Intention attribution guides the cognitively most demanding forms of social learning, such as imitation, thereby scaffolding cumulative cultural evolution. However, it is not thought to be necessary for more basic forms of social learning. Here we present evidence that in marmoset monkeys (Callithrix jacchus) even most basic forms of social learning such as enhancement depend on intention attribution. Marmosets perceived the behavior of a conspecific and a conspecific-like robot, but not that of a moving black box, as goal directed. Their subsequent choice behavior was shaped by social facilitation and stimulus enhancement, that is, by very simple forms of social learning, but only when exposed to the conspecific and robot, which they previously had perceived as intentional agents. We discuss the implications of this finding for contemporary debates about social learning, including emulation learning and ghost control studies, the necessity of goal-directed copying for cumulative cultural evolution, and the limits of current classification systems of social learning for the evolution of social and asocial learning.