Author says persuasiveness in AI evaluations exceeds humans, with no evidence for real-world political persuasion
Where is all the AI persuasion?
Although multiple studies show that frontier models' persuasiveness has surpassed human experts, the author Karthik Tadepalli, after reviewing real-world evidence, believes that AI persuasiveness has not taken effect in high-stakes scenarios such as elections and social media. On the evaluation side, a June 2026 study by UK AISI had frontier models compared with top debaters and professional canvassers; in nearly 20,000 policy conversations, AI shifted opinions by about 5 percentage points more; another study showed that AI's effect in persuading voters before the 2024-25 elections in the United States, Canada and Poland was several times that of typical political advertising. On the real-world side, AI advertising spending in the 2026 U.S. midterm elections was at least $187 million, and campaigns themselves admitted the effect was limited; one campaign team's chatbot received responses from only 1.7% of targeted voters, and the five influence operations recorded in OpenAI's first report all scored no more than 2 on the six-point Breakout Scale.