Study: LLMs endorse political persuasion when given partisan personas
AI models from six global regions show concerning susceptibility to manipulative political language
A new study published on arXiv (2606.05961) by researchers including Alessia Antelmi and Giovanni Da San Martino examines whether large language models (LLMs) are susceptible to political persuasion. The team evaluated six models from different geographic regions, presenting them with messages annotated with real-world persuasion techniques (e.g., emotional appeals, logical fallacies). The models were prompted as either neutral social media users or as users with left- or right-leaning political views. Each message was scored on a 5-point Likert scale for endorsement likelihood.
The results reveal that without partisan conditioning, LLMs generally do not endorse persuasion-infused content—though model-level differences exist. However, when given a partisan persona, models significantly increased endorsement of politically aligned persuasive messages, revealing a polarization effect. Endorsement rates varied widely by persuasion technique and topic. The authors caution that these biases complicate the use of LLMs as reliable proxies for human political cognition and raise red flags for agentic deployments in politically sensitive environments.
- Six LLMs from different geographic regions tested on persuasion-endorsement using annotated real-world media messages
- Without partisan prompting, models mostly reject persuasive content; with left/right personas, polarization increases endorsement
- Endorsement varies by persuasion technique and topic, undermining LLMs as unbiased human simulators in political science
Why It Matters
LLMs' political bias under persona prompting threatens their reliability in social science simulations and sensitive real-world deployments.