What are AI guardrails?

AI guardrails are the controls that limit what an AI system can say or do. Some are instructions given to the AI, such as “never promise a delivery date”. Others are separate software that checks what goes into the AI and what comes out. That software blocks anything that breaks a rule, such as a reply promising a refund the policy does not allow.

Why it matters for a business

The 2 kinds are not equally strong. An instruction is a request the model usually follows and can sometimes be talked out of. A software check holds whatever the model was persuaded to write. Stronger than either is a permission the AI does not have at all. When a product claims guardrails, what matters is which kind, what exactly is checked, and what happens to an output that fails the check.

Example

Let’s say an assistant’s instructions say never to promise a delivery date. A customer asks persistently enough and it promises one anyway. A software check that holds back any reply containing a date not found in the order record stops that reply, however it was worded.

Go deeper

Human in the loop and guardrails: how much checking AI needs

Related terms

• Human in the loop
• Prompt injection
• System prompt
• Responsible AI
• AI governance

All terms, A to Z