate detailed explanations, confidence and large amounts of information with expertise. A chatbot can provide all three almost instantly. But an argument that sounds authoritative can still contain inaccurate information.
Research published in Nature Communications in December 2025 provides another example. The study, titled “Persuading voters using human–artificial intelligence dialogues,” used preregistered experiments examining whether conversations with AI could influence political attitudes. The research included the 2024 U.S. presidential election, the 2025 Canadian federal election and the 2025 Polish presidential election, as well as an experiment involving Massachusetts residents' support for a ballot measure concerning psychedelic legalization.
The researchers found that conversations with AI produced significant changes in candidate preferences and policy attitudes. They also reported that the AI systems appeared to persuade largely through relevant facts and evidence rather than sophisticated psychological persuasion techniques. However, not all of the facts and evidence supplied by the models were accurate.
This combination — high persuasive ability alongside imperfect factual reliability — creates an unusual challenge.
AI persuasion can potentially be used to help people reconsider false beliefs, but the same capability could also be directed toward misleading or unsupported claims.
One frequently discussed example comes from a 2024 Science study by Thomas Costello, Gordon Pennycook and David Rand, which examined personalized conversations between GPT-4 Turbo and people who believed conspiracy theories. The chatbot conversation reduced participants' embrace of their chosen conspiracy theory by almost 17 points on a 100-point scale.
However, there is an important development that must be included in any article published in 2026.
In June 2026, Science issued an Editorial Expression of Concern about that paper. According to the journal, investigators identified inconsistencies between the screening criteria described in the manuscript and the published analysis pipeline, as well as extraneous rows in the public dataset caused by a code-merging error. The authors supplied a corrected analysis pipeline and updated results, and reported that the corrected analysis preserved the original finding's direction, statistical significance and substantive size. Science is still evaluating the corrected analysis.
For that reason, the original study should now be described cautiously. Its reported findings remain scientifically interesting, but its specific numerical results should not be presented as an unquestioned final result while the journal's evaluation remains ongoing.
The broader lesson nevertheless remains important: conversational AI may be capable of tailoring an argument to the specific claims and evidence raised by an individual, potentially making it useful for both correcting misinformation and, if misused, reinforcing it.
The 2026 research provides additional evidence that AI persuasion can extend beyond changes in stated opinions.
In a fundraising experiment, AI was nearly three times as effective as professional canvassers from a UK fundraising firm at raising real-money donations for Save the Children. The researchers describe this as evidence that AI's persuasive advantage can extend to consequential behavior rather than remaining limited to responses in an opinion questionnaire.
That result should still be interpreted within the limits of the experiment.
The researchers studied structured conversational interactions, not every possible real-world situation in which people encounter AI. The finding does not establish that AI will outperform humans in every form of persuasion, nor does it show that AI will automatically determine political outcomes.
But it does demonstrate something increasingly difficult to ignore: conversational AI can influence decisions, not merely generate information.
Another recent study highlights the importance of transparency.
In a preregistered experiment involving 1,500 adults in the United Kingdom, researchers tested whether telling people they were interacting with AI would reduce the chatbot's persuasive effect. Participants discussed one of 60 policy issues with an otherwise identical persuasive chatbot.
Simply disclosing that the participant was interacting with AI had little effect. The control group showed a 12.6-point shift on a 100-point attitude scale, while the AI-identity disclosure group showed a 13.1-point shift. But when participants were told not only that the system was AI but also about its persuasive intent and instructions, the effect fell to 6.3 points — roughly half the size.
That finding suggests that transparency may need to go beyond simply saying “this is an AI.”
People may also need to know what the system is trying to accomplish.
This distinction could become increasingly important as conversational AI spreads into political communication, advertising, customer service, fundraising and other areas in which influencing decisions has direct value.
The research also shows why it would be misleading to describe AI persuasion simply as a story about machines becoming better at manipulating humans.
The evidence is more complicated.
AI can potentially help someone understand an unfamiliar subject, encounter evidence they had not considered or reconsider a false belief. But exactly the same ability to assemble relevant arguments quickly can be used to construct a compelling case for something inaccurate.
That means persuasiveness and truthfulness must be evaluated separately.
A system optimized to maximize persuasion could become more convincing while simultaneously becoming less reliable. The 2025 Science research provides direct evidence of this tension, finding that methods that increased persuasive performance also systematically reduced factual accuracy.
The implications extend beyond politics.
Any activity involving human decisions could potentially be affected: marketing, sales, fundraising, lobbying, education, negotiations and public communication all depend to some degree on persuasion.
