Abstract In the standard levels-of-processing (LOP) paradigm, subjects are presented with a question directing them to attend to surface features (case), phonetic features (rhyme), or semantic features (category) of an upcoming word. Answering the question greatly affects recall and recognition for the word, and this outcome is attributed to differing LOP of the word. Craik (1977) reasoned that if the words appeared before the questions (i.e., the word-first paradigm) under intentional learning instructions, the LOP effect would disappear. However, when he conducted experiments using the word-first paradigm, the effect remained. We conducted 2 experiments to try to explain this puzzle. In the first experiment, we replicated Craik’s findings in one condition, and in other conditions we asked subjects to make judgments of learning (JOLs) in an attempt to eliminate the LOP effect because JOLs seem to promote deep processing. However, the LOP effect still occurred, albeit in attenuated form. In Experiment 2, we tested 2 new conditions to overcome and explain the LOP effect in the word-first paradigm. In one case, subjects were instructed to use idiosyncratic study strategies to memorize the word; the second condition required subjects to perform pleasantness judgments on the words. The pleasantness rating condition eliminated the LOP effect, but the first did not. We conclude that in the word-first paradigm, subjects typically keep the word in mind without semantic processing using rote rehearsal, despite intentional learning instructions.