toxicity detection

Toxicity detection classifies text for hate speech, offensive language, harassment, and harmful content. **Categories**: Hate speech (targeting identity groups), harassment/bullying, threats/violence, sexually explicit, profanity, self-harm content. **Approaches**: **Classifiers**: Trained models outputting toxicity scores per category. **LLM evaluation**: Prompt model to assess content appropriateness. **Rule-based**: Keyword matching for explicit terms. **Models**: Perspective API (Google), OpenAI moderation endpoint, HuggingFace toxic-BERT, Detoxify. **Challenges**: Context dependence (reclaimed language, quotation), evolving language, coded hate speech, cross-cultural variations, false positives on legitimate discussion. **Calibration**: Set thresholds based on use case - strict for child-facing, looser for research. **Multi-lingual**: Toxicity patterns differ across languages, need language-specific training. **Implementation**: Score threshold for blocking, gradual response (warning → block), human review for borderline cases. **Integration points**: Input filtering, output filtering, content moderation queues. Foundation for content safety systems.