Researchers have built an AI system called CareGuard that tries to catch cyberbullying before it escalates, framing the problem as a mental-health issue as much as a moderation one.
The framework combines zero-shot semantic labeling with fine-tuned transformer models - BERT, DistilBERT, and RoBERTa - to classify posts across several cyberbullying categories. To keep costs down, it adds an emotion-aware filter and cosine-similarity screening, so the system only runs its heavier analysis on content that already looks emotionally charged and semantically relevant. On benchmark datasets, the researchers say this combination held detection accuracy steady while cutting computational overhead.
Most cyberbullying detection tools are built for platform moderation, not healthcare. Pitching this explicitly as mental-health infrastructure - something a hospital system or school counselor's office might deploy - is the more interesting move here, since it reframes online harassment as a clinical early-warning problem rather than a content-policy one.
Whether that framing survives contact with real deployment, where false positives carry very different stakes in a clinical setting than on a social app, is the open question the paper does not answer.