A large-scale study of political discussions on Facebook has found that subtle wording in posts can be linked to a higher share of hostile replies. Analysing tens of thousands of posts about American political groups, researchers say using noun labels such as “Senate Republicans” or sweeping generalisations about a party tends to attract more prejudiced comments than adjective-led phrasing like “Republican senators.”
The researchers argue that such linguistic cues connect with broader patterns of affective polarization, the growing disdain and distrust that supporters of opposing parties feel toward one another. They note that when voters view the other side as an entitative, tightly knit group, hostility can rise as people generalise a single member’s actions to the whole party.
The concept of entitativity—seeing a group as a cohesive entity rather than a loose collection of individuals—was central to the study. Carsten W. Sander, a research associate at Humboldt University of Berlin, explained that this perception can make party members appear interchangeable, prompting generalisations and greater outgroup prejudice. “People then generalize one member’s behavior to the whole group, retaliate against one group member for offenses committed by another, and generally show more outgroup prejudice. We reasoned that this could also fuel political polarization, understood as growing prejudice toward the political outparty,” he said.
The work sits at the intersection of linguistics and political psychology, building on evidence that people judge groups beyond the content of explicit statements. Beyond naming groups, the researchers highlight grammatical structures as subtle signals of uniformity. For example, noun labels like “Senate Republicans” frame individuals by a single category, whereas adjective labels like “Republican senators” treat affiliation as one trait among many.
Lab-based experiments have previously shown that noun versus adjective labels can influence attitudes toward unfamiliar or minority groups. A 2020 study found that portraying an ethnic minority member with a noun rather than an adjective nudged readers toward slightly more prejudice, regardless of whether the article leaned positive or negative.
In this study, Sander and Degner turned to real-world social media to test whether these language cues matter in established political groups, using Facebook posts and comments from the Meta Content Library. They excluded posts that referred to institutions or events rather than people, leaving a dataset of 45,780 posts and 769,684 direct replies written by 391,386 unique users across 3,878 pages and profiles. News outlets, public figures and politicians provided the bulk of contributions.
To process hundreds of thousands of messages, five transformer-based AI models were trained and tested on thousands of human-coded examples, with two models further enhanced by 600 synthetic posts matched in all but one feature. One model classified whether a post used a noun or adjective party label, the other identified generic versus non-generic statements.
The team used multilevel statistical modelling to assess whether the wording of a single post predicted the share of prejudiced replies beneath it. By comparing posts on the same page, they controlled for page-specific differences in audience and content.
Across all posts, noun-label wording correlated with a higher proportion of prejudiced comments than adjective-label wording. Before accounting for controls, the odds of a comment showing partisan prejudice were 1.39 times higher under noun-label posts; this dropped to 1.15 after adjusting for other factors, including whether the target was an individual, a subgroup, or the entire party.
On a typical page where 59% of posts used noun labels, the model predicted 16.7% of comments under noun-label posts would be prejudiced, compared with 14.8% under adjective-label posts. In absolute terms, that’s a 1.9 percentage point difference, or about a 13% relative increase.
The second major finding concerned generic language. Posts making broad generic claims about a group—such as “Democrats want climate action”—drew a higher share of prejudiced replies than non-generic phrasing like “some Democrats want climate action.” In the dataset of 26,026 posts and 437,744 comments focusing on groups rather than individuals, generic posts had a 22.2% predicted rate of prejudice versus 19.2% for non-generic posts, an absolute difference of 3.0 percentage points (roughly 16% relative).
At page level, pages with more generic statements tended to attract somewhat fewer prejudiced comments after controls were applied, a pattern the authors attribute to differences between pages rather than the wording itself.
Importantly, noun labels continued to predict higher hostility independently of generic phrasing, and the patterns held whether posts referred to Democrats or Republicans. Even when comparing different comments by the same user across multiple posts, noun and generic labels remained associated with more prejudiced replies, which the authors interpret as evidence that language itself can shape user responses, not merely attract already hostile readers.
The researchers note the findings align with earlier work on generic statements expanding perceived ideological gaps between parties, though that earlier research relied on surveys, whereas the current study identifies links to actual online behaviour.
There are caveats. The study is observational rather than experimental, so it cannot definitively prove that wording causes prejudice. It relied on automated classifiers to detect prejudiced comments, which correctly identified about 73% of prejudiced remarks, with about 68% of flagged comments actually prejudiced, leaving room for error.
Sander emphasised that the results should not be read as proof of causality. “The key caveat is that the findings are observational,” he explained. “Features of the post content that we didn’t control for could be driving both the labeling and the prejudiced comments, so the results shouldn’t be read as showing that labels cause prejudice.” He added that the broader results are consistent with previous experimental work.
Looking ahead, the authors propose real-world trials to test whether altering post labeling can reduce online hostility. They suggest collaborating with an online news outlet to randomly vary party labeling in social media posts, a design that could help establish causality. They also propose testing whether adjective labels, qualifiers like “some,” or person-first phrasing such as “people who voted Republican” might cool hostilities, or whether describing discrete past actions rather than permanent group traits could temper partisan animosity.
The study, titled Constructing the Enemy: Noun and Generic Party Labels Predict Partisan Prejudice in Social Media Comments, was authored by Carsten W. Sander and Juliane Degner.
