Checks placed around a model rather than inside it. A separate system inspects what goes in and what comes out, blocking or flagging anything that breaks the rules. Because these sit outside the model, they can be updated immediately when a problem is found, without any retraining.
Technical view
Typically implemented as classifiers, rule sets or smaller models applied to inputs and outputs, covering categories such as harmful content, personal data leakage, off-topic use and prompt injection patterns. The trade-off is between false positives, which block legitimate use and frustrate users, and false negatives, which let harmful content through. Output-side guardrails add latency, particularly where they require the full response before a judgement can be made.
Related
Part of AI Terminology Index, section: Testing, Safeguarding and Running Models in Practice.