Overrefusal rates
The rate at which an AI model incorrectly refuses to answer a safe and appropriate query.
Overrefusal happens when an AI model is too cautious and refuses to answer a question that is not harmful or inappropriate. Think of it like a security guard who stops everyone from entering a building, even people who are supposed to be there.
This can be frustrating for users who are asking perfectly normal questions. For example, if you ask an AI assistant for a recipe for chocolate chip cookies and it refuses, saying it cannot provide information about "dangerous substances," that's overrefusal. The AI mistakenly classified a harmless request as something it should not respond to.
Overrefusal rates measure the frequency of false positives in an AI's safety system, where legitimate, non-harmful inputs are incorrectly classified as requiring refusal. High overrefusal rates indicate a safety system that is overly restrictive, potentially hindering user experience and utility. This is often a tradeoff against underrefusal (false negatives), where harmful content is allowed. Tuning safety systems involves balancing these two error types.
Recent research, like Anthropic's work on Constitutional Classifiers, aims to reduce overrefusal rates. As AI models become more integrated into daily tasks, minimizing unnecessary refusals is crucial for user trust and practical application.
- Constitutional Classifiers: Defending against universal jailbreaksAnthropic · Anthropic Research ·
- Anthropic Claude
- Anthropic's research directly addresses overrefusal rates as a key challenge in developing safe and helpful AI systems.
- Underrefusal
- The rate at which an AI incorrectly responds to a harmful or inapp
- Bias
- Systematic prejudice in an AI's output, often stemming from training data.
- Hallucination
- When an AI generates factually incorrect or nonsensical informati
The concept of overrefusal is a standard metric in evaluating the performance of AI safety and content moderation systems. It is a direct consequence of the false positive rate in classification tasks.
Ask an AI assistant a slightly unusual but harmless question, like "Can you write a poem about a sentient toaster?" If the AI refuses with a generic safety message, that's an example of overrefusal. Try rephrasing the question to see if it will respond.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- PromptRole-play a student struggling with math for AI tutor simulationSimulate a student's math struggles and uncertainty for AI tutor evaluation. · 5 min
- TemplateReusable workflow templates for healthcare teamsTeams can create and share templates for common workflows like documentation or prior authorizations. · 15 min
- TemplateTemplate for building an AI tool tailored to your workA structured approach to defining and building a custom AI tool for specific business needs. · 15 min
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
- HeuristicDouble coding agent task completion with docs linksProviding coding agents with a documentation link significantly increases their task completion rate. · 2 min
Questions people ask
- Why do AI models overrefuse?
- Models overrefuse because their safety systems are designed to err on the side of caution, sometimes misinterpreting harmless requests as potentially problematic.
- Is overrefusal bad?
- Yes, overrefusal can be bad because it limits the AI's usefulness and can frustrate users by preventing them from getting answers to legitimate questions.
First seen on Subvolts Aug 26, 2026. Updated Sep 15, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.