From FYP To WW3: Analyzing TikTok ’s Role in Polarizing Political Discourse During the 2025 Nato Summit

Team Members

Facilitators: Miazia Schueler (AI Forensics), Salvatore Romano (AI Forensics), Natalie Kerby (AI Forensics), Giovanni Astante

Participants: Pavel Cihlář, Jan Bakker*, Letizia Sacco, João Guilherme Bastos dos Santos, Gianmaria Avellino, Aliki Livada, Goran Kusic, David Schogt

Designers: Benedetta Riccio, Carla d’Antonio

1. Introduction

On the 24-25 June 2025, world leaders landed in The Hague, Netherlands, for the annual North Atlantic Treaty Organization (NATO) summit. NATO is a political and military alliance of 32 countries from North America and Europe, focused on collective defense and cooperation on security issues. The summit comes at a time of heightened political tension. In early June, Geert Wilder’s party exited the ruling coalition in the Netherlands, sparking an upcoming re-election for fall 2025; the week before the summit, war between Israel and Iran began to escalate; and fear around Russia’s war in Ukraine and what that means for NATO is ever-increasing.

Very large online platforms (VLOPs) are known to play a crucial role in shaping the public debate around social and political themes. With over two billion users, TikTok has become one of the most influential VLOPs. TikTok ’s algorithmically driven feed creates an environment with high potential for virality and increased political influence. As a global, high-stakes geopolitical event, the NATO Summit provided a relevant setting to investigate TikTok 's content recommendation algorithm and its role in shaping local political discourse.

Within the Dutch context, we explore NATO-related content on TikTok and analyze how the platform’s algorithms reinforce possibly polarizing narratives in politically sensitive times. The research focuses on personalized For You Page (FYP) data collected from multiple accounts, both leading up to and after the summit. We explore the data through multiple lenses, including visual, topical, textual, geographic, and time-based narratives, to shed light on the way that TikTok drives content surrounding the NATO Summit. We found that while in many instances the platform provides a diversity of content, there is an overall tendency towards military glorification, further enshrining NATO’s war-bent stance.

2. Research Questions

In this project, we ask: How are NATO-related narratives reflected on TikTok? Given TikTok ’s personalized feed, to what extent are these narratives algorithmically represented to users in the Netherlands, especially given the country’s political state?

3. Initial datasets

3.1 Personalized TikTok FYP Data - Manual Collection

The dataset contains 2699 videos, of which 918 are unique, collected from the FYP of 12 individually personalized TikTok accounts, which were each trained through neutral NATO-related searches. The data was collected from 19 June to 4 July 2025, which covered the time period one week before the summit (6 accounts) and the week following the summit (6 accounts).

Located in the Netherlands, we personalized the account by querying for 5 neutral NATO-related terms: ‘nato,’ ‘nato summit 2025,’ ‘nato news,’ ‘nato summit the hague,’ ‘nato leaders’. These terms were selected for their neutrality, with the goal of understanding how TikTok evolves the videos from a neutral to a polarized stance. The researcher would then scroll the search results for each individual query for approximately 30 seconds without clicking into a specific video.

Once the account was query-trained, we proceeded to scroll the FYP for one hour, watching only videos “relevant” to the project and scrolling past any “unrelated” content. Relevant videos included anything that mentioned or speculated about NATO, its member states, and their current politics, news, military, or protests. Further, we also watched videos that had world war-related themes, military content, and specifically videos about the Dutch political context, as this investigation was taking place within the Dutch context.

3.2 Data Augmentation and Processing

We then extracted the watch history of each account through TikTok ’s data request feature, which we then enriched with video-specific metadata using the TikTok Content Scraper. Furthermore, we extracted transcripts from the videos’ audio and thumbnails. The transcripts were extracted using a Python library developed by AI Forensics. Video URLs were imported into 4CAT, a data collection and analysis tool for multiple online platforms, including TikTok (Peeters & Hagen, 2022). With the aid of 4CAT, we collected 893 video thumbnails from the personalized feed. Images depict a variety of content, including people, infographics, maps, objects, commentaries, legacy media footage, and others.

