MORAIDICTIONNAIRE IA
Inférence & Prompting

Guardrails

The control mechanisms that keep an AI safe, useful, and within its limits.

Picture a tightrope walker: guardrails are the safety net stretched beneath them. They are the control mechanisms that surround an AI model so it stays useful, safe, and compliant — without ever stepping past the boundaries set by its designers.

Where do they act?

Guardrails operate at several stages of the inference chain, not just "inside" the model:

Stage Guardrail example
Input Malicious-query filtering, prompt-injection detection
System Framing instructions (the "system prompt") defining role and prohibitions
Output Toxicity classifiers, fact-checking, polite refusals
Post-processing Moderation, redaction of personal data

Key idea: a guardrail is an external layer, separate from the model's weights. It can be updated without retraining the AI.

The fundamental trade-off

Too many guardrails choke usefulness (over-refusal, bland answers); too few invite risk (harmful content, leaks). We therefore seek a balance, often framed as:

$$\text{Score} = \alpha \cdot \text{Usefulness} - \beta \cdot \text{Risk}$$

where $\alpha$ and $\beta$ encode the organization's tolerance.

Best practices

A good guardrail goes unnoticed until it's needed — like a seatbelt, it protects without getting in the way.

Explore the full AI dictionary →