Reading Between the Glyphs: Unmasking Gen Z's Coded Cyberbullying on Social Media with AI

Adolescents increasingly speak in code online—where a peach isn’t fruit and “101” isn’t a class. We ask whether these symbol–number–emoji sequences cause measurable harm, how adults misread them, and whether AI can detect them without over-policing. We execute three phases—Map, Listen, Detect. In Phase I (Map), we analyzed 5,000 posts across platforms: covert-hostility prevalence was 2.52% (95% CI 2.12 – 2.99), higher on short-form feeds and concentrated in evenings (73%). Hostility resided in combinations—hybrids (47.6%), emojionly clusters (34.1%), and alphanumeric obfuscations (18.3%)—with mixed-script tricks overrepresented. In Phase II (Listen), interviews with 45 Gen Z participants showed exposure (91%), frequent experience (62% often/always), and salient impacts. Adults positively engaged with coded mockery in 58% of cases; 73% believed this amplified reach. In Phase III (Detect), our sequence-aware model (CRISP) achieved AUPRC 0.79 and Recall@1% FPR 0.58 (ECE 0.032), outperforming a text-only baseline (AUPRC 0.52) and a commercial API, especially on covert items (0.74 vs 0.47/0.31) while cutting benign-banter false positives by 41%. Decision-curve analysis favored CRISP across costs. We contribute a semiotic framework (symbols→sequences→scenes), a corpus, and pipeline that flags coded hostility—including do whistles for extremism or drugs—without ethically silencing in-group play. A repository is uploaded to GitHub to show a sample of posts (n = 100), an Excel codebook and an AIgenerated video (prototype).