The Partisan Persona Becomes the Persuasion Test
The June 2026 arXiv paper Political Persuasion and Endorsement in Large Language Models, by Alessia Antelmi, Alessia Galdeman, Lucio La Cava, Arianna Pera, and Giovanni Da San Martino, tests whether partisan persona prompts change how LLMs endorse persuasion-infused political content.
The Proxy Problem
The paper, arXiv:2606.05961 [cs.CY], was submitted on June 4, 2026. Its starting point is a problem already visible across computational social science: large language models are being used as stand-ins for human attitudes, but a model that can simulate a persona may also inherit political and rhetorical patterns from training and post-training.
That makes the paper a useful companion to this site's pages on synthetic respondents, LLM social-network simulations, and personality sliders. The question is not whether an LLM has a private political essence. It does not need one. The question is whether a deployed system's prompted role changes what it treats as endorseable in a politically charged setting.
The Test
Antelmi and coauthors evaluate six open-weight, instruction-tuned models from the Hugging Face Model Hub: Llama-3.1-8B-Instruct, Phi-3.5-mini-instruct, Mistral-7B-Instruct-v0.3, Qwen2.5-7B-Instruct, Yi-1.5-9B-Chat, and aya-expanse-8b. The paper groups them as models from different geographic development regions, then tests whether that provenance explains endorsement patterns.
The study uses two data sources. The first is a dataset of tweets collected around the beginning of the Ukraine-Russia conflict, with 29,596 tweets total and labels for persuasion techniques. After filtering to four common techniques and balancing with neutral tweets, the authors use 7,730 tweets. The second is the English portion of the SemEval-2023 Task 3 news corpus. After merging train and dev splits and filtering to single-technique spans for comparability, they use 3,532 news spans.
Each model is asked to rate how likely it would be to endorse a piece of content on a five-point Likert scale. The baseline prompt asks it to behave as a typical social media user. The partisan conditions change that user into a left-leaning or right-leaning social media user. The output is constrained to a single number from one to five, which makes the measurement narrow but reproducible.
What Shifted
The central result is disciplined rather than sensational. Under neutral persona prompting, the models generally do not strongly endorse persuasion-infused content. In the tweet dataset, persuasion is associated with a large drop from the neutral no-persuasion baseline, with the neutral persuasion condition near the low end of the scale. The authors describe Mistral and Qwen as especially cautious across both datasets, while Phi has a higher baseline endorsement score for news.
Partisan persona prompting changes the pattern. In the tweet dataset, left-leaning personas suppress endorsement while right-leaning personas amplify it, especially when the content contains persuasion techniques. The news dataset shows the same direction more weakly: right-leaning personas still score higher than the neutral baseline, but the left-leaning condition is less cleanly separated.
The study also finds that endorsement varies by persuasion technique and topic. Name calling and labeling receive lower endorsement scores across partisan conditions in the tweet dataset and lower endorsement in the news dataset under neutral and left-leaning conditions. Loaded language and appeal to fear or prejudice generally receive somewhat higher scores, especially for tweets. Topic matters too: the news corpus shows more heterogeneous endorsement behavior than the conflict-centered tweet corpus.
Persona Risk
The governance lesson is not that one side is uniquely vulnerable. The paper's useful warning is about persona-conditioned systems. If a civic assistant, forum bot, research simulator, campaign tool, or community agent is told to speak as a political type, the role instruction can reorganize its apparent endorsement behavior. The system may remain fluent and cooperative while its response pattern becomes less stable for political evidence.
That matters for two uses. First, LLMs used as simulated publics can overstate confidence if a simple persona prompt shifts endorsement of persuasion-infused content. Second, public-facing agents can become persuasion amplifiers without generating propaganda themselves. Endorsing, ranking, summarizing, recommending, or "agreeing with" content is enough to move the interface.
The paper is careful about scope. It studies endorsement, not generation; single-turn numeric ratings, not extended conversation; English-only content, not multilingual politics; and simplified left/right personas, not the full structure of political identity. Those limits make the finding more useful, not less. They mark where future safety evaluations should expand before institutions trust persona agents in political settings.
Limits That Matter
The study should not be overread. It does not prove how all current frontier models behave, how people respond to those models, or how endorsement unfolds in a live platform. It does not show that model geographic origin determines endorsement behavior; the authors explicitly report that individual model differences do not map cleanly onto that origin category.
It also should not be treated as a recipe for persuasion. The paper uses previously published datasets, reports aggregate findings, and frames the work as diagnostic. A governance reading should preserve that boundary. The point is to measure susceptibility and simulation reliability, not to optimize political manipulation.
The hardest missing layer is interaction history. A single numeric rating is useful for controlled measurement, but real political interfaces involve memory, repeated exposure, recommendation loops, user adaptation, and community identity. A model that refuses to endorse a slogan once may still normalize a framing through summary, ranking, or repeated polite agreement over time.
Governance Standard
Any politically situated LLM deployment should name its persona policy. That means recording whether the system uses political, demographic, community, brand, therapeutic, religious, or ideological role prompts; what evaluation set was used; which content types were tested; and whether endorsement-like actions were measured separately from generation.
Political evaluations should not ask only whether the model produces false claims or banned content. They should test endorsement, agreement, ranking, summarization tone, refusal asymmetry, topic sensitivity, and role sensitivity. A persona that looks harmless in product copy can become an unlogged policy layer if it changes which claims the interface treats as plausible, urgent, or socially approved.
The practical rule is simple: do not deploy a political or community persona unless the institution can show how that persona changes endorsement behavior. If the answer is unknown, the persona is not a style setting. It is an untested persuasion condition.
Sources
- Alessia Antelmi, Alessia Galdeman, Lucio La Cava, Arianna Pera, and Giovanni Da San Martino, Political Persuasion and Endorsement in Large Language Models, arXiv:2606.05961 [cs.CY], submitted June 4, 2026.
- arXiv PDF for Political Persuasion and Endorsement in Large Language Models, reviewed June 24, 2026.
- Related pages: The Synthetic Respondent Becomes the Public, The LLM Social Network Becomes the Polarization Lab, The Personality Slider Becomes the Belief Interface, The Persuasion Engine Gets a Memory, The Ad Library Becomes Political Memory, and Network Propaganda and the Media Feedback Machine.