The Unseen Gap in Online Safety

What if the algorithms designed to protect our children are fundamentally blind to the way they actually speak? For the 4.95 billion active social media users—roughly 61% of the global population—the digital town square is increasingly a minefield. While artificial intelligence is marketed as a vigilant shield against online abuse, a comprehensive new meta-synthesis reveals a troubling "resource gap" that leaves millions of the most vulnerable users exposed.

A Rigorous Investigation

This study was conducted using the rigorous PRISMA framework, which filtered 279 records down to a final selection of 27 peer-reviewed studies to map the technical landscape of cyberbullying detection. The findings are a wake-up call for a world that moved its social life online: while we have the raw computing power to identify hate, we are failing to apply it equitably across the globe.

The Severe Human Cost

The stakes are not merely technical; they are deeply human. The research highlights critical vulnerabilities:

The psychosocial impact on adolescents aged 9 to 17 includes social anxiety rates of 40-50%, self-harm, and suicidal ideation.
The COVID-19 pandemic sharply worsened this crisis, with reported cyberbullying cases surging by approximately 70%.

The Technical Landscape & Its Leaders

The "gold standard" for cyberbullying detection is shifting. Here is how the leading models compare:

Performance & Accuracy

Key detection technologies and their reported accuracy:

Traditional Models: Support Vector Machines (SVM) achieved a peak accuracy of 0.976 in controlled settings.
Modern Architectures: Newer models have captured the lead in handling nuance.
- Bidirectional LSTM (BiLSTM)
- Transformer-based models like BERT, which maintain an accuracy of 0.80–0.91.
Specialized Architecture: The Salp Swarm Algorithm-Deep Belief Network (SSA-DBN) reported a staggering 0.999 accuracy in specific environments.

The Systemic Bias Problem

However, these impressive numbers mask a deep structural bias that creates a dangerous digital divide in safety.

The Dataset Imbalance

Two major inequities are embedded in the data that powers these systems:

Language Bias: A massive 74.1% of training datasets are English-centric. Low-resource languages like Swahili, Hindi, and Dutch represent only 3.7% each.
Content Imbalance: In widely-used datasets like FormSpring, bullying accounts for just 6.1% of instances versus 93.9% non-bullying content. This "class imbalance" causes models to struggle to find the proverbial needle in a haystack.

The Path Forward

The researchers conclude that most current AI safety systems are post-hoc—reacting after the damage is done—rather than being preventative.