Transcripts and video descriptions were merged into a single text, which was further processed. The pre-processing pipeline involves extracting hashtags, transforming text into lowercase, stripping punctuation, removing emojis and stopwords, lemmatizing and tokenizing the text using Python’s nltk (Bird et al., 2009) and emoji (Kim & Wurster, 2025) modules. In addition, the processed text was used to extract locations (country names, cities, and geographical entities) using Python’s spacy module (Honnibal et al., 2020). The extracted locations are further inspected and cleaned by two annotators in order to fix mismatches and redundant place-names (e.g., sun, south, world). The processed dataset consists of video-ids, the processed text, as well as the cleaned geographic entities mentioned in the text, and is further merged into the dataset of annotated videos.

4. Narrative Analysis

4.1 Methodology

To determine dominant narratives, we created a codebook with a set of labels to manually annotate the unique videos across the 12 query-trained accounts. We developed the following codebook:

1. NATO-Related: Videos that mention or are related to NATO. This includes videos related to the summit itself, such as security measures, traffic restrictions, arrival of NATO leaders, etc. It could also include discussions on NATO as an organization, such as its history, member state speeches, spending discussions, etc.

2. Military-Related: Videos that display military actions, equipment, missions, or weapons.

3. War Speculations: Videos that speculate about or hypothesize possible war scenarios, or videos that mention World War 3.

4. Current Conflicts:

   1. Russia-Ukraine: Videos displaying or discussing the ongoing war between Russia and Ukraine.

   2. United States-Iran: Videos displaying or discussing the United States’ attack on Iran.

   3. Israel-Gaza-Iran: Videos displaying or discussing the war between Israel, Gaza, and Iran.

5. Generative-AI Speculation (GenAI):

   1. Photorealistic: According to these guidelines, videos we suspect to be genAI aim to mimic real-life visuals with realistic lighting, textures, and details that look like actual footage.

   2. Cartoon: Videos we suspect to be genAI, using stylized, simplified shapes, colors, and exaggeration to resemble animated visuals rather than a depiction of reality.

   3. Non-GenAI: Videos we do not suspect to be non-AI generated.

All main label categories were non-exclusively applicable to each video (Nato-related, Military-related, War Speculations, Current Conflicts, GenAI Speculation), whereas the subdivisions of Current Conflicts and GenAI Speculation were respectively single-choice. Limitations include that what we deem AI-generated remains a speculation. We suspect that some may not necessarily be AI-generated but possibly video game snippets or other types of animations. However, due to the lack of accuracy that genAI detection tools provide, as well as a lack of time, we refrained from the methodological inclusion of any detection tools. However, for further research, we plan on including verification steps for the sake of higher accuracy.

4.2 Findings

In general, the predominant narrative we discovered was military glorification. When an account was trained on NATO-related queries, recommended content was frequently involving praise of army equipment, both in the case of NATO and other specific countries, with the US army taking up a big chunk of this space. Another striking result is the high percentage of AI-generated content in the accounts feed (the maximum value found was 12% for a single account). What follows is a breakdown of codebook categories occurrences across the whole dataset of unique videos.

1. NATO-Related: 15.7%

2. Military-Related: 32.17%

3. War Speculations: 13.41%

4. Current Conflicts:

   1. Russia-Ukraine: 9.39%

   2. United States-Iran: 7.96%

   3. Israel-Gaza-Iran: 5.68%

5. Generative-AI Speculation (GenAI):

   1. Photorealistic: 1.44%

   2. Cartoon: 6.44%

Specifically, the count of Generative-AI content can also be broken down according to the topic-relatedlabels:

1. NATO-Related: 9.71%

2. Military-Related: 29.23%

3. War Speculations: 12.21%

4. Current Conflicts:

   1. Russia-Ukraine: 9.39%

   2. United States-Iran: 7.74%

   3. Israel-Gaza-Iran: 4.47%

The numbers confirm that it is likely that accounts trained on NATO-connoted queries end up in a bubble that is pretty much focused on war. The percentage of military-related content is double that of NATO-related content, and War Speculations' presence in the feed is comparable to that of NATO content. Plus, almost a third of Military-Related videos are AI-generated.

