Safety
Guardrails
Safety mechanisms implemented in AI systems to prevent harmful, biased, or policy-violating outputs, including content filters, system prompts, and output classifiers.
Explained at five levels
Level 1
Safety rules that keep the AI from saying mean, dangerous, or wrong things — like bumper lanes at a bowling alley.
Level 2
Safety measures built into AI to prevent it from generating harmful, inappropriate, or dangerous content. They're like rules the AI has to follow.
Level 3
Safety mechanisms implemented in AI systems to prevent harmful, biased, or policy-violating outputs, including content filters, system prompts, and output classifiers.
Level 4
Multi-layered safety controls including input/output classifiers, constitutional training objectives, system-level instructions, and monitoring pipelines designed to keep model behavior within acceptable bounds.
Level 5
A defense-in-depth safety architecture combining pre-deployment alignment (RLHF/Constitutional AI), runtime input/output classifiers, prompt-level constraints, and post-hoc monitoring — evaluated via red-teaming and adversarial robustness benchmarks.
Definitions are educational summaries. Terminology can vary by source and context.
Sources
- NIST Trustworthy and Responsible AI Resource Center: Glossary — National Institute of Standards and Technology. Accessed 2026-07-20.