Low‐frequency earthquakes (LFEs) are detected primarily from continuous seismograms using a matched‐filter technique with an impulsive template waveform in a relatively narrow frequency band. However, this method can also detect events from some kinds of random time sequences without clearly isolated events. Here this fact is demonstrated via simple numerical simulations using the synthetic moment accelerations from a model of broadband slow earthquake, the Brownian slow earthquake (BSE) model, and a totally random noise sequence. The matched‐filter technique identifies time sections including relatively isolated pulse‐like fluctuations, as signals in both time sequences, depending on the threshold. These waveforms, stacked relative to the signal timing, show a clear impulse similar to the assumed template in both time sequences, which highlights that we might potentially misinterpret an original time sequence as containing many isolated pulse‐like events. An important difference exists between the BSE model and random noise at frequencies much lower than the analyzed frequency band, with the stacked BSE sequence containing coherent signals at very low frequencies, which are not visible in the noise. Real observations in the Cascadia subduction zone also contain similar coherent signals at low frequencies, suggesting that these LFE signals are coincident with some slow slip. Therefore, so‐called LFEs might potentially be a misinterpretation due to signal processing, or at least they are the tip of the iceberg, with these signals forming a component of a very broadband slow‐earthquake‐like slip process that possibly occurs over sub‐second to multi‐year timescales.