Refusal (AI) vs Guardrail (AI)
People think a model's refusal is a guardrail blocking the request. Most of the time it is not. A refusal comes from the model's own training: it decides to say no. A guardrail is a separate check placed around the model, on the way in or the way out. A good app uses both, because each one misses things the other catches.
Refusal (AI)
When an AI model says no to a request. Most often it could cause harm.
A refusal is when an AI model declines to do what was asked. Models are trained to refuse requests that could cause serious harm, such as help with weapons. This comes from training, not from a filter added on top. But refusals go wrong both ways. A model may refuse a safe request that only sounds risky, which is called over-refusal, or help with a harmful one that is worded cleverly. A good refusal says why, and offers a safe path where it can.
Guardrail (AI)
A check around an AI model. It blocks bad requests and bad answers.
A guardrail is a check placed around an AI model to stop bad input and output. An input guardrail looks at a request before the model sees it. It can block attacks or requests that are off topic. An output guardrail looks at the answer before the user sees it. It can block harm, hide private data, or check the format. Guardrails can be simple rules, or a second, smaller model that sorts the text.
In one line
- Refusal (AI)
- When an AI model says no to a request. Most often it could cause harm.
- Guardrail (AI)
- A check around an AI model. It blocks bad requests and bad answers.