AI chatbots outpersuaded elite human debaters by producing more claims faster

A preprint covered by Science placed more than 2,000 people in political debates with other people or leading chatbots. ChatGPT, Gemini, and Claude consistently changed opinions more than laypeople and a paid group of 56 elite debaters, including world champions. The models' advantage was not a mysterious new form of wisdom. Persuasion rose with the number of fact-checkable claims, and forcing AI to write human-length messages at human speed brought its performance down to roughly human levels. That mechanism should alarm anyone building political, commercial, or therapeutic chatbots: claim volume can look like evidence even when the facts are weak or false. The researchers also found professional fundraisers were less effective than a persuasive bot at increasing donations in the study. These are controlled experiments with paid participants, not proof of mass persuasion in the wild, but they expose a scalable asymmetry between the speed of assertion and the time humans need to verify it.

AIImpactLab story signal

Analysis byAIImpactLab Editorial Desk

PublishedAugust 21, 2026

Reading time6 minutes

The machines beat more than ordinary participants

The study placed more than 2,000 participants in text debates about policies ranging from a teenage social-media ban to assisted suicide. Leading models consistently shifted views more than human counterparts. The researchers also recruited 56 elite debaters, including world champions, paid them, and tied bonuses to persuasiveness. The models still performed better.

An AI coaching tool improved human performance slightly after eight hours of use, but did not close the gap. In a separate charity test within the project, persuasive AI also generated more donations than professional canvassers who had worked for the organization.

Speed and claim volume explain much of the advantage

The number of fact-checkable claims in a conversation predicted how persuasive it was. Models could write longer answers containing more claims in far less time than a person. When the researchers constrained the systems to human message length and writing speed, the AI advantage fell to roughly human levels.

That result also reveals the failure mode. A listener may experience a dense stream of specific claims as rigor even when verification is impossible in real time. Science notes related findings that models optimized for persuasion can become less truthful, and that inaccurate claims appeared in at least one reported conversation.

Controlled persuasion is not mass persuasion

The experiments paid people to engage and focus on the exchange. Real information environments compete for attention, and many people may ignore an unsolicited chatbot entirely. The work therefore does not establish that a political actor can move an electorate by deploying millions of bots.

It does establish a capability that can be personalized and repeated at low cost. High-stakes deployments should separate fact retrieval from persuasive optimization, expose sponsored or strategic intent, audit truthfulness, limit interaction speed where appropriate, and let independent researchers test the system before evidence-shaped output becomes invisible influence.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

Science — AI chatbots are becoming experts at changing people's minds

The machines beat more than ordinary participants

Speed and claim volume explain much of the advantage

Controlled persuasion is not mass persuasion

Go to the source

Related impact

Chatbots rebut election lies but can still fabricate fraud and miss their own deepfakes

AI writing assistants preserve content while flattening the human signals inside language

AI-exposed young workers face a 19 percent employment gap driven by weaker hiring