noun · AI safety
Guardrails
also called safety filters
Rails along the road. They do not drive the AI, they just keep its input and output from going where they should not.
Checks applied to a model's inputs and outputs, such as filters, validators and policies, that block, modify or flag unwanted content or actions.
Separate checks that run before and after the model: classifiers, rules, validators or a second model. They can block, rewrite or flag a request or a reply, and they sit in the harness, outside the model itself.
Real-life example
You launch a support bot for your shop. One check refuses questions that are off topic, another removes anything that looks like a card number from replies. A third blocks a refund bigger than a set amount until a person approves it. The model is the same, the checks are what keep it in bounds.
Where the analogy breaks
-
Guardrails are not inside the model's head. They are extra code around it, so the model itself is unchanged and can still produce what they catch.
-
They are not perfect. They miss some bad cases and block some good ones, so expect both and test for both.
-
One layer is not enough. Combine input checks, output checks, limited permissions and review for risky actions.