§01· Battle of ideas

The beliefs that won the war

Behind the tug of war, ideas fought for attention. We read 327 viral posts one by one, in eight languages, images and videos included, and mapped what each one asks you to believe. A joke and a statistic can carry the same belief: the joke is only the packaging. Each bubble is a belief. Its size is the attention it earned.

§02· The map

Where each belief stands

Battle lines: beliefs serving the French side on the left, the Japanese side on the right, neutral ones in the middle. Lines are quotes: one post answering another.

§03· What we found

Seven lessons from the battlefield

§04· Duels

When a post answers another

A quote is a duel: someone takes a post and answers it with another one. The ratio is the reach of the answer divided by the reach of the original. Above 1, the answer won.

§05· Ranking

Which beliefs win their duels

An Elo rating, as in chess. Every rebuttal between two beliefs is a match. The belief whose post reached more people wins the match. Duels inside one belief (support against attack) are counted in the lessons above.

§06· Third parties

Who else was fighting

§07· Origin

How one cosplay started it

1. 17 Oct

   High-school protests in France turn into riots and looting in several cities.

2. 18 Oct

   A French post shows a Naruto cosplayer posing at a burning car, captioned as an anime fire attack. About 2.3M views.

3. 18 Oct

   A Japanese account answers: stop soiling Japanese culture to do stupid things. 2.3M views.

4. 18 Oct

   French answers: withdraw every anime inspired by Europe (5.1M views), and a list of Japanese "vices" that becomes the script of the jokes to come.

5. 19 to 23 Oct

   Japanese X frames the riots as a Muslim invasion. On 23 Oct a Japanese post cites a Justice Ministry note: 70% of those jailed are French nationals, nearly all born in France.

6. 21 to 25 Oct

   The French joke wave: Vegeta is Algerian (17.9M), sperm and sex dolls (11.5M), the 65-hour week (10.1M). X auto-translates every one of them, and the English-speaking world watches.

§08· Method

How this was measured

1. Corpus. 59 seed posts (every tracked post above 500K views, plus posts found by hand), 146 posts found by searching X in eight languages (three passes: 500K views and more, then 250K, then 100K for the anime-as-a-protest-symbol idea), 8 posts they quote, 2 origin posts seen in screenshots, and 115 quotes read in the Quotes tab of the 15 most viewed posts. 327 posts, about 356M views. Twelve viral posts were excluded after a relevance audit: off topic or camp unclear.
2. Multimodal reading. Every post was read with its image or video frame, because the joke often lives in the picture. Each post was annotated with one idea, written as a neutral sentence that states the implied meaning, the camp it serves, the community that posted it and its register.
3. Beliefs and stance. The 18 beliefs were induced from the posts, not fixed in advance, following concept induction with language models (LLooM). Each is written as something a person can believe ("The Japanese are suicidal"), and each label passes four checks: someone can say "I believe that", a joke and a serious post can both carry it, one post can support it while another attacks it, and the label names no format. Each post is tagged as supporting, attacking or neutral toward its belief. The register (humor, argument, testimony, data) is kept apart as the packaging. Posts with no belief (questions, racial insults) stay off the rope.
4. Map. Ideas are embedded with a sentence model (all-mpnet-base-v2). Belief centroids are placed by metric MDS on cosine distance, like the intertopic map of pyLDAvis. Battle lines place each belief by the share of its views serving each camp.
5. Can it be automated?
6. Limits. Seed posts were partly chosen by hand. Views measure reach, not agreement. The camp of a post is a reading, never an assumption about the origin of a person. Post text is never displayed: ideas are paraphrased, and insults are counted but not linked.

References: Lam et al., Concept Induction with LLooM, CHI 2024 · Pham et al., TopicGPT, NAACL 2024 · Grootendorst, BERTopic, 2022 · Hwang and Shwartz, MemeCap, EMNLP 2023 · Sievert and Shirley, LDAvis, 2014.