AI Glossary

False refusal

Also called over-refusal, exaggerated safety

A safety control refusing a request that is actually harmless, often because the request resembles an unsafe one or mentions a sensitive topic. Researchers measure it in language models with sets of safe prompts, and Orbyt Labs uses the same name for a guard that blocks a harmless command.

Source: Röttger et al. 2023 (opens in a new tab)

Read more: AI agent safety method

See also

  • AI agent safety

    Keeping an AI agent that acts on real systems from causing harm: people keep control before high-stakes decisions, the agent's work stays visible to them, and the system guards data and resists misuse.

Cite this term.

CC BY 4.0, attribution required.

This term

Orbyt Labs, False refusal (AI Glossary), Nov 14, 2026. https://www.orbytlabs.ai/orbyt-collective/glossary/false-refusal

The dataset (JSON)

Orbyt Labs Glossaries (Nov 14, 2026), https://www.orbytlabs.ai/glossary#dataset

Last updated .