Four forced-choice letter-detection experiments with 164 undergraduates examined the effect on detection latency of noise letters (NLs) that were visually similar to target letters. A single target letter was present in each display. NLs were similar to the target letter present in the display (the signal), to a different target letter assigned to the same response as the signal, or to a target letter assigned to the opposite response from the signal. NLs were present in either relevant or irrelevant display positions, and 2 stimulus sets were used. The experiments were designed to test a prediction of models in which information about NLs is transmitted continuously from the recognition to the decision process. These models predict that responses should be faster when NLs are visually similar to a target assigned to the same response as the signal than when NLs are similar to a target assigned to the opposite response. Statistically reliable effects of the type predicted by continuous models were obtained when NLs appeared in relevant display positions, but not when they appeared in irrelevant positions.