What it means
Guardrails are technical and procedural controls that limit what an AI system can receive, produce or do, such as filtering harmful content, blocking personal data, restricting topics or requiring confirmation for certain actions. They reduce risk but cannot guarantee safe behaviour in every case.
Real-world example
A city's public chatbot has guardrails that detect and block personal data such as tax IDs in user messages, keep answers within municipal services, refuse to give legal advice, and redirect people in crisis to appropriate helplines. The team tests these guardrails with deliberately tricky questions before launch and after every update.
Why it matters for AI adoption
Guardrails are an essential layer of defence, but they can be bypassed, can block legitimate requests and need ongoing testing. They work best alongside careful system design, limited permissions, monitoring and human oversight.
Questions to ask
- What guardrails are in place, and how have they been tested, including with deliberate misuse?
- How do we detect and respond when guardrails fail or block legitimate use?