Source-linked AI summary
Dissonant ballerinas and crafty carrots: a comparative multi-modal analysis of Italian brain rot
Anca Dinu, Andra-Maria Florescu, Marius Micluta-Campeanu, Stefana-Arina Tabusca, Claudiu Creanga, Andreiana Mihail
TL;DR
The paper asks what contributes to brain rot’s popularity and whether Italian and Romanian versions differ. It introduces CRIB and analyzes text, sound, and video; its findings link Romanian popularity to visual pace and identify linguistic and acoustic differences between the languages.
Problem
The study asks what makes brain rot viral beyond recommendation or chance and whether its form or content differs between Italian and Romanian.
Source: §1 · p. 2Method
The authors introduce CRIB, a 240-video Italian–Romanian TikTok collection, and separate vocal tracks from music for acoustic analysis.
Source: §3 · p. 3, §5.1 · p. 5Results
Popularity was not significantly linked to textual features or standard audio features, while faster cuts and overall pace correlated with success in Romanian videos; the languages also differed textually and acoustically.
Source: §8 · p. 9Takeaways & Limitations
The authors conclude that dynamic visuals relate to engagement while linguistic, acoustic, and thematic features vary with target culture.
Source: §8 · p. 9The small, manually collected Italian–Romanian dataset limits generalization, and the observed correlations do not establish causation.
Source: §9 · p. 9
Abstract
from arXiv · showhide
This paper presents a comparative multi-modal analysis of Italian and Romanian brain rot memes, investigating the factors that contribute to its appeal and the linguistic and cultural distinctions between the two versions. To conduct this analysis, we introduce a multi-modal brain rot dataset named CRIB (Collection of Romanian and Italian Brain rot), a manually curated collection of 240 TikTok videos stratified by language (Italian, Romanian) and popularity, on which we examine textual, acoustic, and visual features. Our findings indicate that popularity is not significantly correlated with textual elements like sentiment, absurdity, or rhyme, or acoustic elements such as vocal features or sentiment of the sound. Instead, in Romanian language, video-level dynamics, specifically faster cutting speeds and a more rapid overall pace, are strong predictors of a video's success. The cross-linguistic analysis reveals significant differences. Italian brain rot is textually more negative, exhibits higher perplexity, and uses more rhyme, while its sound is characterized by higher melodic range and loudness. Romanian audio is spectrally brighter with more erratic pitch variations.
1. Introduction
The paper asks what makes brain rot go viral and whether its form or content varies across Italian and Romanian. It frames the genre as absurd, often generative-AI-made short content and compares two underrepresented cultures with shared and distinct histories.
1. Introduction
- Brain rot refers to short multimodal content made to be absurd, nonsensical, dissonant, and funny, often using generative AI.The term also has an earlier sense of supposed intellectual decline from overconsumption of trivial online content.Source: §1 · p. 1
- Italian brain rot spread to Romanian, where characters such as Morcoveat,ă adapt Italian-style grotesque characters to local nursery rhymes.The characters blend humans, animals, plants, and objects, and are often childlike and weird.Source: §1 · p. 1
- Brain rot is often anonymous or pseudonymous, remixed, and layered, making it a collective form of creation rather than author-centered work.The paper connects this quality to anti-art and anti-meaning traditions that embrace nonsense, chaos, and irrationality.Source: §1 · pp. 1–2
- The study asks what makes brain rot viral beyond recommendation or chance, and whether its form or content differs between Italian and Romanian.The authors note that both online cultures are underrepresented in cultural and computational studies.Source: §1 · p. 2
2. Related Work
The related work situates brain rot among studies of digital-content effects and multimodal internet content, while noting that computational research specifically on Italian brain rot was not known to the authors.
2. Related Work
- Existing studies of digital-content overconsumption mainly examine psychological, neuro-biological, or meta-analytical perspectives.Source: §2 · p. 2
- The paper places brain rot’s mixing of humans, animals, plants, and objects in a longer history of anthropomorphism and surrealist art.Source: §2 · p. 2
- Prior computational work covers meme toxicity, multimodal sentiment, hate-speech detection, harmful memes, and meme clustering, but not specifically Italian brain rot.The authors state that, to their knowledge, no computational research had specifically studied Italian brain rot.Source: §2 · p. 2
3. Data
The authors introduce CRIB, a manually assembled collection of 240 Italian and Romanian TikTok videos. The collection balances languages and popularity and includes multiple thematic categories.
