💬 NLP

Moral Outrage Shapes Commitments Beyond Attention: Multimodal Moral Emotions on YouTube in Korea and the US

This study utilizes a fine-tuned vision-language model to analyze 400,000 YouTube news videos from Korea and the US, revealing that moral outrage expressed through other-condemning rhetoric consistently drives higher levels of audience engagement across all commitment levels, from passive viewing to active commenting, thereby highlighting its effectiveness while raising concerns about its potential to deepen societal polarization.

Original authors: Seongchan Park, Jaehong Kim, Hyeonseung Kim, Heejin Bin, Sue Moon, Wonjae Lee

Published 2026-02-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Seongchan Park, Jaehong Kim, Hyeonseung Kim, Heejin Bin, Sue Moon, Wonjae Lee

Original paper licensed under CC BY 4.0 (http://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

Imagine the internet, specifically YouTube, as a giant, noisy marketplace. In this marketplace, the most valuable currency isn't money; it's human attention. News channels are like street vendors trying to shout the loudest to get people to stop, look, and interact with their wares.

This paper is a study by researchers from KAIST in South Korea who wanted to understand: What kind of "shouting" actually gets people to not just look, but also buy (like) and argue (comment)? They looked at news videos from both the United States and South Korea to see if the rules of the game are the same in both places.

Here is the breakdown of their findings, using simple analogies:

1. The Toolkit: Reading Between the Lines (and Pictures)

Most previous studies only read the text (the video titles), like reading a newspaper headline without looking at the photo. But on YouTube, the thumbnail image and the title work together like a movie poster and its tagline.

The Problem: Humans are great at seeing a picture of an angry mob and reading a title about "corruption" and instantly feeling moral outrage. Computers usually struggle to connect the dots between the image and the text.
The Solution: The researchers built a special "digital detective" (an AI model). They taught it by showing it thousands of examples of YouTube thumbnails and titles, labeling them with human help. They taught the AI to spot moral emotions—specifically, feelings about right and wrong.
The Result: They created a tool that can look at a video's cover and title and say, "This is full of moral outrage," "This is about suffering," or "This is neutral." They tested this on nearly 400,000 videos.

2. The Three Levels of Engagement

The researchers looked at three ways people interact with a video, which they compared to levels of commitment:

Views: Just walking by the vendor's stall and glancing at it (Passive).
Likes: Stopping to give a thumbs-up or a nod (Moderate commitment).
Comments: Stopping to start a conversation or a fight (High commitment).

3. The Big Discovery: The "Outrage Engine"

The study found a very clear pattern that was true in both Korea and the US.

The "Outrage" Fuel: When a video used other-condemning rhetoric (language that scolds, shames, or gets angry at someone else's bad behavior), it was like pouring gasoline on a fire.
- It got more views.
- It got even more likes.
- It got the most comments.
- The Analogy: Imagine a vendor yelling, "Look at that thief over there! He's terrible!" People don't just glance; they stop, point, and start arguing with each other about the thief. The anger makes people want to do something, not just watch.
The "Praise" and "Sadness" Effect:
- Videos that praised people (e.g., "Look at this hero!") or focused on suffering (e.g., "Look at this poor victim") generally did not get the same high levels of engagement. In fact, in many cases, they actually got fewer comments and likes than neutral videos.
- The Analogy: A vendor saying, "Look at this nice guy," might get a polite nod, but it rarely gets people shouting or arguing. It's too calm to spark a frenzy.

4. The Cross-Cultural Connection

The researchers were surprised to find that this "Outrage Engine" worked exactly the same way in both South Korea and the United States. Even though the cultures, languages, and specific news stories were different, the human reaction to moral outrage was identical. Whether it was a political scandal in Seoul or a trial in Washington, the anger made people engage more deeply.

5. The Warning Label

The paper concludes with a cautionary note. While moral outrage is a very effective tool for getting attention and engagement, it comes with a risk.

The Danger: Because outrage drives people to form "us vs. them" groups (the "good guys" vs. the "bad guys"), it can deepen divisions in society. If news outlets realize that "yelling about bad guys" gets them the most views and comments, they might be tempted to use more angry, divisive language just to stay popular.
The Takeaway: The study doesn't say this will happen, but it warns that the "attention economy" naturally rewards the loudest, most angry voices, potentially turning the marketplace into a battleground.

Summary

In short, this paper built a smart tool to read YouTube videos and found that anger at others is the most powerful magnet for human attention. It makes people watch, like, and argue more than any other emotion, and this rule applies whether you are in Korea or the US. The researchers warn that while this is great for getting views, it might be bad for keeping society united.

