It depends. In music for example it's often question whether the artist has been exposed to the original work. In that spirit, small language models are less likely to infringe copyright.