AI Political Chatbots Can Sway Voters, New Research Finds
Recent findings by Cornell University and the UK AI Security Institute indicate that AI systems can shift voter preferences by up to 15% in controlled experimental settings. These results were published in the journals Science and Nature, based on pre-registered experiments assessing how large language models (LLMs) influence voter attitudes.
The Nature study involved nearly 6,000 participants across the US, Canada, and Poland, including a US sample of 2,300 people ahead of the 2024 presidential election. Chatbots supporting a candidate reinforced participants' existing preferences, producing larger shifts when supporting the opposing candidate. Policy-focused messages from chatbots had stronger persuasive effects than personality-based messages. However, accuracy varied across conversations, with chatbots backing right-leaning candidates delivering more inaccuracies than those supporting left-leaning ones.
Complementing this, the Science study examined responses from 76,977 UK adults on more than 700 political issues using 19 language models. It found that prompting techniques influenced persuasion more than model size. Prompts encouraging new information increased persuasion but reduced accuracy, with providing new information identified as the most persuasive factor.
Additionally, a survey by the Heartland Institute and Rasmussen Reports found that younger conservatives (ages 18–39) were more willing than liberals to grant AI authority over major government decisions. Conservatives showed the highest support for AI-assisted policy-making, constitutional interpretation, and military commands. Donald Kendal of the Heartland Institute cautioned that AI models are biased and not neutral, warning that trust in AI may be misplaced due to corporate decisions shaping content distribution across platforms.
These studies highlight concerns that AI-driven political persuasion could influence elections and democratic processes. Researchers emphasize the importance of pre-registered experiments to rigorously assess how LLMs affect voter attitudes.
