💻 computer science

Bilder ohne Worte: Can AI Identify Political Parties via Unspoken Visual Signatures in the 2025 German Election?

This paper introduces PREF, an interpretable vision-language model that successfully classifies political party affiliation from 2025 German election campaign media by verbalizing perceptual cues, achieving performance comparable to black-box embeddings while quantifying the extent to which party identity persists even after removing explicit branding.

Original authors: Ali Abusaleh, Alexander Mehler, Mevlut Bagci

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Abusaleh, Alexander Mehler, Mevlut Bagci

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern political arena, the most powerful messages are often the ones never spoken. While campaigns still rely on policy speeches and written manifestos, the daily rhythm of political communication has shifted toward short videos and images shared on social media. These visual snippets carry a wealth of information beyond the words on the screen. Just as a person's posture, clothing, and the tone of their voice can reveal their mood or background, the visual style of a political campaign can signal its identity. Researchers have long known that humans can make quick, surprisingly accurate guesses about a person's political leanings based on brief visual cues. The question now is whether a computer can do the same, not by reading the text, but by understanding the silent language of the image itself.

This is the challenge faced by a team of researchers at Goethe University Frankfurt, who set out to see if artificial intelligence could identify German political parties solely by looking at the visual and auditory signatures of their campaign media. They were not interested in guessing the private beliefs of individual voters or politicians. Instead, they wanted to know if the coordinated style of a party's official posts—the way they frame a shot, the colors they choose, the formality of the clothing, and the tone of the voice—creates a unique fingerprint that a machine can recognize. To test this, they gathered over 15,000 images and videos posted by politicians during the 2025 German federal election. These posts came from the major parties, ranging from the conservative bloc to the progressive greens and the liberal free democrats.

The researchers built a system they call PREF, which stands for a Perceptual-Verbal Feature model. Instead of feeding the raw images directly into a complex, opaque computer program that acts like a black box, they first asked an advanced AI to describe what it saw and heard in plain language. The system analyzed the videos and images to generate a list of observable details: how many people were in the frame, whether the setting was indoors or outdoors, if the speaker was wearing a suit or casual clothes, the emotional tone of their voice, and the presence of specific objects like flags or charts. These descriptions were then fed into a transparent classifier, a type of computer program that makes decisions based on clear, understandable rules, to determine which party had posted the content.

The results were striking. The system correctly identified the political party behind a post nearly 85 percent of the time. This level of accuracy matched the performance of much more complex, "black box" systems that rely on hidden mathematical patterns rather than human-readable descriptions. More importantly, the researchers wanted to know how much of this success depended on obvious branding, such as party logos, flags, or campaign colors. To find out, they ran a second test where they stripped away all those explicit symbols from the AI's description, leaving only the subtle cues of style and behavior. Even without the logos or the specific campaign colors, the system still identified the correct party with high accuracy, dropping only slightly to about 84 percent. This suggests that political parties have developed distinct, unspoken visual languages. A conservative party might consistently use formal attire, serious tones, and podium settings, while a progressive party might favor casual clothing, diverse crowds, and a more measured tone. These patterns are so consistent that they remain recognizable even when the obvious signs of identity are removed.

To ensure these findings were not just an artifact of the computer, the researchers also asked human volunteers to look at a small sample of the same images and videos. The humans were able to identify the political party based on the visual cues alone with a high degree of agreement, confirming that the signals the machine detected were real and observable by people. The study also showed that the visual and auditory cues were just as effective, if not slightly more so, than the text captions that usually accompany these posts. This indicates that the visual strategy of a campaign carries its own heavy weight in shaping public perception.

The study concludes that political identity is encoded in a structured, perceptual signature that can be decoded without relying on text. While the technology used in this research is designed to analyze public, official campaign materials, the ability to automatically detect these patterns raises important questions about the future of political communication. The researchers emphasize that their work focuses on the institutional branding of parties, not the private ideology of individuals, and they call for strict ethical safeguards to prevent such tools from being used to profile private citizens or manipulate political messaging. Ultimately, the work demonstrates that in the digital age, the way a party looks and sounds is just as much a part of its message as the words it chooses to say.

Technical Summary: "Bilder ohne Worte: Can AI Identify Political Parties via Unspoken Visual Signatures in the 2025 German Election?"

Problem Statement
Computational political communication research has historically remained text-centric, despite the shift of political campaigning toward short-form video and images on platforms like X (formerly Twitter). Existing multimodal approaches often rely on black-box embeddings (e.g., CLIP, standard VLM embeddings) that lack interpretability and transparency—critical requirements for political analysis. Furthermore, these embedding-based methods struggle with semantic opacity, lack of interpretability, and "congruence blindness" (insufficient modeling of cross-modal alignment). The authors pose two central questions: (1) Can an interpretable representation match the performance of black-box embeddings in classifying party affiliation from campaign media? (2) To what extent does party identifiability persist when explicit branding (logos, flags, colors, text) is removed, leaving only subtle, coordinated aesthetic and behavioral cues?