The central issue is therefore not simply whether AI can change minds. The evidence increasingly shows that it can.
The more difficult question is how that capability should be used.
If AI systems are allowed to optimize continuously for persuasion, they may become extremely effective at presenting information in ways that influence human decisions. Without safeguards, however, the system's ability to persuade could outpace the user's ability to recognize inaccurate claims or understand why the system is attempting to influence them.
The emerging research points to a relatively simple explanation for part of AI's advantage: machines can communicate information at machine speed.
When that advantage is removed and AI is forced to communicate more like a human, the gap can disappear.
That finding provides an important perspective on the future of persuasion. The challenge may not be that AI possesses an entirely new form of psychological power. Instead, it may be that AI can combine speed, information volume, conversational flexibility and rapid adaptation in a way that human persuaders cannot easily match.
As these systems become more widespread, the most important safeguard may therefore be not simply making AI less persuasive, but ensuring that its persuasive abilities remain connected to accurate information, transparent intentions and meaningful human control.
Key Points Summary
AI systems can outperform expert human persuaders in controlled conversational experiments.
A 2026 study involving 18,978 conversations and 6,923 people found AI reliably more persuasive than several groups of human persuaders.
The strongest evidence points to information throughput as a major part of AI's advantage.
Restricting AI to human-like response length and speed removed its advantage over coached expert humans.
Research also shows a major warning: increased AI persuasiveness can come with reduced factual accuracy.
What This Means
Why it matters: AI is increasingly capable of influencing not only what people read, but also what they believe and what they choose to do.
Who may be affected: Voters, consumers, donors, students, social-media users and anyone interacting with conversational AI could encounter increasingly persuasive machine-generated arguments.
What to watch next: Researchers and policymakers will need to pay close attention to accuracy, disclosure of AI involvement, disclosure of persuasive intent, transparency and safeguards against manipulation.
advertisement
Frequently Asked Questions
Can AI chatbots really change people's minds?
Yes. Multiple controlled studies have found that conversations with AI can change attitudes and preferences. The magnitude of the effect varies depending on the system, topic and experimental design.
Why can AI be more persuasive than humans?
Recent research points strongly toward information throughput. AI can rapidly introduce many relevant pieces of information and fact-checkable claims during a conversation. When researchers restricted AI to human-like response lengths and speeds, its advantage over coached expert humans disappeared.
Does AI persuasion mainly depend on personalization?
Not necessarily. The 2025 Science study found that post-training and prompting had larger effects on persuasion than personalization or simply increasing model scale.
Does a persuasive AI always provide accurate information?
No. This is one of the central concerns in the research. The 2025 Science study found that methods that increased AI persuasion also systematically reduced factual accuracy.
Can AI persuasion influence real-world behavior?
There is evidence that it can. In the 2026 fundraising experiment, AI was nearly three times as effective as professional UK canvassers at raising real-money donations for Save the Children.
Does telling people they are talking to AI stop persuasion?
Not necessarily. A 2026 study of 1,500 UK adults found that AI-identity disclosure alone had little effect, while disclosure of both AI identity and persuasive intent roughly halved the persuasive effect.
What happened to the 2024 study about AI reducing conspiracy beliefs?
The study reported substantial reductions in conspiracy belief following personalized GPT-4 Turbo conversations. However, Science issued an Editorial Expression of Concern in June 2026 over data-handling and reproducibility issues. The authors have supplied a corrected analysis, which they say preserves the original result's direction, significance and substantive size, but Science is still evaluating it.
Does being persuasive mean an AI is intelligent or correct?
No. Persuasiveness and factual accuracy are separate properties. A system can produce an extremely convincing argument while still making inaccurate claims.
Sources
Science — Kai Kupferschmidt, “Powers of persuasion: AI chatbots are becoming experts at changing people’s minds. What gives them an edge?” — August 20, 2026.
Read the Science feature
Additional Verified Sources
Hackenburg et al. — “AI systems out-persuade expert humans” — 2026 preprint.
Read the research paper on arXivHackenburg et al. — “The levers of political persuasion with conversational artificial intelligence” — Science, 2025.
Read the study on ScienceLin et al. — “Persuading voters using human–artificial intelligence dialogues” — Nature, 2025.
Read the Nature studyCostello, Pennycook & Rand — “Durably reducing conspiracy beliefs through dialogues with AI” — Science, 2024.
Read the study on PubMedScience — Editorial Expression of Concern regarding the 2024 conspiracy-belief study — June 11, 2026.
Read the editorial noticeRauchfleisch & Jungherr — “Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion” — 2026.
Read the research paper on arXiv
Thank you !
