ITG NowHome › AI glossary › Guardrails
What are AI guardrails?
Guardrails are rules and checks that keep an AI system from giving harmful, off-topic or unsafe answers.
- Category Safety & law
- Also called guardrail
- Updated 20 Oct 2026
Guardrails, explained simply
Guardrails work like the rails on a mountain road. They don't drive, but they keep the car from going off the edge.
They can be built into the model during training. They can also be extra filters that check questions and answers.
You notice them when a chatbot declines a request. Businesses add their own guardrails too, so a support bot stays on topic.
On ITG Now, our free tool guides sometimes mention guardrails that a tool adds to keep its answers on track.
Related terms
AI safetyAI alignmentJailbreakPrompt injectionSystem promptRed teaming
Where you'll see it on ITG Now
Questions people ask about Guardrails
Can guardrails be bypassed?
Sometimes. People try tricks to get around them, which is why companies keep testing and updating them.
Why does a chatbot refuse some requests?
Guardrails tell it not to help with harmful or unsafe things. Sometimes it refuses by mistake, and you can rephrase.
Do all AI tools have guardrails?
Most big ones do. Open models may have fewer, so the person using them is more responsible.