Guardrails are the controls placed around an AI model to keep what goes in and what comes out within set limits. They can block certain topics, strip out personal data, check that an answer is in the right format, or require a person to approve an action before it happens. For example, a chatbot on a charity’s website might refuse to give medical advice, remove phone numbers from its answers, and wait for a staff member to approve before it sends any email for a visitor.
Why it matters
Guardrails lower risk, but they are not a guarantee. People can find ways around them, for example with a jailbreak. That is why guardrails work best in layers, with more than one check between the model and anything that matters. A tool might combine a content filter, a safety classifier and careful instructions in its system prompt, plus a person who signs off on important steps. When a vendor tells you their AI tool is safe, ask what guardrails it has, what happens when one of them fails, and whether a person approves actions such as sending messages or changing records.
Published Updated