It would be reasonable to expect that these kinds of videos are the ones that get compensated with more views by users, and AI-Generation would be pretty convenient to create engagement with low effort. This is an aspect we quantitatively investigated when doing the text analysis, which also validated the codebook categories by finding word clusters that fit with them.

6. Thumbnail Analysis

6.1 Methodology

Among the quantitative analyses performed, the video thumbnails were used to complement the qualitative work carried out with the narrative annotations. To get an idea of the visual themes found in the video corpus, we used PixPlot (Leonard, 2017/2025), an image analysis tool. Pixplot visualizes images by embedding them in a higher-dimensional vector space where visually similar images appear in proximity. The embedding space is then projected into two dimensions using dimensionality reduction techniques (UMAP in particular).

We used a two-dimensional image plot to interpret clusters based on visual themes. The theme interpretation was conducted collectively among the researchers and was informed by the previous annotation process. Hence, the cluster/theme extraction is conducted in a computational grounded theory framework (Nelson, 2020), whereas interpretations are conducted inductively and grounded in the data. The image analysis complements the micro-understanding of watching and annotating videos.

6.2 Findings

We observe seven main visual clusters. One includes videos where the NATO logo explicitly appears in the thumbnail. Another cluster revolves around maps, in which videos provide historical insights into war and conflicts, along with hypothetical scenarios for future wars or geopolitical analyses. Three clusters focus on different types of military equipment: military carriers and ships, war planes, fighter jets, helicopters, and stealth planes, and military tanks and trucks. Moreover, a cluster is based on video interviews conducted on the street, commonly found in online audiovisual platforms, and another one includes images of police officers and videos of citizen protests. Finally, a cluster features media coverage videos on politicians delivering speeches, conducting press conferences, or giving interviews.

We underline that we observed the presence of women primarily in non-political clusters in the video thumbnails. Furthermore, content depicting women was found to be mainly involved in left-wing political discourse. In general, politicians were found to be almost exclusively men. While partially this reflects the NATO leader gender distribution, it could also point to a relation of war-related content with the manosphere, an online subculture promoting masculinity, anti-wokeness, and conspiratorial discourse. This is further hinted at by thumbnails pointing to conspiratorial thinking and videos depicting cash or stock time series.

Lastly, supporting the previous narrative analysis, we underline that the number of videos depicting military equipment stands out compared to all other clusters. These observations further raise questions on the curation of content on TikTok, particularly in the context of sensitive global political events.

7. Sentiment Analysis

7.1 Methodology

This phase of the project involved the construction of a complete, reproducible pipeline for Sentiment Analysis and Emotion Extraction. The pipeline was applied using a fine-tuned large language model to the descriptions and transcripts (where present) extracted from the TikTok data.

The sentiment classification approach employed was bidimensional (Valence-Arousal) and grounded in the Valence-Arousal-Dominance (V.A.D.) framework, which interprets valence as a value to determine the positivity, neutrality, or negativity of an input and arousal as the degree to which such input is expressed (excited-calm).

More specifically, we utilised a transformer-based, Twitter-finetuned XLM RoBERTa that is currently available on HuggingFace (cardiffnlp/twitter-xlm-roberta-base-sentiment). Such a model allowed us to produce valence and arousal scores for each text input (i.e., description-transcript), which we chose to scale as integers in a range from 1 (i.e., negative, low-intensity) to 9 (i.e., positive, high-intensity), respectively. Hence, and in other words, the computational logic of our pipeline was inherently bi-dimensional, aiming to capture both emotional polarity (i.e., valence: negativity, neutrality or positivity of the input) and the intensity (i.e., arousal: the degree to which the input is expressed) of the affective states reflected in each video’s content.

Prior to the analysis, the textual input extracted from the video description and transcripts columns of the dataset underwent preprocessing. This preliminary phase of the analysis involved multiple steps: a normalisation of Unicode characters (through ftfy), HTML and noise removal (through BeautifulSoup and regular expressions), URL and social media handle stripping, emoji conversion (through the emoji library), punctuation normalisation (while simultaneously preserving more challenging scripts such as Arabic and Cyrillic), and spacing normalisation. Moreover, we implemented language-specific lemmatisation through spaCy libraries tailored for each detected language, including English, Spanish, French, Italian, Arabic, Dutch, German, and Russian. Arabic texts also underwent additional normalisation through the removal of diacritics and Tatwil characters, which ensured a better linguistic consistency.

The language detection process was pursued through the use of the standalone Language Identification tool named LangID, and the texts were subsequently routed to their respective spaCy language models. Emoji-based emotions, which were common in our data, were instead explicitly identified and directly mapped onto valence-arousal coordinates. In addition, the model’s tokenisation limit (i.e., 512 tokens) was bypassed through the automatic segmentation of the input data into manageable chunks, individually analysed and aggregated through the average of their resulting scores.

After the preprocessing, we thus proceeded to obtain valence-arousal coordinates for each entry, and further classify these coordinates into complex emotional categories through a clustering method (KMeans with seven clusters). Each cluster corresponds to an emotion: Gratitude, Admiration, Anger, Fear, Hope, Disappointment, Joy. These labels were further assigned based on the centroid positions of clusters in the valence-arousal space, and then utilised for a deeper interpretive investigation.

In conclusion, we generated additional visualisations and analysed, including t-SNE projections (through sklearn.manifold.TSNE) and inter-emotion distance maps using classical Multidimensional Scaling (MDS). This allowed us to examine the semantic and affective relations between and among emotions in more detail and accuracy. Future directions aimed at enhancing the pipeline for greater accuracy may involve a deeper grounding of the classification approach in theories of appraisal as well as an additional account of complex communicative forms such as irony and sarcasm, a multimodal extension through computer vision applications in video-sentiment analysis and an adhoc fine-tuning of the language model through Low-Rank Adaptation methods (LORA adapters).

7.2 Findings

Key findings regarding emotional dynamics in the data were found particularly in the relative temporal distribution of the emotions between pre- and post-summit. The distribution was as follows: gratitude (1781 entries, 38.7%), joy (858 entries, 18.7%), hope (758 entries, 16.5%), admiration (394 entries, 8.6%), disappointment (351 entries, 7.6%), fear (266 entries, 5.8%), and anger (195 entries, 4.2%). Crucially, when comparing emotional shifts between pre-summit and post-summit phases, we explicitly accounted for relative weights by calculating each emotion as a percentage of the total number of posts in each respective phase. This allowed us to measure shifts accurately.

Admiration slightly decreased from representing 9.1% of pre-summit posts to 8.2% post-summit, marking a moderate decrease of 0.9 percentage points. Thus, the post-summit context showed slightly less emphasis on Admiration compared to before the summit. Anger, similarly, increased from 3.6% pre-summit to 4.7% post-summit, a rise of 1.1 points, reflecting increased negative, high-arousal sentiment after summit events. Similar increases in 1.2 and 1.3 points affected disappointment and fear, respectively, which may indicate moderately unmet expectations as well as heightened expressions of concern and anxiety amongst the user base in response to the summit’s outcomes. Minimal shifts, specifically of +0.2 and -0.3, respectively, affected hope and joy. The biggest shift, however, was found in the expression of gratitude, which declined from 40.2 to 37.7 points after the summit, a 2.5% reduction that allows us to observe a less positive sentiment proportionally post-summit.

8. Sound Analysis

8.1 Methodology

In order to better understand the affective and rhetorical dimensions of TikTok ’s NATO-related content, we conducted a qualitative auditory analysis of the videos included in our dataset. This involved a close reading approach aimed at identifying trending sounds, vocal characteristics, and sound design choices. This method privileged a human-centered assessment, instead of automated audio classification tools, aimed to capture how the music choices may contribute to the emotional tone of the content.

