Karo, G., Divon, T., & Hallinan, B. (2026). The TikTok caliphate: How jihadist supporters exploit algorithmic recommendations and evade content moderation. Social Media + Society, 12. https://doi.org/10.1177/20563051251412167
Summary
This paper offers one of the first in-depth empirical studies of jihadist supporter activity on TikTok, examining how backers of ISIS and al-Qaeda embed extremist propaganda within the platform’s participatory culture to evade a stated zero-tolerance moderation policy. Through a modified grounded theory approach and an algorithmically-assisted snowball sample drawn from a purpose-built research account, the authors inductively develop a typology of five communicative evasion strategies. They argue these tactics constitute a form of “everyday extremism” in which spectacle and ritual collapse into mundane platform vernaculars — memes, trending sounds, and emoji — thereby exposing structural vulnerabilities in TikTok’s recommendation architecture and the limits of content-based moderation and voluntary industry self-regulation.
Key Contributions
- One of the first empirical analyses of jihadist (rather than far-right) activity on TikTok, addressing a notable gap in extremism research.
- An adaptable, ideology-independent typology of five moderation-evasion strategies suited to cross-platform and cross-group comparison.
- The concept of “everyday extremism” describing how spectacle and ritual dissolve into ordinary platform vernaculars.
- A shift of analytical focus from content removal toward infrastructural critique of recommendation loops and uneven moderation.
- Concrete policy recommendations, including TikTok’s inclusion in GIFCT, hiring contextually expert moderators, and user-driven annotation mechanisms like Community Notes.
Methods
The authors use a modified grounded theory approach informed by prior work on jihadist symbolism and evasion. They created a dedicated anonymous TikTok account (accessed via a US-based VPN) and deliberately “radicalized” its feed through algorithmically-assisted snowball sampling — searching English and Arabic hashtags, emojis, and references to canonical jihadist figures and events, then liking content to extend recommendations. Over 300 pro-jihadist videos were bookmarked between January and July 2024, analyzed through iterative constant comparison and multimodal analysis (treating strategies as configurations of semiotic resources drawing on Kress and Van Leeuwen). A follow-up check in October 2025 found only 25 videos remained, serving as evidence of platform ephemerality.
Findings
- Audio camouflage: pitch/tempo manipulation of nasheeds (e.g., ISIS’s “Dawlati Baqiah,” al-Qaeda’s “Tora-Bora”) and disguised audio metadata titles.
- Meme infiltration: extremist messaging embedded in Roblox, Fortnite, and Inside Out 2 memes and trending sounds (e.g., Mareux’s “Lovers from the Past” glorifying the Sri Lanka attacker).
- Blurred intent: deliberate visual obfuscation of banned figures (Bin Laden, al-Baghdadi, al-Adnani) and ISIS flags to bypass image detection while remaining legible to in-groups.
- Emoji codes: heavy use of the black flag and raised index finger (tawhid) to signal allegiance and bypass text-based moderation.
- Bait-and-switch: videos opening with innocuous content (Pride Month, July 4th fireworks, Ronaldo highlights) before pivoting to extremist messaging.
- Basic moderation flaws: simple textual alterations (e.g., “1S1 S”) defeat keyword filters that block direct searches for “ISIS.”
- Strategies frequently overlap within single videos, making extremist content appear mundane rather than spectacular.
Connections
This paper sits within a growing body of work on adversarial creativity and moderation evasion on short-video platforms; its claim that evasion tactics are ideology-independent resonates with studies of far-right and mainstream-far-right dynamics such as Askanius2026-de and Grusauskaite2026-po. Its infrastructural critique of recommendation loops and platform architecture connects to research on algorithmic amplification and platform governance including Efstratiou2025-gs and Marwick2025-ov. Its methodological reliance on a dedicated research account probing an opaque recommender speaks to broader debates on platform data access and observation raised by Rieder2026-pp.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify · Apple Podcasts