Understanding the Political Ideology of Legislators from Social Media Images

Di Ma; Jason Anastasopoulos; Jungseock Joo; Marcus Liou; Nan Xi; Zachary C. Steinert-Threlkeld

arxiv: 1907.09594 · v1 · pith:VU25NGMWnew · submitted 2019-07-22 · 💻 cs.SI · cs.CV· cs.HC· cs.MM· stat.AP

Understanding the Political Ideology of Legislators from Social Media Images

Nan Xi , Di Ma , Marcus Liou , Zachary C. Steinert-Threlkeld , Jason Anastasopoulos , Jungseock Joo This is my paper

Pith reviewed 2026-05-24 17:23 UTC · model grok-4.3

classification 💻 cs.SI cs.CVcs.HCcs.MMstat.AP

keywords political ideologysocial media imagesdeep learningparty affiliationU.S. CongressFacebook photographsimage content analysis

0 comments

The pith

Deep learning on Facebook photos of U.S. legislators produces probabilities that proxy left-right ideology.

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

The paper trains a deep learning model to classify whether a photograph comes from a Republican or Democratic member of the 114th Congress. The resulting class probabilities are presented as a continuous measure of ideological content along the liberal-conservative axis. Single-image accuracy reaches 59.28 percent and rises to 82.35 percent when scores from up to 150 photographs of the same legislator are averaged, after statistical controls for gender and race. Separate content analysis shows that conservative images more often depict status-quo institutions, dominant social groups, and happier expressions.

Core claim

Predicted class probabilities from our model function as an accurate proxy of the political ideology of images along a left-right (liberal-conservative) dimension. After controlling for the gender and race of politicians, our method achieves an accuracy of 59.28% from single photographs and 82.35% when aggregating scores from multiple photographs (up to 150) of the same person.

What carries the argument

A deep learning classifier trained to output Republican versus Democratic probabilities from individual Facebook photographs, with the probability scores treated as ideology measures.

If this is right

Ideological signals in images can be measured at scale across many politicians without hand coding.
Aggregating multiple images per person substantially increases the reliability of the visual ideology proxy.
Conservative images differ systematically from liberal ones in support for hierarchy and emotional tone.

Where Pith is reading between the lines

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

The method could be extended to track how visual rhetoric changes over time or across platforms.
If the visual proxy holds, it offers a new way to study ideology in populations that produce fewer text records.
The same pipeline might be tested on images from non-legislator political actors to check generalizability.

Load-bearing premise

Party labels remain a clean left-right proxy once gender and race are controlled, and the photographs legislators post are an unbiased sample of their visual ideological expression.

What would settle it

Apply the trained model to a fresh set of congressional photographs and test whether the resulting ideology scores correlate with independent measures such as DW-Nominate roll-call scores or survey-based ideology ratings for the same legislators.

read the original abstract

In this paper, we seek to understand how politicians use images to express ideological rhetoric through Facebook images posted by members of the U.S. House and Senate. In the era of social media, politics has become saturated with imagery, a potent and emotionally salient form of political rhetoric which has been used by politicians and political organizations to influence public sentiment and voting behavior for well over a century. To date, however, little is known about how images are used as political rhetoric. Using deep learning techniques to automatically predict Republican or Democratic party affiliation solely from the Facebook photographs of the members of the 114th U.S. Congress, we demonstrate that predicted class probabilities from our model function as an accurate proxy of the political ideology of images along a left-right (liberal-conservative) dimension. After controlling for the gender and race of politicians, our method achieves an accuracy of 59.28% from single photographs and 82.35% when aggregating scores from multiple photographs (up to 150) of the same person. To better understand image content distinguishing liberal from conservative images, we also perform in-depth content analyses of the photographs. Our findings suggest that conservatives tend to use more images supporting status quo political institutions and hierarchy maintenance, featuring individuals from dominant social groups, and displaying greater happiness than liberals.

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.

Desk Editor's Note private letter to a colleague

The paper gets above-chance party prediction from congressional Facebook photos even after gender and race controls, plus some content-analysis differences, but the control procedure is not described.

read the letter

The main takeaway is that a standard deep-learning classifier trained on 114th Congress Facebook photos can separate Republican from Democratic images at 59% accuracy on single shots and 82% when averaging up to 150 photos per person, and that this holds after they control for gender and race. They also report that conservative images more often show status-quo institutions, hierarchy cues, dominant-group members, and higher happiness expressions. That combination of quantitative proxy and qualitative breakdown is the concrete addition here. The work is straightforward supervised learning against party labels, so the accuracies are not derived from any new theoretical reduction; they are empirical results on this particular corpus. The content analysis is manual and appears to be the part that tries to interpret what the model is picking up. The soft spot is the missing description of exactly how the gender and race controls were applied. The abstract states the accuracies survive the controls, but without knowing whether they used stratification, regression adjustment, or matching, it is hard to judge whether residual demographic correlation remains. The assumption that posted Facebook photos are an unbiased window into ideological expression is also taken as given rather than tested. For readers working on computational political communication or visual rhetoric, the numbers and the content findings are worth seeing. The paper is coherent on its own terms and shows honest engagement with the data it has. I would bring it to a reading group for discussion of the control issue and the sampling frame. I would not cite it in my own work unless the methods section clarifies the controls. A serious editor should send it to peer review so the control procedure and any robustness checks can be examined.

