Guardrails (or protection barriers) are the set of restrictions, filters, and rules applied to an artificial intelligence system to narrow its behavior and prevent it from generating harmful, illegal, or inappropriate responses. They function as a control layer that delimits what the model can and cannot do, both in what it receives as input and in what it produces.
Their importance lies in the fact that generative models do not distinguish by themselves between legitimate and dangerous use. Guardrails help mitigate risks such as:
- The generation of violent, discriminatory, or illegal content.
- The leakage of sensitive or personal data.
- Manipulation attempts through prompt injection or jailbreaking.
In practice, they are implemented at various levels: adjustments in model training, filters on user requests, and validations on outputs before displaying them. It is worth remembering that no guardrail is infallible: they are a risk reduction measure, not an absolute guarantee, and must be continuously reviewed and updated.