pith. sign in

arxiv: 2605.11166 · v1 · submitted 2026-05-11 · 💻 cs.CV

Unpacking the Eye of the Beholder: Social Location, Identity, and the Moving Target of Political Perspectives

Elena Sirotkina This is my paper

Pith reviewed 2026-05-13 07:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords visual sentiment analysispolitical identityprotest imageryaudience differencesmachine learningcomputational social scienceimage evaluation
0
0 comments X

The pith

Viewer identity changes perceived violence and engagement in protest imagery

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

This paper develops a classifier that predicts how people with specific political and social identities will evaluate political images. It does so by learning from over 82,000 ratings provided by 5,575 U.S. adults, preserving disagreements across groups instead of averaging them into one score. When used to revisit earlier studies on visual sentiment, the classifier shows that assessments of violence in protest photos and the emotional reasons for engaging with them change once identity is factored in. The result is that political visuals function as moving targets whose content depends on the audience.

Core claim

The paper establishes that the evaluative meaning of a political image varies systematically with the social and political identities of its viewers. The Perspectivist Visual Political Sentiment classifier is built to capture these variations by returning profiles of agreement and divergence along identity dimensions rather than a single aggregated sentiment score. Reanalysis of prior work on protest imagery demonstrates that both perceived violence levels and the emotional mechanisms of engagement shift when audience identity is incorporated.

What carries the argument

The Perspectivist Visual Political Sentiment (PVPS) classifier that models identity-conditioned responses to images from crowdsourced evaluations.

Load-bearing premise

The evaluations from 5,575 U.S. adults are representative enough to train predictions that hold for new images and for identity groups not exactly matching the training sample.

What would settle it

Collect independent ratings on a fresh set of political images from participants grouped by the same identity categories and compare them directly to the classifier's output profiles; systematic mismatches would disprove the generalizability claim.

Figures

Figures reproduced from arXiv: 2605.11166 by Elena Sirotkina.

Figure 1
Figure 1. Figure 1: Illustration of how PVPS classification works. The same image yields structured disagreement across certain demographic axes but not others. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of PVPS classifier architecture. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evaluative profiles for three held-out images. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Extended pipeline for sociodemographic axes. The main PVPS classifier (Fig￾ure 2) fails on sociodemographic dimensions because the per-image evaluative gaps are too noisy. This extension combines two complementary prediction strategies per axis. Path A predicts the evaluative gap from visual content alone. Path B predicts individual ratings from image features and respondent attributes, then derives the ga… view at source ↗
Figure 5
Figure 5. Figure 5: Predicting attention to BLM images over the range of evoked emotions. Top row replicates Casas and Webb Williams (2019, [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correlates of perceived violence in protest images (n = 2,343, UCLA Protest Image Dataset test set). Left panel replicates Won et al. (2017) [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Full evaluative profiles Military at Border LGBTQ+ Rally January 6 STANDALONE POLITICAL Party Dem. Rep. 97% Dem. Rep. 97% Dem. Rep. 95% Ideology Lib. Con. 98% Lib. Con. 97% Lib. Con. 97% Therm. Dem-lean. Rep-lean. 96% Dem-lean. Rep-lean. 95% Dem-lean. Rep-lean. 94% STANDALONE SOCIAL Age Young Old Young Old Young Old Gender Female Male Female Male Female Male Education LowEdu HighEdu LowEdu HighEdu LowEdu H… view at source ↗
read the original abstract

Political and social identities structure how people evaluate political information, a finding decades deep in political science and routinely discarded by computational tools that often produce single scores that treat a piece of text, an image, or a video as if it means the same thing to everyone. This paper shows that it does not, and that the difference is consequential. To address this problem, I develop the Perspectivist Visual Political Sentiment (PVPS) classifier, which learns from approximately 82,000 evaluations by 5,575 U.S. adults to predict how audiences defined by political and social identities will evaluate the same image. Unlike standard tools that average systematic disagreement away, PVPS preserves it, returning an evaluative profile that records who agrees, who diverges, and along which identity lines. Applied to several influential studies of visual sentiment, PVPS shows that perceived violence in protest imagery and the emotional mechanisms behind protest image engagement both change substantively once audience identity is taken into account. It follows that what a political image conveys is a moving target, and measuring it requires knowing whom it is moving.