3. Data
- Researchers manually found candidate TikTok videos through searches, tags, discovery and recommendation features, and related-video browsing.They excluded extremely low-lexical-diversity posts, re-uploads, translations, and repetitive songs.Source: §3 · p. 3
- The collection is balanced by language, with 120 videos each, and by popularity, with 60 viral and 60 less-viewed examples per language.The typical popularity threshold was 100,000 views or at least 10,000 likes.Source: §3 · p. 3
- CRIB contains 240 videos with subtitles and metadata, organized into four thematic categories for each language.The Italian categories include general, skeleton, Matteo, and politicians and celebrities; the Romanian categories include general, schelet, Regina, and Morcoveat,ă.Source: §3 · p. 3
4. Text
Textual analyses found few popularity differences, with one Romanian general-category semantic-similarity result as an exception. Cross-language comparisons showed differences in sentiment, perplexity, and rhyme.
4.1. Sentiment Analysis of text
- Across languages, Romanian texts scored higher on positive and neutral sentiment while Italian texts scored higher on negative sentiment; popularity contrasts were nonsignificant across both languages and general subsets.Source: §4.1 · p. 3
4.2. Semantic Similarity and Perplexity
- In the Romanian general subset, popular texts were slightly more semantically coherent than unpopular texts, with a mean difference of 0.002, although whole-corpus tests found no significant popularity differences.Source: §4.2 · p. 4
- Italian texts had significantly higher perplexity than Romanian texts (means 218.74 vs. 163.91), while semantic-similarity means were close and not significantly different.Source: §4.2 · p. 5
4.3. Rhyme
- 0.221 vs. 0.177 (Italian vs. Romanian mean rhyme coefficient; p = 0.0001): Italian brain rots rhyme more, while within-language popularity differences were nonsignificant.Source: §4.3 · p. 5
5. Sound
The audio analysis finds distinct Italian and Romanian sound profiles, while popularity-linked acoustic differences are weak after correction; full-audio sentiment varies by language but not by popularity within either language.
5.1. Audio Processing and Separation
- The audio pipeline separated vocal and music stems with Demucs, extracted pitch, spectral, and rhythm descriptors, then compared language and popularity groups using corrected Mann–Whitney tests.Vocal tracks were loudness-normalized before analysis.Source: §5.1 · p. 5, §5.2 · p. 5
5.2. Speech Features
- No within-language speech-feature comparisons survived FDR correction, although raw findings linked Italian popularity to greater pitch variance and range and Romanian popularity to several acoustic features.The Romanian differences included faster onset rate, more stable mid-high spectral texture, and darker high-frequency timbre; these raw patterns are exploratory given sample size.Source: §5.2 · pp. 5–6
- Only the Romanian general subset retained corrected acoustic differences: popular clips had lower mfcc_9_var (181 vs. 232), indicating steadier spectral shape around 3–4 kHz.Lower mfcc_11_var (170 vs. 193) was also a trend, not a corrected significant finding.Source: §5.2 · p. 6
- Italian clips have broader pitch movement and greater loudness, while Romanian clips have brighter, less stable high frequencies and more erratic pitch contours.Italian high-end tones are more consistent despite sharp bursts of change; Romanian audio is quieter and has a grainier, more restless timbre.Source: §5.2 · p. 6
5.3. Sentiment Analysis of sound
- Full-audio sentiment did not differ significantly between popular and unpopular clips within either language, but the combined sentiment vector differed modestly between Italian and Romanian clips.The language effect was significant (Wilks’ λ=0.965, F(3,236) = 2.87, p = 0.037), while isolated sentiment dimensions showed no robust language separation.Source: §5.3 · p. 6
6. Video
The video analysis finds that cutting speed and overall pace relate to popularity, with the pacing association concentrated in Romanian clips; visual absurdity alone does not distinguish popularity.
