Several experiments have shown that the difference in masked threshold between signals in the S0M0 and SπM0 conditions, the masking‐level difference (MLD), increases markedly with reduction in masker bandwidth. However, there are two inconsistencies this result. First, apparently similar experiments conducted by different laboratories have produced quite different results. Second, the MLD‐bandwidth relation appears to be nonmonotonic; with very narrow‐band maskers (e.g., a pure tone), the MLD is often small. The results of the experiments reported here suggest that these problems arise from anomalies in the S0M0 data alone. It appears that in some narrow‐band masking conditions the SπM0 threshold is spuriously low, resulting in deceptively small MLDs. The proposed explanation is that observers detect signal energy which falls outside the masker band, for when this energy is attenuated by filtering the signal, the S0M0 threshold rises dramatically. The SπM0 threshold is unaffected by this manipulation. With specially filtered signals, the nonmonotonicity of the MLD‐bandwidth relation is considerably reduced, and MLDs with continuous pure‐tone maskers are observed which are nearly as large as the largest narrow‐band noise MLDs ever reported (28 dB).