'Entitativity': How a Single Word Choice Fuels Partisan Bias in Political Conversations
Key Takeaways
- •Researchers Carsten W. Sander and Juliane Degner used fine-tuned transformer models to automatically classify partisan prejudice across roughly 46,000 Facebook posts and approximately 770,000 comments.
- •Posts containing noun labels, such as 'Senate Republicans,' received higher proportions of prejudiced comments than posts using adjective labels, such as 'Republican senators.'
- •Generic phrasing, such as 'Democrats want,' was associated with more prejudiced replies than non-generic phrasing, such as 'some Democrats want.'
- •The findings extend the concept of entitativity, first introduced by Donald Campbell in 1958, from earlier experiments on fictional groups to established real-world political groups.
- •Because the dataset came entirely from Facebook, the authors note that whether the same grammatical pattern applies on other platforms remains an open question for future research.
The French philosopher Michel Foucault argued that language does not neutrally describe reality — it actively shapes it. A recently published scientific study of political language offers empirical support for that thesis, finding that subtle choices in wording can substantially influence the level of partisan bias in online discussions of sensitive political issues.
At the center of the findings is a single psychological concept: entitativity — the degree to which a group of people is perceived as a cohesive, unified and “real” entity. The concept has a long history in social psychology, dating back to Donald Campbell’s introduction of the term in 1958 to capture what makes a collection of individuals feel like a single, coherent group rather than a loose assembly of separate people.
“Research on novel groups suggests that perceived entitativity is driven by how others label these groups,” Carsten W. Sander and Juliane Degner recently wrote in the journal Social Psychological and Personality Science. “Using fine-tuned transformer models to automatically classify Facebook posts (N = 45,780) and comments (N = 769,684), we show that this effect also holds for established political groups. Specifically, posts containing noun labels (e.g., “Senate Republicans”) rather than adjective labels (e.g., “Republican senators”) and using these labels generically (e.g., “Democrats want”) rather than non-generically (e.g., “some Democrats want”) receive higher proportions of comments expressing partisan prejudice.”
Earlier research on the topic relied on fictional political groups that participants had never encountered. In those experiments, phrasing that subtly encouraged entitativity led participants to view the groups as single units, while phrasing that did not led them to see the members as separate individuals. To test whether the pattern extended to real-world political discourse, the researchers analyzed roughly 46,000 Facebook posts and approximately 770,000 comments. That is exactly what they found. The fine-tuned transformer models cited in the paper belong to the same family of AI language models that power modern text-analysis tools, and they made it possible to classify prejudice automatically across a comment volume that would be impractical to code by hand. Because the dataset was drawn entirely from Facebook, whether the same grammatical pattern holds on other platforms remains a question for future research.
“In essence, perceiving a group as entitative makes group members appear more interchangeable,” Sander, a research associate at Humboldt University of Berlin, told PsyPost. “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.”
For those who produce and consume political commentary, the findings point to a small but measurable editorial lever: in this dataset, posts using noun labels and generic phrasing drew higher proportions of prejudiced replies than their adjective-based, non-generic counterparts. The research aligns with observations from Dr. Steven Hassan, a cult expert who spoke with AlterNet in August about President Donald Trump and his political movement. Hassan explained how such movements subtly manipulate their followers to create cult-like cohesion and loyalty. Beyond the way views are worded in casual online posts, he noted, political movements also build media ecosystems in which followers develop distorted perceptions of reality through repeated positive reinforcement from their social networks.
“It really comes down to what information sources they're relying on and who their network is that they identify with,” Hassan told AlterNet. “So if they really love Joe Rogan, they're going to follow a lot of his views. And consequently, when he turns against Trump, Trump loses a lot of followers. What I'm trying to convey is that we're living in an age of digital manipulation that's never been seen before.”