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

1 major / 1 minor

Summary. The paper develops the Perspectivist Visual Political Sentiment (PVPS) classifier trained on approximately 82,000 evaluations from 5,575 U.S. adults to predict how different political and social identity groups evaluate the same political image. It argues that this approach preserves systematic disagreement unlike standard single-score tools, and demonstrates its utility by re-analyzing influential studies on visual sentiment, finding substantive changes in perceived violence in protest imagery and the emotional mechanisms of protest image engagement when identity is accounted for.

Significance. If the PVPS predictions generalize beyond the training set, this work is significant for highlighting that political images convey different meanings to different audiences, challenging the use of averaged sentiment scores in computational analysis. It provides a method to incorporate identity into visual political sentiment analysis, which could lead to more accurate and nuanced understandings in political science and computer vision.

major comments (1)
  1. [Abstract] The central claim that PVPS reveals substantive shifts in perceived violence and engagement mechanisms relies on the model's ability to accurately predict for identity-defined audiences on new images. However, no details are provided on held-out image performance, cross-validation by image ID, or out-of-distribution testing, raising concerns that reported changes could be artifacts of the training distribution rather than robust findings.
minor comments (1)
  1. [Abstract] The number of evaluations is given as 'approximately 82,000' and participants as '5,575'; providing exact figures would improve precision.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback, which highlights an important aspect of model robustness. We address the single major comment below and commit to revisions that strengthen the presentation of our validation approach.

read point-by-point responses

  1. Referee: [Abstract] The central claim that PVPS reveals substantive shifts in perceived violence and engagement mechanisms relies on the model's ability to accurately predict for identity-defined audiences on new images. However, no details are provided on held-out image performance, cross-validation by image ID, or out-of-distribution testing, raising concerns that reported changes could be artifacts of the training distribution rather than robust findings.

    Authors: We agree that explicit reporting of held-out image performance, image-ID cross-validation, and out-of-distribution testing is necessary to support the claims about substantive shifts in perceived violence and engagement mechanisms. The current manuscript does not include these details, which is a presentational gap. In the revised version we will add a dedicated validation subsection that reports (1) performance metrics when entire images are held out during training, (2) results from k-fold cross-validation stratified by image ID to prevent leakage, and (3) any available out-of-distribution evaluations on political images drawn from sources outside the original training corpus. These additions will demonstrate that the observed changes are not artifacts of the training distribution. revision: yes

Circularity Check

0 steps flagged

No circularity: PVPS is standard supervised learning on human labels

full rationale

The paper trains a classifier on 82,000 human evaluations to output identity-conditioned evaluative profiles for images. No equations, derivations, or self-citations are described that would make predictions equivalent to inputs by construction. The central claim rests on empirical model application to existing studies rather than tautological redefinition or fitted-input renaming. This is self-contained against external benchmarks of held-out performance and does not match any enumerated circularity pattern.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities beyond the general assumption that identity categories structure visual evaluations.

pith-pipeline@v0.9.0 · 5488 in / 1060 out tokens · 59404 ms · 2026-05-13T07:02:59.759111+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

153 extracted references · 153 canonical work pages

  1. [1]

    Transactions of the Association for Computational Linguistics , year =

    Pavlick, Ellie and Kwiatkowski, Tom , title =. Transactions of the Association for Computational Linguistics , year =

    work page

  2. [2]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year =

    K. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year =

    work page

  3. [3]

    Grabe, Maria Elizabeth and Bucy, Erik Page , title =

    work page

  4. [4]

    Political Analysis , volume=

    Computer-assisted text analysis for comparative politics , author=. Political Analysis , volume=. 2015 , publisher=

    work page 2015

  5. [5]

    , title =

    Wolpert, David H. , title =. Neural Networks , volume =

    work page

  6. [6]

    Machine Learning , volume =

    Breiman, Leo , title =. Machine Learning , volume =

    work page

  7. [7]

    Perspectives on Psychological Science , volume =

    Yarkoni, Tal and Westfall, Jacob , title =. Perspectives on Psychological Science , volume =

    work page

  8. [8]

    and Chan, Alexander and Joo, Jungseock , title =

    Steinert-Threlkeld, Zachary C. and Chan, Alexander and Joo, Jungseock , title =. Journal of Politics , year =

    work page

  9. [9]

    and Goren, Amir and Hall, Crystal C

    Todorov, Alexander and Mandisodza, Anesu N. and Goren, Amir and Hall, Crystal C. , title =. Science , volume =

    work page

  10. [10]

