Guardrails
Definition
Technical or procedural controls designed to constrain AI behaviour or prevent undesirable actions or outputs.
Put simply
Controls intended to keep an AI system inside acceptable boundaries.
What boards and leaders should understand
Guardrails is a broad term and can create false reassurance.
A guardrail might filter an output, restrict access, require approval or prevent certain actions. Different controls protect against different failures.
The useful question is therefore not whether “guardrails are in place”, but what those controls can and cannot actually prevent.
Questions to consider
- Which behaviours are our guardrails designed to prevent, and which still depend on detection or human intervention?
