Safety & Alignment
Guardrails
Quick Answer
Systems or rules that prevent LLMs from generating harmful, toxic, or inappropriate content.
Guardrails are safety mechanisms preventing problematic outputs. Guardrails include: content filters, topic restrictions, and refusal patterns. Guardrails might block certain keywords or topics. Guardrails can be too restrictive (over-filtering) or too lenient (under-filtering). Balancing safety and usability is challenging. Guardrails are necessary but not sufficient. Guardrails are typically learned during training. Guardrails require ongoing maintenance.
Last verified: 2026-04-08