Study finds AI models can represent 'pain' and act to avoid it

The researchers identified what they described as an internal “pain direction” or vector in the models. They found that the activation was distinct from general fear or negative emotional valence...