AI Chatbots Sway Voters More Effectively Than Political Ads

A collaboration between the AI Security Institute and academics from Cornell, Oxford, Stanford, MIT, and the London School of Economics found that AI chatbots are more effective at swaying political opinions than traditional political advertisements. Studies published in Nature and Science involved participants in the United Kingdom, United States, Canada, and Poland interacting with models from OpenAI, Meta, Google, and xAI.

In the U.S., chatbots shifted Donald Trump supporters toward Kamala Harris by 3.9 points on a 100-point scale, an effect roughly four times greater than previous television ads. Pro-Harris bots moved 1 in 21 opposition voters toward her, while pro-Trump bots moved 1 in 35. Even larger shifts occurred in Canada and Poland, where approximately one in 10 respondents shifted their voting intentions. In the UK, the persuasive effects were durable, with 36% to 42% remaining after one month.

Researchers discovered that the most persuasive bots were those providing high volumes of information and evidence-backed arguments, regardless of their truthfulness. Approximately 19% of claims in the UK study were predominantly inaccurate, and chatbots advocating for right-leaning candidates generally made more inaccurate claims than those supporting left-leaning candidates. Newer, larger models like GPT-4.5 showed a decline in accuracy compared to older versions. Experts warn these dynamics could be exploited to promote radical ideologies or compromise independent political judgment in future elections.