Journal article · Journal of Computational Social Science · June 2025 # The face of populism: examining differences in facial emotional expressions of political leaders using machine learning Sara Major¹, **Aleksandar Tomašević**¹ ¹ Department of Sociology, University of Novi Sad - [Paper](https://doi.org/10.1007/s42001-025-00392-w) - [Preprint](https://arxiv.org/abs/2304.09914) - [Code](https://github.com/atomashevic/face-of-populism) - [Annotation data](https://github.com/atomashevic/fop-annotation) - [Cite](#cite) **In short:** We ran a facial emotion recognition model over YouTube videos of 44 political leaders from 15 countries. In videos of leaders from populist parties, negative emotions took up a larger share of the face and neutral expressions a smaller one. The gap was widest between strongly populist and strongly pluralist parties. [](https://atomasevic.xyz/images/papers/face-of-populism-figure.webp) **One dot per video.** Average share of negative emotion (left) and neutral expression (right) across the frames of each video, for leaders of populist and pluralist parties. Figure 2 of the paper. ## Why look at politicians' faces? Populist rhetoric is often described as heated and full of strong, mostly negative emotion. Most of the evidence for that comes from words: speeches, posts and tweets. Text is hard to compare across languages, and most sentiment tools work best in English. Faces travel better. Party channels on YouTube are full of well-lit close-ups of leaders, and these videos show how leaders choose to present themselves. We asked whether leaders of populist parties look different from leaders of pluralist parties when they do. ## What we did We used the Global Party Survey, a 2019 expert survey of parties worldwide, to place parties on a scale from strongly pluralist to strongly populist. In each of 15 countries with competitive elections, from Argentina to the United States, we picked three parties with different levels of populist rhetoric and one leader for each. For every leader we took five videos from the official party or leader channel: speeches, press conferences and ads, with the face in view and no editing by news outlets. - **44** political leaders - **15** countries - **220** YouTube videos - **300** frames per video From each video we took 300 evenly spaced frames. An open-source pipeline in Python found the leader's face in each frame and passed it to a small neural network trained to recognise six emotions (anger, disgust, fear, happiness, sadness, surprise) and a neutral expression. For each frame the model returns seven scores that add up to one. We summed anger, disgust, fear and sadness into one negative emotion score and averaged everything over the frames of each video. Seventeen videos had several faces the pipeline could not tell apart, so the analysis uses 203. To check the model, five coders labelled 514 random frames as positive, negative or neutral. The coders often disagreed with each other. Where a majority label existed, the model matched it in 55% of frames, and 60% when the face was clearly visible. ## What we found **Negative emotion dominated overall.** Across all videos, negative emotions made up more than half of the average facial expression. Anger had the highest average score, followed by sadness and fear. **Populist leaders showed more of it.** In videos of populist leaders, negative emotion averaged 0.62 of the expression, against 0.50 for pluralist leaders. Negative emotion was the dominant expression in 73.8% of populist videos and 45.8% of pluralist ones. **Pluralist leaders looked neutral more often.** The average neutral score was 0.37 for pluralist leaders and 0.29 for populist leaders. This difference was smaller than the one for negative emotion. - Negative emotions, populist leaders: 0.62 (95% confidence interval 0.584 to 0.649) - Negative emotions, pluralist leaders: 0.50 (95% confidence interval 0.464 to 0.537) - Neutral expression, populist leaders: 0.29 (95% confidence interval 0.26 to 0.318) - Neutral expression, pluralist leaders: 0.37 (95% confidence interval 0.338 to 0.409) Each video gets the average score across its frames; dots are group means of those averages, lines are 95% confidence intervals. When we split parties into four groups, the typical video moved steadily toward more negative emotion from strongly pluralist to strongly populist. The clearest differences were between strongly populist leaders and both pluralist groups for negative emotion, and between strongly pluralist leaders and both populist groups for neutral expression. These differences were larger than the differences we found between ideological groups. ## What it does not show The sample is a snapshot of 44 leaders and is not globally representative. Only five leaders are women, so we could not look at gender. The populism measure comes from a single expert survey that stops in 2019. The pattern was not the same in every country. We chose videos partly by view count, so audiences that prefer angrier videos could shape what we see. The emotion model is modest: about 66% accurate on its standard benchmark, with moderate agreement with our coders, and it may be more sensitive to negative expressions than to positive ones. Averaging over 300 frames per video softens the effect of single misread frames, but cannot remove it. ## Why it matters The result matches what text-based studies report about populist communication, and it comes from a source that does not depend on language. It also shows that this kind of study can run on free tools that individual researchers can use without institutional computing support. The code, the list of videos, the frame-by-frame emotion scores and the human-labelled frames are all public, so others can test better models on the same material or extend it to more leaders, platforms and years. Abstract Populist rhetoric employed in online media is characterized as deeply impassioned and most often imbued with strong emotions. This paper investigates the differences in affective non-verbal communication of political leaders. We use a deep-learning approach to process a sample of 220 YouTube videos depicting political leaders from 15 different countries, analyze their facial expressions of emotion, and then examine differences in average emotion scores representing the relative presence of six emotional states (anger, disgust, fear, happiness, sadness, and surprise) and a neutral expression for each frame of the processed YouTube video. Based on a sample of manually coded images, we find that this machine learning approach has 53−60% agreement with human annotation. We observe statistically significant differences in the average score of expressed negative emotions between groups of leaders with varying degrees of populist rhetoric. Overall, our contribution provides insight into the characteristics of non-verbal emotional expression among political leaders, as well as an open-source workflow for further computational studies of their affective communication. ## Cite Major, S., & **Tomašević, A.** (2025). The face of populism: examining differences in facial emotional expressions of political leaders using machine learning. *Journal of Computational Social Science*, 8(3), 62. ``` @article{Major_2025, title={The face of populism: examining differences in facial emotional expressions of political leaders using machine learning}, volume={8}, ISSN={2432-2725}, url={http://dx.doi.org/10.1007/s42001-025-00392-w}, DOI={10.1007/s42001-025-00392-w}, number={3}, journal={Journal of Computational Social Science}, publisher={Springer Science and Business Media LLC}, author={Major, Sara and Tomašević, Aleksandar}, year={2025}, month=June } ``` ## Related - [Dynamic exploratory graph analysis of emotions in politics](https://advances.in/psychology/10.56296/aip00021/) (A later paper with Sara Major on emotions in politics, using dynamic exploratory graph analysis.) - [transforEmotion: An Open-Source R Package for Emotion Analysis Using Transformer-Based Generative AI Models](https://atomasevic.xyz/papers/transforemotion.html) (Computational Communication Research, 2026) - [A Latent Oscillator Measurement Model to Simulate Emotional-Expression Score Dynamics in Video](https://atomasevic.xyz/papers/latent-oscillator.html) (arXiv, 2026) - [All publications](https://atomasevic.xyz/publications.html#2025-the-face-of-populism)