pith. sign in

arxiv: 2605.12526 · v1 · pith:TAX73HZCnew · submitted 2026-04-10 · 💻 cs.SI · cs.CY

"F*** You Biden": Cross-Partisan Electoral Toxicity on X

Danishjeet Singh , Anindya Mondal , Filippo Menczer This is my paper

Pith reviewed 2026-05-14 21:48 UTC · model grok-4.3

classification 💻 cs.SI cs.CY
keywords political toxicitycross-partisan repliessocial media2024 electionX platformpartisan asymmetryreply volumeonline discourse
0
0 comments X

The pith

Republican-leaning posts on X are more toxic than Democratic ones, yet Democratic posts attract more toxic replies because Republicans generate most cross-partisan replies.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper analyzes millions of original posts and replies on X collected during the 2024 U.S. presidential election. It classifies each post and user as Republican- or Democratic-leaning with a human-validated large language model and scores toxicity with the Perspective API. The central finding is an asymmetry: Republican-leaning posts score higher in toxicity, but Democratic-leaning posts receive higher-toxicity replies overall. The difference arises because Republican users reply to Democratic posts far more often than the reverse, even though cross-partisan replies are only slightly more toxic than same-party replies for both sides. This pattern matters for understanding whether online electoral hostility stems from how partisans speak or from how they engage the other side.

Core claim

Republican-leaning posts are significantly more toxic than Democratic-leaning posts, yet Democratic-leaning posts attract significantly more toxic replies. Cross-partisan replies are slightly but significantly more toxic than same-party replies for both groups, but Republican users account for the large majority of replies to Democratic posts while Democrats account for a minority of replies to Republican content. Therefore the elevated toxicity directed at Democratic content is better explained by the volume of Republican cross-partisan replies.

What carries the argument

Asymmetry between outgoing toxicity of posts and incoming toxicity of replies, measured by comparing same-party versus cross-partisan reply volumes and toxicity scores.

If this is right

  • Cross-partisan replies carry modestly higher toxicity than same-party replies regardless of the original post's alignment.
  • The bulk of toxic replies to Democratic content comes from Republican users rather than from elevated per-reply toxicity.
  • Moderation strategies focused only on post toxicity would miss the reply-volume driver of incoming hostility.
  • Electoral periods amplify the observed asymmetry because cross-partisan engagement rises.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If reply volume is the dominant factor, platforms could reduce perceived toxicity by limiting cross-partisan reply reach rather than by lowering post toxicity thresholds.
  • The pattern suggests that interventions aimed at encouraging same-party replies might lower overall toxicity more effectively than uniform toxicity filters.
  • Similar volume-driven asymmetries may appear in other high-stakes topics such as public health or climate policy where one side engages the other more aggressively.

Load-bearing premise

The large language model correctly identifies the political alignment of posts and users, and the Perspective API measures toxicity without systematic partisan bias.

What would settle it

Re-running the analysis after swapping the alignment classifier for an independent human-coded sample or a different model and finding that the toxicity-volume asymmetry disappears.

Figures

Figures reproduced from arXiv: 2605.12526 by Anindya Mondal, Danishjeet Singh, Filippo Menczer.

Figure 1
Figure 1. Figure 1: Kernel density distributions of post toxicity scores (0–1, higher = more toxic). [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Kernel density distributions of reply toxicity scores (0–1, higher = more toxic) by [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
read the original abstract

Political discourse on social media has grown increasingly toxic, with electoral periods amplifying partisan hostility and cross-group attacks. Yet it remains unclear whether toxicity in online political speech reflects how partisans communicate within their own circles, or how aggressively they engage with the opposition. Disentangling these dynamics is critical for understanding online political hostility and for designing effective content moderation. We examine this question at scale using a large collection of original posts and replies from X (formerly Twitter), collected during the 2024 U.S. presidential election. Using a human-validated large language model to classify the political alignment of posts and users, and the Perspective API for toxicity scoring, we uncover a striking asymmetry: Republican-leaning posts are significantly more toxic than Democratic-leaning posts, yet Democratic-leaning posts attract significantly more toxic replies. To interpret this finding, we compare the toxicity of same-party and cross-partisan replies. While cross-partisan replies are slightly but significantly more toxic than same-party replies, this is true for both Democratic and Republican posts. However, Republican users account for a large majority of replies to Democratic posts, while Democrats account for a minority of replies to Republican content. Therefore, the elevated toxicity directed at Democratic content is better explained by the volume of Republican cross-partisan replies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript analyzes a large corpus of X posts and replies from the 2024 U.S. presidential election. Using a human-validated LLM to label political alignment of posts and users and the Perspective API to score toxicity, it reports that Republican-leaning original posts are significantly more toxic than Democratic-leaning ones, yet Democratic-leaning posts receive significantly more toxic replies. The authors attribute the latter pattern to the much higher volume of Republican cross-partisan replies rather than to any difference in the toxicity of cross- versus same-party replies.