    Todorov, Alexander , title =

    work page

  11. [11]

    Political Research Quarterly , volume =

    Casas, Andreu and Webb Williams, Nora , title =. Political Research Quarterly , volume =. 2019 , doi =

    work page 2019

  12. [12]

    , title =

    Webb Williams, Nora and Casas, Andreu and Wilkerson, John D. , title =. 2020 , series =

    work page 2020

  13. [13]

    2018 , address =

    Mason, Lilliana , title =. 2018 , address =

    work page 2018

  14. [14]

    , title =

    Iyengar, Shanto and Lelkes, Yphtach and Levendusky, Matthew and Malhotra, Neil and Westwood, Sean J. , title =. Annual Review of Political Science , volume =

    work page

  15. [15]

    , title =

    Boxell, Levi and Gentzkow, Matthew and Shapiro, Jesse M. , title =. Review of Economics and Statistics , volume =

    work page

  16. [16]

    Sides, John and Tesler, Michael and Vavreck, Lynn , title =

    work page

  17. [17]

    and MacWilliams, Matthew and Nteta, Tatishe , title =

    Schaffner, Brian F. and MacWilliams, Matthew and Nteta, Tatishe , title =. Political Science Quarterly , volume =

    work page

  18. [18]

    and Lee, Taeku , title =

    Hajnal, Zoltan L. and Lee, Taeku , title =

    work page

  19. [19]

    and Collingwood, Loren and Valenzuela, Ali A

    Reny, Tyler T. and Collingwood, Loren and Valenzuela, Ali A. , title =. Public Opinion Quarterly , volume =

    work page

  20. [20]

    Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

    Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya , title =. Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

    work page

  21. [21]

    , title =

    Kornblith, Simon and Shlens, Jonathon and Le, Quoc V. , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  22. [22]

    and Bhatt, Umang and Weller, Adrian , title =

    Collins, Katherine M. and Bhatt, Umang and Weller, Adrian , title =. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing (HCOMP) , year =

    work page

  23. [23]

    , title =

    Iyengar, Shanto and Westwood, Sean J. , title =. American Journal of Political Science , year =

    work page

  24. [24]

    and Anastasopoulos, Jason and Joo, Jungseock , title =

    Xi, Nan and Ma, Di and Liou, Marcus and Steinert-Threlkeld, Zachary C. and Anastasopoulos, Jason and Joo, Jungseock , title =. Proceedings of the 14th International AAAI Conference on Web and Social Media (ICWSM) , year =

    work page

  25. [25]

    and Carmines, Edward G

    Layman, Geoffrey C. and Carmines, Edward G. , title =. The Journal of Politics , year =

    work page

  26. [26]

    , title =

    Rodden, Jonathan A. , title =. 2019 , address =

    work page 2019

  27. [27]

    , title =

    Cramer, Katherine J. , title =. 2016 , address =

    work page 2016

  28. [28]

    2007 , address =

    Prior, Markus , title =. 2007 , address =

    work page 2007

  29. [29]

    , title =

    Graham, Jesse and Haidt, Jonathan and Nosek, Brian A. , title =. Journal of Personality and Social Psychology , year =

    work page

  30. [30]

    and Watts, Duncan J

    Hofman, Jake M. and Watts, Duncan J. and Athey, Susan and Garip, Filiz and Griffiths, Thomas L. and Kleinberg, Jon and Margetts, Helen and Mullainathan, Sendhil and Salganik, Matthew J. and Vazire, Simine and Vespignani, Alessandro and Yarkoni, Tal , title =. Nature , year =

    work page

  31. [31]

    Signs: Journal of Women in Culture and Society , year =

    McCall, Leslie , title =. Signs: Journal of Women in Culture and Society , year =

    work page

  32. [32]

    2018 , note =

    Gelman, Andrew , title =. 2018 , note =

    work page 2018

  33. [33]

    and Judd, Charles M

    McClelland, Gary H. and Judd, Charles M. , title =. Psychological Bulletin , year =

    work page

  34. [34]

    Proceedings of the 1st Conference on Fairness, Accountability and Transparency (FAT*) , year =

    Buolamwini, Joy and Gebru, Timnit , title =. Proceedings of the 1st Conference on Fairness, Accountability and Transparency (FAT*) , year =

    work page

  35. [35]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , year =

    Kearns, Michael and Neel, Seth and Roth, Aaron and Wu, Zhiwei Steven , title =. Proceedings of the 35th International Conference on Machine Learning (ICML) , year =

    work page

  36. [36]