8.2 Findings

Our analysis revealed a consistent use of distinct audio aesthetics in TikTok videos related to war and geopolitics. More specifically, we observed in our dataset that content related to war and geopolitical tension circulated predominantly with slowed and reverb audio, loud and dramatic soundtracks, and most notably, AI-generated male voices, typically deeper than standard text-to-speech outputs.

This stands in contrast to the prevalence of female voices in other platforms like Siri, Google Maps, or Alexa, where women’s voices have historically been preferred for their “helpful,” less authoritative tone. In TikTok ’s geopolitical content, however, the deep male voice adds gravitas and institutional authority, aligning more closely with traditional performances of power. The slowed and reverbed aesthetic further deepens vocal tone and injects emotional weight into otherwise straightforward content.

It is important to note that these stylistic patterns likely emerge as a product of the platform’s structural and cultural affordances. TikTok heavily emphasizes audio through both its editing tools and its recommendation system, which tends to reward content that reuses and builds upon trending sounds. TikTok can be seen as an affect-centered platform, where the emotional resonance of content becomes a key metric. The algorithm recommends what performs well, but what performs well is shaped by what the format encourages. In politically sensitive times, such as during coverage of war or conflict, what rises to the top is content that is most potent affect-wise, often made powerful and viral through sound.

9. Geographical Analysis

9.1 Methodology

Building on the location extraction, we counted how many times each location appeared in the corpus using Python. The resulting 201 individual locations were geocoded using the geopandas Python library (Jordahl et al., 2020), enabling spatial visualisation and analysis. Utilising ArcGIS Pro software, we calculated the aggregate mention count for all locations within each country, along with the weighted mean sentiment scores. Location mentions count served as weights for sentiment calculations, allowing us to identify not only the most frequently mentioned countries and regions but also the associated sentiment valence for each geographic entity.

9.2 Findings

The geographical analysis provides some important insights into which parts of the world are visible and which remain hidden in users' FYPs. Regarding our central question about how NATO-related narratives are algorithmically represented to Dutch users, the geographic data reveals a highly selective worldview.

Country names dominated the geographic references, with the USA (1,821 mentions), Russia (1,190 mentions), and Ukraine (791 mentions) appearing most frequently. Specific cities and regions appeared primarily in two contexts: locations related to the Gaza conflict (e.g., Gaza, West Bank, Jerusalem) and Dutch cities hosting or connected to the NATO summit (e.g., The Hague, Rotterdam, Amsterdam). When aggregated at the country level, the USA remained dominant (1,975 mentions), followed by the Netherlands (1,349 mentions).

More striking than what is mentioned is what remains absent from the discourse. While the focus on the USA, Russia, Ukraine, and the Netherlands aligns with expectations (Given that these videos were shown to Dutch users interested in NATO content), the stark disparity in coverage is notable. NATO's activities in the Baltic and Balkan regions receive minimal attention (119 and 75 location mentions, respectively). Beyond the Gaza conflict, other ongoing global conflicts remain virtually absent from the discourse, with negligible mentions of crises in Congo, Sudan, or Myanmar. This geographic selectivity reduces the NATO-related discourse on TikTok to a simplified narrative centred on the USA, Russia, Ukraine, and the Gaza conflict, a pattern with significant implications for users' geopolitical understanding.

The weighted sentiment analysis reveals predominantly negative sentiment across most countries (analysis limited to countries with >50 mentions for reliability). This pattern may reflect methodological limitations: TikTok 's multimodal nature means that the audio of the video often carries significant emotional weight not captured in our text-based analysis. Military content accompanied by nationalist or triumphant music can, for example, register as negative in the textual sentiment analysis. Nevertheless, the finding that Russia appears in less negative contexts than Ukraine warrants further investigation. An interesting spatial pattern emerges wherein countries geographically proximate to the Netherlands tend toward more positive sentiment scores, though this correlation requires more rigorous quantitative validation.

These findings demonstrate that NATO-related content on TikTok exhibits significant geographic unevenness. Certain regions receive disproportionate attention while others remain largely invisible. This selective geographic framing likely shapes users' understanding of global geopolitical dynamics and NATO's international role.

