In a recent paper published in Nature Communications, Robb Willer, former graduate student Jan Voelkel, and colleagues ask whether large language models (LLMs) can generate persuasive policy messages that measurably shift public opinion.

Across three pre-registered online experiments with 4,829 U.S. participants (fielded in late 2022), respondents reported their support for a policy, read a short message, and then reported support again. The messages were either generated by GPT-3/GPT-3.5, written by lay humans, or (in some conditions) selected by humans from a set of LLM messages. Control groups read neutral, unrelated text. The policies ranged from less polarized (a public smoking ban) to highly polarized (an assault weapons ban), and Study 3 tested multiple issues (carbon tax, child tax credit, paid parental leave, automatic voter registration).

Across studies, exposure to LLM-generated persuasive messages produced small but reliable attitude changes (typically about 2–4 points on a 0–100 support scale) relative to controls. Importantly, the LLM messages were about as effective as messages written by lay people, suggesting LLMs have reached an “everyday human” benchmark for persuasion. Exploratory analyses indicate effects were larger among Democrats and among participants already somewhat supportive of the policy, aligning with prior work on the role of identity and prior attitudes in persuasion.

The substantive implication is not that effects are huge, but that persuasive political messaging can now be produced quickly, cheaply, and at scale, raising new concerns about mass influence, impersonation, and the regulation and governance of AI-generated political communication.