We propose a framework based on evolutionary game theory to model feedback loops in supervised learning, and use it to study how different learning settings affect long-term outcomes.