10. Text Analysis

10.1 Methodology

For the text analysis, we used original codes in R languages to operate hierarchical clustering techniques both to clean the text data from TikTok transcriptions and to identify (i) the most important perspectives and vocabularies (lexical clusters); (ii) how each vocabulary (clusters) is related to specific attention and engagement metrics; and (iii) if the vocabularies identified automatically match totally or partially the human tagging made during the Summer School. The hierarchical clustering is represented in a structure similar to a “genealogical tree” of vocabularies called a dendrogram. The dendrogram shows how clusters are related; the closer the connection among branches, the more lexically similar, showing different levels of “kinship”. For each cluster, we have the words most characteristic of its vocabulary (measured by chi2, represented by “n”) and the proportion of the cluster’s tokens (“%”) considering all tokens of the corpus.

Considering the following dendrogram, the online conversation surrounding the NATO meeting in the Netherlands can be seen as a convergence of three major branches, encompassing six lexical clusters. The three branches are social/political axis, local and ceremonial strand and geopolitical-military dimension.

Dendrogram 1

The first branch is a social/political axis (clusters 1, 2, 3) intertwining discussions about global leaders, humanitarian crises, and public opinion. Here, figures like Donald Trump, Erdoğan, and Meloni are invoked alongside concerns about Gaza, children, and alleged war crimes. The tone ranges from formal political critique to casual, opinion-driven commentary, suggesting a polarized yet interconnected discourse.

Wordclouds 1 and 2

Second, a local and ceremonial strand (cluster 4) captures the Netherlands context, marked by references to The Hague, the King of the Netherlands, and NATO’s meeting in the country. These posts often appear in Dutch and emphasize the symbolic and protocol elements of hosting the summit.

Wordclouds 3 and 4

Finally, a geopolitical-military dimension (clusters 5 and 6) frames NATO’s activities in terms of external threats, spending, and defense commitments. One side of this narrative (Cluster 5) focuses on external adversaries such as Russia, Iran, and China, emphasizing nuclear and missile capabilities. The other side (Cluster 6) highlights the summit’s organizational aspects — defense spending, collective security, and alliance management.

Wordclouds 5 and 6

Each of these lexicons are related with specific engagement ranges, being cluster 5 (the military debate around geopolitical threats, Russia and the possibility of nuclear war, bringing nuclear, iran, russia, china, missile, strike, aircraft, bases as the most representative tokens) the most impactful, having the highest engagement in all measures (3.28 billions plays and lading in likes, shares and comments). A closer look show that it brings (i) aircraft, bombing, troops, (ii) billion euros contracts involving weapons and its costs, (iii) geopolitical topics around Iran, Russia, Ukraine, Israel, United States, (iv) prime ministers and presidents discussing attacks that violate international law and humanitarian agreements, (v) BRICS and global market dominance, (vi) Russian and OTAN tensions and fears involving troops in the borders of the war against Ukraine.

Cluster 3 (the colloquial/opinionated conversation bringing dont, yeah, think, gonna, like, really, see, good as representative tokens) also appear as an important one, and a loser look in the corpus shows that beyond the overall colloquial discussion about NATO and public opinion, it involves a discussion about (i) the increase in defense spending, its impact on trade and position of different countries about it (around 10% of cluster 3 tokens) and also (ii) tensions between US and Russia around the war in Ukraine (around 19% of cluster 3 tokens). The colloquial debate involves topics related to the NATO Summit, either international, like hate, nuclear weapons, racism, or local, involving passports, the police, buses, etc. It has the second-highest engagement in all metrics (with 1.55 billion views).

Heatmap 1

The same posts were qualitatively tagged by researchers during the Summer School [categories being Contemporary Politics, History, 2025 NATO Summit the Hague, Conspiracy, History, Israel-Iran-Palestine, Mentions WW3, Mentions NATO, Military-weapons, NATO Spendings, NATO Summit safety measures, Previous NATO Summit, Propagandist, Protest, Ukraine-Russia, US-Iran, War Speculations]. The following heatmap brings the same lexical clusters, crossing them with the human-tagged categories.