Significance. If the measurement pipeline is unbiased, the result supplies concrete evidence of asymmetric toxicity flows in electoral discourse and shows that reply-volume effects can dominate per-reply toxicity differences. The scale of the data collection and the explicit decomposition into same-party versus cross-partisan replies are strengths that could inform both academic understanding and platform moderation design.

major comments (2)
  1. [Methods] Methods section on LLM political classification: the paper states the model is 'human-validated' but reports neither validation sample size, per-class F1 scores, nor agreement rates stratified by toxicity level. Because the central asymmetry claim rests on accurate separation of Republican- versus Democratic-leaning content, differential error rates correlated with aggressive language would directly undermine both the original-post toxicity gap and the reply-volume explanation.
  2. [Results] Results and discussion of Perspective API scores: no robustness check or stratified human validation is provided to confirm that toxicity scores are comparable across partisan lexicons. If the API systematically flags right-leaning terms more readily, the reported Republican original-post toxicity advantage and the volume-based account of reply toxicity would both be artifacts of measurement rather than substantive patterns.
minor comments (2)
  1. [Abstract] The abstract and title use censored profanity; consider whether this is necessary for the journal's readership or if a neutral phrasing would suffice.
  2. [Figures] Figure captions and axis labels should explicitly state the exact toxicity threshold or percentile used when binarizing Perspective scores, if any such threshold appears in the analysis.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive comments on our manuscript. We have carefully considered each point and revised the paper to strengthen the methodological transparency and robustness of our findings. Below we respond point by point.

read point-by-point responses

  1. Referee: [Methods] Methods section on LLM political classification: the paper states the model is 'human-validated' but reports neither validation sample size, per-class F1 scores, nor agreement rates stratified by toxicity level. Because the central asymmetry claim rests on accurate separation of Republican- versus Democratic-leaning content, differential error rates correlated with aggressive language would directly undermine both the original-post toxicity gap and the reply-volume explanation.

    Authors: We agree that the current description of the LLM validation is insufficiently detailed. In the revised manuscript we have expanded the Methods section to report the validation sample size, per-class F1 scores, and inter-rater agreement rates stratified by toxicity quintiles. These metrics show balanced performance across toxicity levels and no systematic differential error that would artifactually inflate Republican toxicity or distort the reply-volume interpretation. revision: yes

  2. Referee: [Results] Results and discussion of Perspective API scores: no robustness check or stratified human validation is provided to confirm that toxicity scores are comparable across partisan lexicons. If the API systematically flags right-leaning terms more readily, the reported Republican original-post toxicity advantage and the volume-based account of reply toxicity would both be artifacts of measurement rather than substantive patterns.

    Authors: We share the referee's concern about possible partisan bias in the Perspective API. We have added a new subsection in the Results that presents a stratified human validation of 1,200 posts (balanced by party and toxicity) and a sensitivity analysis that removes the most partisan lexical items. Both checks confirm that the API scores remain comparable across partisan lexicons and that the main asymmetry findings are robust to these controls. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical analysis with external tools

full rationale

The paper conducts an observational study of X posts during the 2024 election using a human-validated LLM for political alignment classification and the Perspective API for toxicity scoring. No equations, fitted parameters, or derivations appear in the abstract or described full text. Claims rest on direct statistical comparisons of observed toxicity levels and reply volumes across partisan groups. No self-definitional loops, fitted-input predictions, or load-bearing self-citations are present. The analysis is self-contained against external benchmarks (data collection and off-the-shelf classifiers) and does not reduce any result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on two key domain assumptions about the reliability of the classification and toxicity tools; no free parameters or invented entities are introduced.

axioms (2)
  • domain assumption The human-validated LLM accurately classifies political alignment of posts and users
    Invoked to label the large collection of posts and replies; validation details not provided in abstract.
  • domain assumption Perspective API toxicity scores are reliable and unbiased across partisan groups
    Used as the primary toxicity metric without additional calibration or bias checks described.

pith-pipeline@v0.9.0 · 5527 in / 1276 out tokens · 45329 ms · 2026-05-14T21:48:20.433397+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Political Polarization on Twitter , url =

    Michael Conover and Jacob Ratkiewicz and Matthew Francisco and Bruno Gon. Political Polarization on Twitter , url =. Proc. 5th International AAAI Conference on Weblogs and Social Media (ICWSM) , doi =

    work page

  2. [2]

