Children's bowling lanes have bumpers so the ball can't end up in the gutter. AI has bumpers too.
In a children’s bowling lane, bumpers pop up along the gutters. Throw a wobbly ball and it bounces back towards the pins. You still bowl. You just can’t end up in the gutter.
Guardrails are those bumpers for AI. They’re rules and checks around an AI that stop it saying or doing things it shouldn’t.
Kinds of guardrails
- training the AI to refuse harmful requests
- a system prompt with clear rules
- filters that check questions and answers for anything off-limits
- for AI agents, limits on what they can do without a person's approval
Why they matter
An AI will try to be helpful with almost anything. Guardrails stop it giving dangerous instructions, leaking private data, making promises a company can’t keep or taking risky actions.
What it isn’t
It isn’t a guarantee. Determined people sometimes find ways around guardrails, for example with prompt injection. Several layers work better than one.
Where the comparison breaks down
Bowling bumpers are the same height for everyone. AI guardrails have to judge context: the same question can be fine from a nurse and worrying from someone else.
How the words fit together
Guardrails include the system prompt and other checks. They defend against prompt injection and keep AI agents in bounds.
Guardrails sometimes block perfectly harmless requests. If an AI refuses something reasonable, rephrasing with more context often helps.

When you'll hear it
“It hit a guardrail.” The AI refused, because a safety rule kicked in.
“We need guardrails before launch.” We need checks in place so the AI can't say or do something harmful.
“The guardrails are too strict.” The safety rules are blocking things that should be allowed.
“What must our AI assistant never say or do? How do we check it sticks to that?”

