Guardrails
Guardrails: Technical and procedural controls that constrain an AI system’s inputs, outputs and actions to reduce risk. They include validation, policies, filters, permissions and monitoring. They must match the application and be tested against errors and attacks; they do not replace model quality or human oversight.
In italiano: GuardrailsLegge III · Come una rete impara dai datiThey include validation, policies, filters, permissions and monitoring. They must match the application and be tested against errors and attacks; they do not replace model quality or human oversight.
Examples
- Mandatory output schema
- Human approval before a payment
- Sensitive-data filtering