We develop a model of algorithmic trading in which a better-informed algorithm competes with human traders who choose whether to acquire costly information and participate in the market. We compare a model-free algorithm, which learns how to trade based on realized profits alone, with a model-based algorithm that learns from a model of its price impact. We show that model-based learning induces more aggressive trading, raises the algorithm's profits, and can reduce price informativeness by crowding out human information acquisition. This implies that mandates for "interpretable" algorithms can lead to worse market outcomes. Restrictions on algorithmic trading intensity or access to information can improve price informativeness when the resulting increase in human participation outweighs the loss of information from algorithmic trading.