Guardrails

In one sentence

Rules and filters that prevent AI from generating harmful, biased, or inappropriate content. They act as safety boundaries that keep AI systems operating within acceptable limits.

In context

When you use ChatGPT and it refuses to explain how to make something dangerous, that's a guardrail in action. Companies add their own guardrails too—a customer service bot might be programmed to never discuss competitor products, always escalate suicide mentions to a human, and avoid giving medical or legal advice. Guardrails can work at the input level (blocking certain requests), the output level (filtering responses), or both.

See also