Refusal Training
Refusal training is the post-training step that teaches a model to decline requests that violate its policies (illegal, harmful, privacy-invasive, etc.) — and, critically, to comply with the much larger set of legitimate requests that superficially resemble refused ones.
Full definition
Done well, refusal training reduces both true-positive harms and false-positive refusals. Done poorly, it produces models that refuse benign chemistry homework, medical questions, or fiction with violence. Vendors publish refusal benchmarks (Anthropic's helpful-harmless-honest, OpenAI's spec, XSTest for over-refusal). Modern frontier models have reduced over-refusal rates from 20-30% in 2023 to <5% by 2025 while maintaining or improving harm-refusal rates.
Why it matters
Over-refusal is the most common reason enterprise pilots fail. A model that refuses 8% of legitimate queries makes a customer-service automation impossible. Buyers must run domain-specific refusal evaluations, not rely on vendor benchmarks alone.
Example
A health-insurance chatbot built on a 2023-era model refuses 14% of plan-coverage questions citing "medical advice"; switching to a 2025 model with calibrated refusal training drops that to under 1% with no rise in unsafe outputs.
Related terms
- Reinforcement Learning from Human Feedback (RLHF)Reinforcement Learning from Human Feedback (RLHF) is a training technique that aligns a language model's outputs with human preferences by training a reward model on human comparisons of candidate responses, then optimizing the LLM against that reward.
- Constitutional AIConstitutional AI (CAI) is a training technique, introduced by Anthropic in 2022, in which a model critiques and revises its own outputs against a written set of principles, then learns from those revisions instead of from human-labeled preferences alone.
- Alignment TaxThe alignment tax is the performance cost — measured on capability benchmarks — of making a model safer, more honest, or more aligned with human values, relative to the same model trained only for capability.
- AI GuardrailsAI guardrails are the layered controls — input filters, output classifiers, policy engines, schema validation, and structured generation — that constrain an AI system to safe, on-policy, and on-task behavior.
Source & further reading
Primary source: Röttger et al. — "XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in LLMs" (NAACL) (2024).
Citation policy: this entry is part of the AIDOLS AI Implementation Glossary and may be quoted for research, journalism, and education with attribution to aidolsgroup.com/de/glossary/refusal-training/.