    , title =

    Dalton, Russell J. , title =. Comparative Political Studies , year =

    work page

  37. [37]

    and Mancini, Paolo , title =

    Hallin, Daniel C. and Mancini, Paolo , title =. 2004 , doi =

    work page 2004

  38. [38]

    and Miyamoto, Yuri , title =

    Nisbett, Richard E. and Miyamoto, Yuri , title =. Trends in Cognitive Sciences , year =

    work page

  39. [39]

    and Nisbett, Richard E

    Chua, Hannah Faye and Boland, Julie E. and Nisbett, Richard E. , title =. Proceedings of the National Academy of Sciences , year =

    work page

  40. [40]

    Scientific Reports , volume =

    Kosinski, Michal , title =. Scientific Reports , volume =. 2021 , doi =

    work page 2021

  41. [41]

    and Argyle, Lisa P

    Bail, Christopher A. and Argyle, Lisa P. and Brown, Taylor W. and Bumpus, John P. and Chen, Haohan and Hunzaker, M. B. Fallin and Lee, Jaemin and Mann, Marcus and Merhout, Friedolin and Volfovsky, Alexander , title =. Proceedings of the National Academy of Sciences , year =

    work page

  42. [42]

    American Economic Review , year =

    Allcott, Hunt and Braghieri, Luca and Eichmeyer, Sarah and Gentzkow, Matthew , title =. American Economic Review , year =

    work page

  43. [43]

    , title =

    Bail, Christopher A. , title =

    work page

  44. [44]

    and Barber\'

    Guess, Andrew M. and Barber\'. The Consequences of Online Partisan Media , journal =. 2021 , volume =

    work page 2021

  45. [45]

    Birds of the Same Feather Tweet Together:

    Barber\'. Birds of the Same Feather Tweet Together:. Political Analysis , year =

    work page

  46. [46]

    Journalism & Mass Communication Quarterly , year =

    Geise, Stephanie and Xu, Yi , title =. Journalism & Mass Communication Quarterly , year =

    work page

  47. [47]

    The International Journal of Press/Politics , year =

    Peng, Yilang , title =. The International Journal of Press/Politics , year =

    work page

  48. [48]

    Journal of Communication , year =

    Peng, Yilang , title =. Journal of Communication , year =

    work page

  49. [49]

    Images, Politicians, and Social Media:

    Farkas, X\'. Images, Politicians, and Social Media:. The International Journal of Press/Politics , year =

    work page

  50. [50]

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

    Joshi, Amogh and Buntain, Cody , title =. Proceedings of the International AAAI Conference on Web and Social Media , year =

    work page

  51. [51]

    Social Media + Society , year =

    Kim, Minchul and Bas, Ozen , title =. Social Media + Society , year =

    work page

  52. [52]

    and Norenzayan, Ara , title =

    Henrich, Joseph and Heine, Steven J. and Norenzayan, Ara , title =. Behavioral and Brain Sciences , year =

    work page

  53. [53]

    PLOS ONE , year =

    Gasparyan, Olga and Sirotkina, Elena , title =. PLOS ONE , year =

    work page

  54. [54]

    Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future , pages =

    Basile, Valerio and Fell, Michael and Fornaciari, Tommaso and Hovy, Dirk and Paun, Silviu and Plank, Barbara and Poesio, Massimo and Uma, Alexandra , title =. Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future , pages =

    work page

  55. [55]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Cabitza, Federico and Campagner, Andrea and Basile, Valerio , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

    work page

  56. [56]

    Political Analysis , volume =

    Learning to See: Convolutional Neural Networks for the Analysis of Social Science Data , author =. Political Analysis , volume =. 2022 , publisher =

    work page 2022

  57. [57]

    and Athalye, Anish and Mueller, Jonas , title =

    Northcutt, Curtis G. and Athalye, Anish and Mueller, Jonas , title =. Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , year =

    work page

  58. [58]

    Are We Done with

    Beyer, Lucas and H. Are We Done with. 2020 , eprint =

    work page 2020

  59. [59]