Heatmap 2

In the heatmap, cluster 1 shows stronger connections to the “Contemporary Politics”, “2025 NATO Summit The Hague”, and “Mentions NATO” categories, pointing to the relevance of leaders related to the Summit in contemporary politics debate. It is also possible that these discussions are tied to geopolitical and diplomatic disputes involving these leaders and their interactions within international forums. Clusters 2 and 3 are also strongly related to “Contemporary Politics” (1st) and “Mentions NATO” (3rd), but also strongly connected to “Military/Weapons” (2nd). It makes sense considering the focus on contemporary humanitarian crises, military conflicts, bombings, genocide, taxes, and colloquial discussion. The similarity among categories attributed to posts in clusters 2 and 3 can be understood by the fact that they are in the same branch, showing coherence.

Cluster 4 contains terms clearly pointing to the Netherlands, the “2025 NATO Summit in The Hague,” and the “Summit safety measures.” The presence of Dutch-language terms alongside references to NATO leaders suggests coverage targeted at not just international but also domestic audiences. This cluster likely represents local reporting on the summit, including logistical details, political statements, and mentions or official appearances by Dutch figures such as King Willem-Alexander.

Cluster 5 includes words which link it specifically to war-related military debates, combining “Military-weapons” (1st), “Contemporary Politics” (2nd), and “War Speculations” (3rd). This cluster is centered around military escalation, advanced weapon systems, and global security threats, strategies, and possible scenarios. On the other hand, cluster 6 posts are tagged with categories “Mentions NATO”, “Contemporary Politics”, and “2025 NATO Summit in The Hague”. This cluster reflects discussions about financial commitments and military budgets within the alliance, as well as political debates over how resources should be allocated. It may also capture debates within national politics about the costs and benefits of NATO membership, its costs and its expectations towards the 2025 Summit.

10.2 Findings

Together, the three layers of text analysis (vocabularies, engagement, and relation with categories) reveal a multi-layered debate: posts related to war speculations whose vocabulary was identified in the hierarchical clustering as cluster 5 and main topics confirmed by human tagging are clearly more willing to attract engagement and be more successful in spreading through TikTok. It is possible that the presence of sensationalist or alarmist narratives about a possible large-scale global conflict, or curiosity around new powerful weapons, drives this success. Its “twin” vocabulary, also relevant (3rd biggest metrics), cluster 6 brings discussion about NATO spending, management, and meetings. Together, they compose what we identified as the geopolitical-military branch.

The debates essentially centered on political leadership (cluster 1), children, bombing, war crimes, genocide (cluster 2), and the national Dutch debate about local measures and the impact of the Summit (cluster 4) are set aside by this branch. The only one coming closer to cluster 5 is the more colloquial debate around NATO-related topics (cluster 3) putting together diverse posts from small talk expressions to NATO spendings, and even this cluster could not reach half of the plays of the former and a bit more of half of the comments (though coming close when it comes to likes and shares).

Cluster 5 stands as a clear proxy for “War Speculations”, “WW3” and “Military/Weapons” categories, concentrating more posts in this category than all other clusters combined (and even in this scenario having a surplus of at least 10 posts), at the same time that it is behind clusters 2 and 3 when it comes to “NATO Spendings” and “Conspiracy”, behind clusters 3, 4 and 2 when it comes to “NATO Summit Safety Measures”, behind clusters 2 and 4 when it comes to “Previous NATO Summits”, nearly tied with cluster 1 in the last position when it comes to mentioning “2025 NATO Summit The Hague” and not having a single post about “Protests”.

Therefore, war speculations are at the center of users’ attention (plays and comments). This group of posts can be automatically identified using text analysis in post transcriptions, a method also able to separate war speculations from discussions about humanitarian crises, protests, spending, or even political leadership. NATO itself stays out of focus when not connected to the possibility of a new world war.

