Guardrails
1 min read · Reviewed 17 Oct 2026
In short
The limits built into an AI tool to stop it doing certain things.
In plain English
Rules layered on top of a model — refusing certain requests, filtering outputs, staying on topic. Some come from the model's own training, some are added by whoever built the product.
A real example
A customer-service bot that politely declines to discuss anything but your products has guardrails. So does a model that refuses to give medical dosages.
Why it matters
Guardrails are helpful and imperfect. They can be too strict, blocking legitimate work, and they can be worked around. Never treat them as your only safeguard.
