All terms
Safety & Alignment
Harmlessness
An alignment goal of avoiding outputs likely to cause harm.
Definition
Harmlessness is an alignment objective requiring that a model avoid outputs likely to cause harm, such as instructions for weapons, support for self-harm, or offensive and deceptive content. It is one of the three goals in the helpful, harmless, honest framework. Balancing harmlessness against helpfulness is a core tension, since overly cautious training can lead to unnecessary refusals.