Using the oral and manual Stroop tasks we tested the claim that retrieval of meaning from a written word is automatic, in the sense that it cannot be controlled. The semantic interference effect (greater interference caused by color-related words than color-neutral words) was used as the index of semantic activation. To manipulate the level of attentional control over the task of reading, the proportion of nonreadable, neutral trials (a row of #s) was varied (75% vs. 25%). In all four experiments a high-neutral proportion magnified the interference caused by word distractors. With the color-associated words presented in incongruent color (e.g., LEMON in blue), the semantic Stroop effect was weak and did not interact with neutral proportion (Experiment 1 and 2). Experiment 3 and 4 used color names (e.g., GREEN) not in the response set, and here the semantic interference effect was more robust, and the effect was magnified in the high-neutral proportion condition. We take these results to argue that semantic retrieval is controlled by endogenous attention in the Stroop task.