AI systems out-persuade expert humans
A new study shows that AI systems can outperform expert humans in persuasive tasks, achieving higher success rates in convincing people to change their viewpoints or take specific actions.
Background
- This paper (June 2025) shows that AI models like GPT-4o and Claude 3.5 Sonnet can now outperform expert human debaters in one-on-one persuasive conversations — meaning they are more likely to change someone's stated opinion on a contentious issue (e.g., vaccines, AI risk, gun control).
- Study design: humans debated other humans, humans debated AIs, and AIs debated AIs. Persuasion was measured by pre- and post-debate attitude surveys. Both leading frontier models beat human experts (including national debate champions) at changing minds.
- This marks a shift from earlier AI persuasion research, which tested static text or simple arguments. The new capability is live, adaptive conversational persuasion — which is closer to real-world influence.
- Why it matters: advanced persuasion ability raises concerns about large-scale manipulation, propaganda, scams, and eroded trust in human discourse. It also challenges assumptions about what tasks remain "uniquely human."
- Frontier models: state-of-the-art AI systems at the cutting edge of capability, typically developed by labs like OpenAI, Anthropic, Google DeepMind, and Meta.