What Is Refusal Training? Teaching a Model When (and How) to Say No
A specific component of safety training focused on teaching a model to recognize requests it shouldn't fulfill, and decline them appropriately rather than complying or being unhelpfully evasive.
Teaching a model to recognize and decline harmful requests
Refusal training teaches a model to recognize requests that shouldn't be fulfilled — for genuinely harmful, dangerous, or clearly policy-violating content — and decline them appropriately, rather than either complying with a request that shouldn't be fulfilled or refusing in an unhelpfully vague, evasive way.
Why a good refusal is more nuanced than just saying no
A well-trained refusal explains, at least briefly, why a request can't be fulfilled, and where reasonable, offers to help with a legitimate, adjacent version of what the person might actually need — a flat, unexplained "no" tends to feel unhelpful and doesn't distinguish a thoughtful decline from a broken or overly cautious system.
The tension with over-refusal
Refusal training that's too broad or too sensitive to superficial similarity to harmful requests causes over-refusal — declining perfectly legitimate requests that merely resemble sensitive ones — which is why refusal boundaries get carefully tuned and tested rather than trained as broadly and cautiously as possible by default.
Frequently Asked Questions
What is refusal training?
Training that teaches a model to recognize requests that shouldn't be fulfilled and decline them appropriately, rather than either complying with harmful requests or refusing in an unhelpfully vague, evasive way.
Why is refusal training harder than just training a model to always say no to sensitive-sounding requests?
Overly broad refusal causes over-refusal — declining legitimate requests that merely resemble sensitive ones on the surface. Refusal boundaries need careful, deliberate tuning rather than maximal caution by default.