Play occurs universally in human beings, and it has been a subject of considerable academic scrutiny for over 100 years. In this article, we propose a cognitive theory of play building on recent advances in cognitive and computational neuroscience that portray the human brain as an advanced prediction machine. Central to the theory is the idea that when an agent is free from the demands of certain competing cognitive systems, it may deliberately seek out and create surprising situations that gravitate towards sweet-spots of relative complexity. We argue that this framework can explain why humans play and why playing is so fun and rewarding.