Toxicity refers to AI outputs that are harmful, offensive, discriminatory, or otherwise inappropriate. It is not always intentional — models trained on large web datasets can reflect harmful patterns present in that data, producing outputs that reinforce stereotypes, use offensive language, or provide harmful information.
Guardrails, content filters, and output evaluation systems are used to catch and prevent toxic outputs from reaching users. In customer-facing or high-stakes enterprise applications, the standard is effectively zero tolerance — because a single harmful output can cause real damage to users and organizations alike.