Referee Report

2 major / 0 minor

Summary. The paper claims that a deep learning model trained on Facebook photographs posted by members of the 114th U.S. Congress can predict Republican vs. Democratic party affiliation, with the resulting class probabilities serving as a proxy for the left-right ideological content of the images. After controlling for gender and race, it reports 59.28% accuracy from single photographs and 82.35% accuracy when aggregating predictions across up to 150 photographs per legislator. A follow-up content analysis finds that conservative images more often support status-quo institutions, feature dominant social groups, and display greater happiness than liberal images.

Significance. If the central claim holds after proper controls and validation, the work would supply a replicable, image-based measure of visual ideological rhetoric that could be applied to large-scale social-media corpora in political communication research. The reported gain from aggregation (single-image to multi-image) is a concrete, falsifiable strength, and the content-analysis component supplies qualitative grounding for the quantitative results.

major comments (2)

[Abstract] Abstract: the accuracies of 59.28% (single photograph) and 82.35% (aggregated) are asserted to hold 'after controlling for the gender and race of politicians,' yet no description is given of the control procedure (stratified sampling, regression adjustment, matched pairs, or otherwise). Because the central claim is that the model captures ideology rather than demographic correlates of party, the absence of this detail is load-bearing; residual confounding would undermine the proxy interpretation.
[Abstract] Abstract: the manuscript supplies no information on neural-network architecture, training-set size, cross-validation scheme, or hyperparameter selection. Without these details the reported accuracies cannot be assessed for overfitting, data leakage, or robustness, directly affecting evaluation of the proxy claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive report. We address each major comment below and commit to revisions that strengthen the manuscript's clarity and reproducibility.

read point-by-point responses

Referee: [Abstract] Abstract: the accuracies of 59.28% (single photograph) and 82.35% (aggregated) are asserted to hold 'after controlling for the gender and race of politicians,' yet no description is given of the control procedure (stratified sampling, regression adjustment, matched pairs, or otherwise). Because the central claim is that the model captures ideology rather than demographic correlates of party, the absence of this detail is load-bearing; residual confounding would undermine the proxy interpretation.

Authors: We agree that the control procedure must be described explicitly, as residual demographic confounding would weaken the ideological-proxy interpretation. The current manuscript does not provide this detail. In the revised version we will add a clear description of the procedure (including whether stratified sampling, covariate adjustment, or another method was used) both in the abstract and in a dedicated methods subsection. revision: yes
Referee: [Abstract] Abstract: the manuscript supplies no information on neural-network architecture, training-set size, cross-validation scheme, or hyperparameter selection. Without these details the reported accuracies cannot be assessed for overfitting, data leakage, or robustness, directly affecting evaluation of the proxy claim.

Authors: We concur that these implementation details are necessary for evaluating the reliability of the reported accuracies. The manuscript currently omits them. We will insert a new methods section that specifies the neural-network architecture, training-set size, cross-validation scheme, and hyperparameter selection process, thereby allowing readers to assess potential overfitting or leakage. revision: yes

Circularity Check

0 steps flagged

No significant circularity; standard supervised learning against external labels

full rationale

The paper trains a deep learning classifier to predict party affiliation (Republican/Democrat) from Facebook images of legislators and reports classification accuracies (59.28% single-image, 82.35% aggregated) after demographic controls. No equations, derivations, or self-referential definitions appear in the provided text that would reduce these accuracies or the ideology-proxy claim to fitted parameters by construction. The reported results rely on empirical performance against held-out party labels rather than tautological renaming or self-citation chains. The control procedure for gender and race is not detailed in the abstract, but this is a methodological transparency issue rather than circularity in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard machine-learning assumptions plus domain assumptions about political imagery; no invented physical entities are introduced.

free parameters (1)

neural-network weights and hyperparameters
Deep-learning model parameters are fitted to the labeled image data; exact count and regularization choices are not stated in the abstract.

axioms (2)

domain assumption Party affiliation is a valid proxy for left-right political ideology
Used to label training images and interpret model outputs as ideology scores.
domain assumption Facebook photographs constitute an unbiased sample of legislators' visual rhetoric
Implicit in treating the collected images as representative for both prediction and content analysis.

pith-pipeline@v0.9.0 · 5788 in / 1367 out tokens · 28360 ms · 2026-05-24T17:23:08.313189+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Using deep learning techniques to automatically predict Republican or Democratic party affiliation solely from the Facebook photographs... ResNet-34 architecture... Grad-CAM... Google Vision API
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

predicted class probabilities... proxy of the political ideology... accuracy of 59.28% from single photographs and 82.35% when aggregating

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.