Harmful, offensive, or abusive content in model output.
Toxicity detection is part of responsible deployment.