We present our analysis of a significant data artifact in the official\n2019/2021 ASVspoof Challenge Dataset. We identify an uneven distribution of\nsilence duration in the training and test splits, which tends to correlate with\nthe target prediction label. Bonafide instances tend to have significantly\nlonger leading and trailing silences than spoofed instances. In this paper, we\nexplore this phenomenon and its impact in depth. We compare several types of\nmodels trained on a) only the duration of the leading silence and b) only on\nthe duration of leading and trailing silence. Results show that models trained\non only the duration of the leading silence perform particularly well, and\nachieve up to 85% percent accuracy and an equal error rate (EER) of 15.1%. At\nthe same time, we observe that trimming silence during pre-processing and then\ntraining established antispoofing models using signal-based features leads to\ncomparatively worse performance. In that case, EER increases from 3.6% (with\nsilence) to 15.5% (trimmed silence). Our findings suggest that previous work\nmay, in part, have inadvertently learned thespoof/bonafide distinction by\nrelying on the duration of silence as it appears in the official challenge\ndataset. We discuss the potential consequences that this has for interpreting\nsystem scores in the challenge and discuss how the ASV community may further\nconsider this issue.\n