Methodology: The PREF Framework
The authors introduce PREF (Perceptual–Verbal Feature model), a modular pipeline designed to convert raw media into structured, human-readable feature schemas. The framework operates in three stages:

Multimodal Feature Extraction: Using the Qwen2.5-VL Vision-Language Model (VLM), the system processes images and video frames to generate structured semantic descriptions. These descriptions cover three modalities: Vision (what is presented), Auditory (what is heard in video), and Temporal (how frames cohere). The output adheres to a strict JSON schema.
Hybrid Feature Elicitation: To ensure the feature taxonomy is both theoretically grounded and data-driven while mitigating researcher bias, the authors employ a Multi-Model Consensus strategy. An ensemble of six state-of-the-art LLMs (including GPT-5, Llama3.2, DeepSeek-R1, Mistral, Phi-4, and GPT-OSS) independently proposes observable audiovisual features. Features are retained only if they achieve consensus (suggested by at least two models), creating an objective schema.
Interpretable Classification: The extracted perceptual-verbal features are fed into transparent classifiers (Decision Trees, Random Forests, Logistic Regression, etc.). The authors prioritize Random Forests for their ability to provide feature importance and decision paths, ensuring the model's reasoning is auditable.

Experimental Setup
The study utilizes the PolTweetCorpus, a dataset of 39,546 tweets collected during the 2025 German federal election campaign (November 6, 2024 – February 23, 2025). The analysis focuses on 15,043 labeled media items (images and videos) posted by verified representatives of seven major German political parties (labeled as Groups A–G) and an "Others" category.

Data Split: 80% training (12,034 samples), 20% testing (3,009 samples).
Ablation Study (H2): A "controlled de-branding" procedure systematically removes explicit branding features (Group G1: logos, flags, campaign colors, text overlays) to test if the model can still classify parties based solely on subtle cues (Group G2: attire, scene setting, tone, motion).
Human Validation: A study on a random subset ( $n=140$ ) involved human annotators labeling the same 12 semantic dimensions used by PREF to measure Inter-Annotator Agreement (IAA) and LLM-Human alignment.

Key Results

Performance vs. Baselines: The PREF framework, using a Random Forest classifier on perceptual-verbal features, achieved 84.98% accuracy and an AUC of 0.9219. This performance matches or exceeds that of text-based models and opaque embedding-based baselines (e.g., k-NN, Neural Networks on raw embeddings), which struggled with minority class sensitivity and achieved lower AUCs (e.g., 0.8267 for RF on raw embeddings).
De-Branding Robustness (H2): Crucially, when explicit branding features (G1) were removed, the model's accuracy only dropped marginally to 84.11% (AUC 0.8977). This confirms that political parties possess distinct, measurable aesthetic and behavioral "signatures" (G2) that persist even without logos or text.
Human Validation: Human annotators achieved substantial agreement on the PREF schema (Krippendorff's $\alpha = 0.67$ overall; $\alpha = 0.78$ for visual party identification). The VLM (Qwen2.5-VL) showed moderate alignment with humans ( $\alpha = 0.50$ ), with high agreement on party-relevant dimensions ( $\alpha = 0.73$ ), confirming that the model's cues are human-verifiable and not merely artifacts of the VLM's internal geometry.
Model Efficiency: The study found a "saturation effect" where the 3B-parameter Qwen2.5-VL model performed comparably to larger 32B and 72B variants, suggesting that smaller models can efficiently extract task-relevant knowledge for this specific domain.

Significance and Claims
The paper claims three primary contributions:

Interpretable Parity: It demonstrates that a transparent, verbalized-feature framework can match the predictive power of black-box embeddings in political classification, offering a path toward auditable AI in sensitive sociopolitical domains.
Decomposition of Branding vs. Style: Through the de-branding ablation, the study provides the first quantification of how much party identity survives the removal of explicit symbols. It concludes that party affiliation is encoded not just in logos, but in a "visual language" of coordinated aesthetics (attire, tone, setting) that is empirically detectable.
Bias-Mitigated Feature Elicitation: By using a multi-model consensus to define the feature schema, the authors address researcher subjectivity, creating a more robust and objective taxonomy for political media analysis.

The authors explicitly frame their work as analyzing institutional campaign branding and coordinated production values, rather than inferring the private ideology of individuals. They emphasize that the "unspoken visual signatures" identified are properties of the communication strategy itself, offering a method to decode political identity in an era of de-verbalized, multimodal discourse.

Technical Summary: "Bilder ohne Worte: Can AI Identify Political Parties via Unspoken Visual Signatures in the 2025 German Election?"

More like this