Interference in picture naming by task-irrelevant words in Stroop-like paradigms has been interpreted as evidence of automatic semantic processing of unattended words. However, this effect may have been due to attentional rather than automatic processing. To test this hypothesis, subjects were instructed either to ignore task-irrelevant words (selective attention) or to attend to them (divided attention). Three separate dependent measures (interference in picture naming, recognition memory, and confidence ratings) provided converging evidence about subjects’ attentional strategies. It was predicted that semantic interference would occur in the divided attention condition, but not when subjects attended selectively to the picture and ignored the word. Contrary to predictions, semantic interference occurred in both instruction conditions. Implications for future research on automaticity are discussed.