# Over-Refusal

> Over-refusal is a model declining safe requests because they superficially resemble harmful ones, such as refusing to explain how to kill a Python process.

- Identifier: PTL-0176
- Category: Failure Modes & Evaluation
- Canonical URL: https://protologue.com/t/over-refusal/
- Also known as: exaggerated safety, false refusal, overrefusal
- Introduced: 2023

## Description

It is the cost side of safety training, since overly cautious models frustrate users and withhold legitimate help. Test suites such as XSTest pair safe prompts with unsafe look-alikes to measure the balance, and methods such as deliberative alignment aim to reduce refusals while improving robustness.

## Related terms

- [Guardrails](https://protologue.com/t/guardrails/)
- [Reinforcement Learning from Human Feedback](https://protologue.com/t/rlhf/)
- [Deliberative Alignment](https://protologue.com/t/deliberative-alignment/)
- [Constitutional Classifiers](https://protologue.com/t/constitutional-classifiers/)

## Sources

- Röttger et al. (2023). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. https://arxiv.org/abs/2308.01263

## Cite this entry

Protologue. (2026). Over-Refusal. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0176). https://protologue.com/t/over-refusal/

License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)