# Refusal (AI)

> When an AI model says no to a request. Most often it could cause harm.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** AI Refusal, Model Refusal, Over-Refusal

## Think of it like

It is like a shop worker who will not sell knives to a child.

## Definition

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.

## Example

A nurse asks a model for the safe top dose of a common pill. A model that is too careful says no, though the answer is on the box.

## Why it matters

Too few refusals let harm through. Too many make the model useless and push people to worse tools.

## Common mistake

Thinking more refusals means more safety. Saying no to safe questions is a failure too.

## Three principles

- **Trained in.** Part of the model, not a filter.
- **Two ways to fail.** Too many, or too few.
- **Explain the no.** Offer a safe path.

## Related terms

- [Guardrail (AI)](https://learn.benibix.com/term/ai/guardrail)
- [Jailbreak (AI)](https://learn.benibix.com/term/ai/jailbreak)
- [AI Alignment](https://learn.benibix.com/term/ai/ai-alignment)
- [Constitutional AI](https://learn.benibix.com/term/ai/constitutional-ai)
- [Reinforcement Learning from Human Feedback (RLHF)](https://learn.benibix.com/term/ai/reinforcement-learning-from-human-feedback)
- [AI Content Moderation](https://learn.benibix.com/term/ai/ai-content-moderation)

## Told apart from

- [Refusal (AI) vs Guardrail (AI)](https://learn.benibix.com/compare/refusal-vs-guardrail): 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.

![Three requests to a model. A safe request is answered. A harmful request is refused, which is right. A safe request that only sounds risky is also refused, which is wrong and is called over-refusal.](https://learn.benibix.com/diagrams/ai4-refusal.svg)

---

Source: https://learn.benibix.com/term/ai/refusal
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