11. Discussion & Conclusions

This research set out to investigate how NATO-related narratives are represented and algorithmically reinforced on TikTok during a period of heightened political tension surrounding the 2025 NATO Summit in The Hague. Recognizing TikTok ’s growing role as a VLOP with significant influence over political discourse, particularly in moments of geopolitical sensitivity, we examined how the platform’s recommendation system reflects and potentially amplifies dominant framings of NATO, especially within the Dutch context. To do so, we employed a multi-method analytical framework, combining narrative analysis, thumbnail analysis, sentiment analysis, and geographical analysis. Together, these approaches allowed us to unpack not only what kinds of content circulated on TikTok in relation to NATO but also how it was framed through algorithmic exposure on the FYP.

From a narrative perspective, a fitting example of a political narrative in this context would be the framing of NATO as a peacekeeping organization, since NATO is an alliance that presents itself as being positioned around the maintenance of global stability and security. However, the predominant narrative that emerged from our analysis was one of military glorification. When user accounts were exposed to NATO-related queries, the recommendation algorithm tended to prioritise content that highlighted and praised military equipment, both from NATO forces and individual countries, with a particular emphasis on the U.S. Army.

The thumbnail analysis revealed seven distinct clusters in how NATO-related content was visually presented. One cluster featured thumbnails with the NATO logo; another revolved around maps explaining historical or hypothetical war scenarios. Three major clusters focused on military hardware: (1) carriers and ships, (2) aircraft and helicopters, and (3) tanks and trucks. Other clusters included man-on-the-street interview formats, protest imagery involving police presence, and political media coverage (e.g., speeches and press conferences). Notably, women were predominantly present in non-political clusters, while female content creators were mainly linked to left-wing discourse. Politicians, by contrast, were almost exclusively male, suggesting a potential connection between war-related content and the manosphere, a digital subculture promoting hypermasculinity and conspiratorial thinking.

The sentiment analysis employed a bi-dimensional Valence-Arousal classification approach to compare video content before and after the NATO Summit. Upon closer examination of emotional trends over time, notable patterns were observed, particularly in videos uploaded between January 6th and 7th. For instance, expressions of gratitude, which appeared in over 200 posts prior to the 6th, sharply declined to slightly above 50 on that day, before rising again on the 7th to nearly 200 posts. Although on a smaller scale, similar fluctuations were evident across all labelled emotions included in the analysis. These variations suggest that emotional responses were highly event-driven rather than grounded in stable, value-based perspectives. In other words, users’ reactions appeared to be shaped by external, time-specific triggers rather than consistent ideological positions. A particularly sharp decline in positive emotions such as admiration and gratitude following the summit may reflect unmet public expectations. This shift seems to align with the nature of the trending content during the summit period, which largely focused on sensational or superficial topics, such as speculation regarding Meloni’s behavior or the Dutch Queen mocking US President Donald Trump. The prominence of such content likely redirected public attention away from substantive political discourse, influencing the overall emotional tone of engagement.

The sound analysis offers a deeper understanding of this affective dynamic. Content related to war and geopolitical tension frequently employed slowed + reverb audio, cinematic soundtracks, and notably, AI-generated deep male voices. These elements added a sense of institutional gravitas and emotional weight to content that might otherwise seem neutral or mundane. This reflects TikTok ’s structural prioritization of sound, both in its editing affordances and in the algorithmic promotion of reusable audio. As a result, emotionally resonant and often dramatized content is more likely to gain traction, especially in politically charged contexts.

Regarding the geographical analysis, our exploration involved counting how often each location appeared in the dataset and calculating average sentiment scores for each country using spatial tools (like geopandas and ArcGIS Pro). The findings pointed to a strong Western-Atlantic bias in public discourse, with the U.S. and Russia receiving the most attention. The Israeli invasion was also heavily featured. In contrast, other geopolitical crises, such as those in Congo or Myanmar, as well as NATO’s activities in the Baltic and Balkan regions, were barely mentioned. Such characteristics in the data allow us to observe that public opinion's attention significantly reflected the specific topics, subjects, and crises covered in the Summit. As a consequence, the public's interest disproportionately promoted geopolitically dominant worldviews, reaffirming historical power imbalances in international relations.