What is the role of an LLM guardrail?
Why D is the answer
To prevent negative outcomes, such as unsafe output
A guardrail checks inputs and outputs so that unsafe or unwanted content is stopped before it causes harm.
In detail
- Guardrails prevent negative outcomes such as unsafe output.
- They are implemented as part of development, not added later.
- Guardrails sit outside the model, so they apply to every reply.
Use case
An illustrative scenario, not a real incident.
A public chatbot could be asked for harmful advice. A guardrail checks input and output, and it stops the reply before anyone sees it.
See it in different ways
flowchart LR A["User input"] --> B["Input guardrail"] B --> C["LLM"] C --> D["Mask or filter output"] D --> E["User"]
mindmap
root((Governance))
Input
Guardrails
Output
Masking
Data
Licensing
PermissionssequenceDiagram participant P0 as User participant P1 as Guardrail participant P2 as LLM P0->>P1: Request P1->>P2: Allowed request P2-->>P1: Reply P1-->>P0: Checked reply
Key point: Guardrails prevent negative outcomes.
Why the other options are wrong
- A. To make the model train faster
- Guardrails can add a little work rather than speed things up.
- B. To lower the cost of each model call
- Cost is not their purpose.
- C. To pick the best embedding model
- They do not choose models.