“Meet the new boss – Same as the old boss”: A longitudinal study of political post sentiment and Facebook engagement

Summary

This paper tracks a decade (2013–2023) of Facebook activity from Norway’s right-wing populist Progress Party and its two successive leaders, Siv Jensen and Sylvi Listhaug, to examine how the sentiment of political posts relates to audience engagement over time. Larsson uses a hybrid pipeline — GPT-4 zero-shot classification of 11,160 Norwegian-language posts, validated against human coders — to categorize posts as negative, positive, or neutral, then compares shares, comments, and likes across sentiment, account, and year. The central argument is twofold: negativity in populist Facebook content has grown (especially after Listhaug’s 2021 takeover) and reliably amplifies shares and comments, while likes follow a different, more positivity-driven logic. The title’s nod to “the new boss” captures the finding that leadership transitions reshape party-level communication style toward greater negativity.

Key Contributions

• A rare 10-year longitudinal account of populist Facebook communication outside the Anglo-American context.
• A reproducible hybrid LLM + human coding workflow for sentiment annotation in Norwegian, with strong reliability (α = .91 intracoder, .82 intercoder).
• Empirical disaggregation of engagement types, showing likes diverge from shares and comments in their relation to sentiment.
• Integration of supply-side (party/leader production) and demand-side (audience engagement) analyses in one design.
• Evidence bearing on negative campaigning, permanent campaigning, and political professionalization debates.

Methods

CrowdTangle was used to harvest 11,160 public posts from the Progress Party page and the personal pages of Jensen and Listhaug (2013–2023). Each post was classified as negative/positive/neutral campaigning by GPT-4 (gpt-4-0314) using a Norwegian-language zero-shot prompt; reliability was tested by re-running the LLM on a 20% sample and by independent human coding of a 5% sample. Engagement (shares, comments, likes) was analyzed with medians, Kruskal-Wallis, and Dunn’s tests due to non-normal distributions, accompanied by visual breakdowns across account, sentiment, and year.

Findings

• Negative content rose markedly over the decade, with Listhaug shifting toward negativity from 2017 and the party page following after her 2021 ascension.
• Negative posts produced the highest median shares across nearly all accounts and years (H2a confirmed).
• Negative posts also produced more comments, with 2021 (election year) as a notable exception (H2b largely confirmed).
• Likes diverged: positive posts dominated likes for Jensen (2020–2023) and Listhaug (2021–2022), suggesting “lightweight” engagement obeys a different affective logic.
• Election years no longer produced clear engagement surges, consistent with a shift toward permanent campaigning.
• Posting volume mirrored leadership status: Listhaug ramped up after 2021, Jensen wound down.

Connections

This paper sits naturally alongside other work using LLMs as annotation instruments for political and media content — particularly Balluff2026-if and Le-Mens2025-qz on validating LLM-based text classification, and Tornberg2025-ir / Tornberg2026-lc on LLMs in political and social analysis. Its non-English (Norwegian) application speaks to multilingual evaluation concerns raised in Nguyen2026-vm. Substantively, the focus on populist negativity and platform dynamics resonates with DeVerna2025-dl on social media political content, though most other papers in this register address LLM capabilities rather than political communication directly.

Podcast

A research-radio episode discusses this paper: Listen