Technical Summary: Moral Outrage Shapes Commitments Beyond Attention

Problem Statement
In the attention economy, social media platforms prioritize content that generates high user engagement. While prior research establishes that moral-emotional content drives participation and polarization, significant gaps remain in understanding these dynamics. Existing studies predominantly rely on unimodal, text-based analyses, failing to capture the inherently multimodal nature of platforms like YouTube where thumbnail images and titles jointly influence perception. Furthermore, most research is concentrated on Western contexts, leaving the cross-cultural generalizability of moral-emotional expression and engagement unclear. This study addresses these gaps by examining how multimodal moral-emotional framing by mainstream news channels influences user behavior across South Korea and the United States.

Methodology
The research employs a computational social science framework involving data collection, multimodal modeling, and regression analysis.

Data Collection: The authors constructed a dataset comprising approximately 400,000 YouTube videos (292,136 from Korea, 105,761 from the US) uploaded between January and December 2024. The data includes 7 Korean and 16 US news channels, selected for political diversity. For each video, the dataset captures thumbnail images, title text, and engagement metadata (views, likes, comments), with metadata retrieved during a fixed eight-day window to ensure comparability.
Human Annotation: To train the models, a gold-standard dataset was created through a multi-stage sampling strategy. Native speakers annotated 1,276 Korean and 900 English thumbnail-title pairs into six categories: other-condemning (moral outrage), other-praising, other-suffering, self-conscious, neutral, and non-moral emotion. Inter-annotator agreement was measured using Cohen's, Fleiss', and Krippendorff's kappa, yielding higher agreement for Korean data (avg. $\kappa \approx 0.75$ ) than English data (avg. $\kappa \approx 0.53$ ).
Multimodal Modeling: The study developed binary classifiers for each of the six moral emotion categories using Multimodal Large Language Models (MLLMs). The authors evaluated both closed-source (Gemini-2.5-Flash-Lite) and open-source (Qwen2-VL-7B-Instruct) models under zero-shot, few-shot, and fine-tuning settings. The fine-tuned Qwen2-VL-7B-Instruct model achieved the best performance, particularly for the Korean dataset (Average Accuracy: 0.8981, Macro F1: 0.8010). The study validated that incorporating visual cues significantly outperformed text-only baselines.
Regression Analysis: The predicted moral emotion probabilities were used as independent variables in negative binomial regression models to predict three dependent variables representing increasing levels of commitment: views, likes, and comments. Models controlled for video duration, channel baseline, month, and weekday.

Key Results
The analysis reveals a consistent cross-cultural pattern regarding the impact of moral emotions on engagement:

Other-Condemning Rhetoric: Expressions of moral outrage (criticizing others' morality) consistently increased all forms of engagement in both countries. Crucially, the effect size strengthened as the level of commitment increased. In the US, other-condemning rhetoric was associated with a 94% increase in views, 128% increase in likes, and 204% increase in comments (Incidence Rate Ratios: 1.94, 2.28, 3.04). Similar, though slightly less pronounced, trends were observed in Korea.
Other Emotions: Other-praising rhetoric generally suppressed engagement, reducing views and comments. Other-suffering showed mixed results, increasing views in Korea but decreasing likes and comments; in the US, it showed consistently negative associations across all metrics. Neutral content generally reduced engagement, though it increased views in the US while reducing likes and comments.
Commitment Gradient: The study confirms that moral outrage is unique in its ability to drive not just passive attention (views) but also active participation (likes and comments), with the strongest impact observed on commenting behavior.

Key Contributions

Methodological Innovation: The authors developed and publicly released the first multimodal moral emotion classifiers for both Korean and English contexts. These models leverage fine-tuned MLLMs to process thumbnail images and titles jointly, demonstrating that visual-textual integration is essential for accurate moral emotion detection in digital media.
Cross-Cultural Empirical Evidence: The study provides robust evidence that the dynamics of moral-emotional engagement generalize across distinct cultural settings (Korea and the US). It identifies other-condemning rhetoric as a universal driver of high-commitment engagement.
Theoretical Insight: By mapping moral emotions to specific engagement levels, the research refines the understanding of the attention economy, showing that moral outrage functions as a powerful motivator for active participation, not just passive consumption.

Significance and Claims
The paper claims that moral outrage is a particularly effective emotional strategy for media outlets, capable of securing both attention and active participation. However, the authors caution that this effectiveness carries significant societal risks. Because other-condemning rhetoric inherently divides audiences into in-groups and out-groups, its amplification through engagement algorithms may deepen political polarization. The authors suggest that malicious actors and misinformation networks could exploit this mechanism to diffuse false or polarizing narratives.

The study concludes that while moral emotions are fundamental to social coordination, the specific mechanism of moral outrage in the digital attention economy requires careful monitoring. The authors propose that future work should investigate how platforms can mitigate excessive outrage without undermining accountability and how to integrate these detection capabilities into frameworks for identifying trolls and misinformation. The paper maintains a modest stance, acknowledging that as an observational study, it identifies correlations rather than causality, and notes limitations in annotation reliability for English data and potential biases in pre-trained models.