    Proceedings of the ACM on Human-Computer Interaction , year =

    Schaekermann, Mike and Goh, Joslin and Larson, Kate and Law, Edith , title =. Proceedings of the ACM on Human-Computer Interaction , year =. doi:10.1145/3274423 , note =

    work page doi:10.1145/3274423

  60. [60]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , year =

    Fornaciari, Tommaso and Uma, Alexandra and Paun, Silviu and Plank, Barbara and Hovy, Dirk and Poesio, Massimo , title =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , year =

    work page 2021

  61. [61]

    Two Contrasting Data Annotation Paradigms for Subjective

    R. Two Contrasting Data Annotation Paradigms for Subjective. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , year =

    work page 2022

  62. [62]

    The Journal of Politics , volume =

    When Conservatives See Red but Liberals Feel Blue: Labeler Characteristics and Variation in Content Annotation , author =. The Journal of Politics , volume =. 2026 , publisher =

    work page 2026

  63. [63]

    doi:10.1177/0956797615594620 , author =

    Barber. Tweeting From Left to Right: Is Online Political Communication More Than an Echo Chamber? , journal =. 2015 , volume =. doi:10.1177/0956797615594620 , publisher =

    work page doi:10.1177/0956797615594620 2015

  64. [64]

    and Kam, Cindy D

    Kinder, Donald R. and Kam, Cindy D. , title =. 2009 , address =

    work page 2009

  65. [65]

    , title =

    Kahan, Dan M. , title =. Judgment and Decision Making , year =

    work page

  66. [66]

    Proceedings of the 25th ACM International Conference on Multimedia , pages =

    Protest Activity Detection and Perceived Violence Estimation from Social Media Images , author =. Proceedings of the 25th ACM International Conference on Multimedia , pages =

    work page

  67. [67]

    1936 , publisher =

    Mannheim, Karl , title =. 1936 , publisher =

    work page 1936

  68. [68]

    Perspectives on Politics , volume =

    Ange-Marie Hancock , title =. Perspectives on Politics , volume =. 2007 , doi =

    work page 2007

  69. [69]

    and Wills, Julian A

    Brady, William J. and Wills, Julian A. and Jost, John T. and Tucker, Joshua A. and Van Bavel, Jay J. , title =. Proceedings of the National Academy of Sciences , volume =

    work page

  70. [70]

    and Malhotra, Neil and Pan, Jennifer and Barber\'

    Guess, Andrew M. and Malhotra, Neil and Pan, Jennifer and Barber\'. How Do Social Media Feed Algorithms Affect Attitudes and Behavior in an Election Campaign? , journal =

    work page

  71. [71]

    and Bernstein, Michael S

    Jia, Tiziano and Piccardi, Tiziano and Saveski, Martin and Bail, Christopher A. and Bernstein, Michael S. and Gentzkow, Matthew and Leskovec, Jure and Nagler, Jonathan and Nyhan, Brendan and Tucker, Joshua A. , title =. Science , year =

    work page

  72. [72]

    and Hindman, Matthew , title =

    Yang, JungHwan and Davis, Richard A. and Hindman, Matthew , title =. Journal of Communication , volume =

    work page

  73. [73]

    Hameleers, Michael and van der Meer, Toni G. L. A. and Krouwel, Andr\'. The Nature of Visual Disinformation Online: A Qualitative Content Analysis of Alternative and Social Media in the. Political Communication , volume =

    work page

  74. [74]

    and van der Linden, Sander , title =

    Rathje, Steve and Van Bavel, Jay J. and van der Linden, Sander , title =. Proceedings of the National Academy of Sciences , volume =

    work page

  75. [75]

    , title =

    Milli, Smitha and Carroll, Micah and Wang, Yike and Dragan, Anca D. , title =. Science , volume =

    work page

  76. [76]

    Journal of Politics , volume =

    Barari, Soubhik and Lucas, Christopher and Munger, Kevin , title =. Journal of Politics , volume =

    work page

  77. [77]

    Political Analysis , volume =

    Samuel Bestvater and Brendan Monroe , title =. Political Analysis , volume =. 2023 , doi =

    work page 2023

  78. [78]

    and Luckmann, Thomas , title =

    Berger, Peter L. and Luckmann, Thomas , title =. 1966 , publisher =

    work page 1966

  79. [79]

    , title =

    Grimmer, Justin and Stewart, Brandon M. , title =. Political Analysis , year =

    work page

  80. [80]

    , title =

    Joo, Jungseock and Steinert-Threlkeld, Zachary C. , title =. Computational Communication Research , year =

    work page

Showing first 80 references.