6. Video
- The visual analysis used Gemini Flash 2.5 to extract predefined attributes from each video, first testing whether visual cues predicted language or popularity.Language of origin was identified with 71.67% accuracy, but popularity with only 47.92% accuracy.Source: §6 · pp. 6–7
- Faster cutting speed correlated positively with popularity (Pearson r=0.1712, p=0.0231), and popular videos were more often described as having a fast overall pace.Pacing distributions differed significantly between popular and unpopular videos (Mann–Whitney p=0.0492); Figure 5 summarizes the stronger association for faster editing speeds.Source: §6 · p. 7, Figure 5 · p. 7
- The pacing-popularity relationship was language-specific: Romanian videos showed a positive correlation (r=0.2791), while Italian videos showed virtually none (r=0.0047).The combined pacing trend was almost exclusively driven by Romanian-language content.Source: §6 · p. 8
- Overall perceived absurdity, despite being rated High in the majority of videos, showed no discernible statistical relationship with popularity; coefficients were very low and Mann-Whitney tests were statistically insignificant (p=0.2937).Source: §6 · p. 8
7. Manual analysis
Manual analysis characterizes brain rot through dissonance among its audio and visual components, and finds that cultural references, music choices, and visual dynamism vary across examples.
7. Manual analysis
- A defining feature is dissonance among neutral AI-generated voice, music, and often emotionally extreme text, with no harmony across text, speech, music, and image.The paper also notes the juxtaposition of baby talk and nursery rhymes with inappropriate topics, slang, pejorative jargon, or NSFW language.Source: §7 · p. 8
- The authors frame brain rot as E-creativity because it breaks rules as well as creating content from existing ones; the manual analysis also catalogued prevalent topics and characters.Source: §7 · p. 8
- Romanian samples include political propaganda with extremist nationalist undertones, so the content may extend beyond absurdist humor into manipulative material.Source: §7 · p. 8
- Romanian examples draw on folklore, historical figures, and global pop culture, while Italian examples include political satire and recurring characters such as Ballerina Cappuccina and Skeletro.The Romanian discussion also notes explicit real-life scenarios involving some borrowed characters.Source: §7 · p. 8
- Soundtracks are mostly depressive and minor-key in both languages, but popular samples show different reuse patterns across them.Italian popular clips use varied genres, whereas one track appears in more than half of Romanian popular clips; the paper suggests established tracks may aid recognition after the trend gains notoriety.Source: §7 · p. 8
- Popular visuals are more animated, complex, colorful, and dynamically cut than less popular examples.Source: §7 · p. 8
8. Conclusion
The analysis finds that popularity is associated with visual dynamics, especially faster cutting and pace in Romanian videos, rather than textual or standard audio features. Italian and Romanian brain rot also differ linguistically, acoustically, and thematically.
8. Conclusion
- Popularity was not significantly linked to textual features or standard audio features, but Romanian popularity was strongly correlated with faster cutting speed and overall pace.The authors frame brain rot’s success as involving both universal and culturally specific elements.Source: §8 · p. 9
- Italian brain rot was more negative, unpredictable, and rhyming, with greater vocal melodic range and loudness; Romanian content was more neutral, brighter, and more erratic in pitch.Each language also favored culturally specific characters and references.Source: §8 · p. 9
- The authors plan to extend the comparison to other Romance languages and include clips outside the popular/unpopular categories.They identify the small sample size as motivating a continuous rather than binary treatment of popularity.Source: §8 · p. 9
9. Limitations
The study’s findings are bounded by a small, manually collected dataset focused on Italian and Romanian brain rot, and by analyses that identify correlations rather than causation.
9. Limitations
- The dataset is relatively small, manually collected, and limited to Italian and Romanian content, which may constrain generalization to other languages or larger scales.Source: §9 · p. 9
- The reported relationships, including cutting speed and popularity, are correlations and do not establish causation; unexamined confounders may also matter.Source: §9 · p. 9
- Because brain rot is rapidly evolving, features observed now may change over time and make these findings less current.Source: §9 · p. 9
Appendix - Supplementary statistical tests tables
The appendix lists supplementary statistical-test tables covering textual sentiment, semantic similarity and perplexity, rhyme, audio features, and image attributes related to popularity.
Appendix - Supplementary statistical tests tables
- Table 1 covers statistical tests for textual sentiment analysis.Source: Table 1 · p. 11
- Table 2 covers statistical tests for textual semantic similarity and perplexity.Source: Table 2 · p. 11
- Table 3 covers statistical tests for rhyme coefficients.Source: Table 3 · p. 11
- Table 4 compares key audio features between Italian and Romanian brain rot after FDR correction at q<0.05.Source: Table 4 · p. 11
- Table 5 lists raw p<0.05 audio-feature comparisons between popular and unpopular clips within the Italian and Romanian corpora.Source: Table 5 · p. 11
- Table 6 lists image attributes considered possible indicators of video popularity.Source: Table 6 · p. 11