    Jana Belschner , title =

    work page

  3. [3]

    2022 , note=

    Political Resources and Online Political Hostility: How and Why Hostility Is More Prevalent Among the Resourceful , author=. 2022 , note=

    work page 2022

  4. [4]

    ArXiv , year=

    LLMs left, right, and center: Assessing GPT's capabilities to label political bias from web domains , author=. ArXiv , year=

    work page

  5. [5]

    2024 , url =

    OpenAI , title =. 2024 , url =

    work page 2024

  6. [6]

    The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution , volume =

    Nachar, Nadim , year =. The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution , volume =. Tutorials in Quantitative Methods for Psychology , doi =

    work page

  7. [7]

    2022 , journal =

    A New Generation of Perspective API: Efficient Multilingual Character-level Transformers , author=. 2022 , journal =

    work page 2022

  8. [8]

    Presidential Election on Twitter/X , author=

    A Public Dataset Tracking Social Media Discourse about the 2024 U.S. Presidential Election on Twitter/X , author=. Proceedings of the International AAAI Conference on Web and Social Media , year=

    work page 2024

  9. [9]

    A., Boerner, T

    Hancock, David Y. and Fischer, Jeremy and Lowe, John Michael and Snapp-Childs, Winona and Pierce, Marlon and Marru, Suresh and Coulter, J. Eric and Vaughn, Matthew and Beck, Brian and Merchant, Nirav and Skidmore, Edwin and Jacobs, Gwen , title =. Practice and Experience in Advanced Research Computing 2021: Evolution Across All Dimensions , articleno =. 2...

    work page doi:10.1145/3437359.3465565 2021

  10. [10]

    & Westwood, S

    The Origins and Consequences of Affective Polarization in the United States , journal =. doi:10.1146/annurev-polisci-051117-073034 , author =

    work page doi:10.1146/annurev-polisci-051117-073034

  11. [11]

    doi:10.1177/0956797615594620 , author =

    Tweeting From Left to Right , journal =. doi:10.1177/0956797615594620 , author =

    work page doi:10.1177/0956797615594620

  12. [12]

    doi:10.1073/pnas.2024292118 , author =

    Out-group animosity drives engagement on social media , journal =. doi:10.1073/pnas.2024292118 , author =

    work page doi:10.1073/pnas.2024292118

  13. [13]

    doi:10.1007/s11109-022-09850-x , author =

    Partisanship on Social Media: In-Party Love Among American Politicians, Greater Engagement with Out-Party Hate Among Ordinary Users , journal =. doi:10.1007/s11109-022-09850-x , author =

    work page doi:10.1007/s11109-022-09850-x

  14. [14]

    A.et al.Exposure to opposing views on social media can increase political polarization.Proceedings of the National Academy of Sciences115, 9216–9221 (2018)

    Exposure to opposing views on social media can increase political polarization , journal =. doi:10.1073/pnas.1804840115 , author =

    work page doi:10.1073/pnas.1804840115

  15. [15]

    URLhttps://www.science

    Reranking partisan animosity in algorithmic social media feeds alters affective polarization , journal =. doi:10.1126/science.adu5584 , author =

    work page doi:10.1126/science.adu5584

  16. [16]

    doi:10.1038/s41467-024-53868-0 , author =

    Patterns of partisan toxicity and engagement reveal the common structure of online political communication across countries , journal =. doi:10.1038/s41467-024-53868-0 , author =

    work page doi:10.1038/s41467-024-53868-0

  17. [17]

    doi:10.1140/epjds6 , author =

    Partisan asymmetries in online political activity , journal =. doi:10.1140/epjds6 , author =

    work page doi:10.1140/epjds6

  18. [18]

    In: 2024 Conference on Computer-Supported Cooperative Work and Social Computing

    Characterization of Political Polarized Users Attacked by Language Toxicity on. Companion Publication of the 2024 Conference on Computer-Supported Cooperative Work and Social Computing , pages =. doi:10.1145/3678884.3681849 , author =

    work page doi:10.1145/3678884.3681849 2024

  19. [19]

    Proceedings of the International AAAI Conference on Web and Social Media , author =

    Proceedings of the International AAAI Conference on Web and Social Media , number =. doi:10.1609/icwsm.v19i1.35819 , volume =

    work page doi:10.1609/icwsm.v19i1.35819

  20. [20]

    Proceedings of the National Academy of Sciences 120(30)

    Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2305016120 , author =

    work page doi:10.1073/pnas.2305016120

  21. [21]

    Can Large Language Models Transform Computational Social Science?

    Can Large Language Models Transform Computational Social Science? , journal =. doi:10.1162/coli_a_00502 , author =

    work page doi:10.1162/coli_a_00502