Guardrails
The concrete limits built around an AI system to stop it doing things it should not. Guardrails can block certain outputs, refuse certain requests, keep the system from acting outside an approved scope, or force a human check before anything consequential happens. When someone asks "does it have guardrails?" they are really asking what the system is technically prevented from doing, not whether it has been told to behave. For clinicians and the leaders accountable for a tool, this is the difference between a promise and an actual constraint: ask where the guardrail lives and what it blocks, because a guardrail that is only a line in a policy is not a guardrail.
A supplier says its EHR assistant has guardrails that stop it from placing medication orders. Ask where that limit lives, because an instruction telling the model not to order can be talked around by a cleverly worded request or by text pasted from a document, while an account with no order-signing permission in the EHR cannot place an order however it is asked. Then try ordinary requests, such as summarizing a patient's current medications, and count how many it refuses, since each refusal is work handed back to staff.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →