In our recent article "Polygraph-based deception detection and Machine Learning. Combining the Worst of Both Worlds?", we critically exposed the drawbacks of the tendency to apply machine learning (ML) methods to ad hoc convenience data. To illustrate our arguments, we referred to a recent publication on polygraph-based deception detection by Asonov et al. Our main argument was that training ML models on data with human-assigned labels, rather than actual ground truth, does not meet the requirements for developing and validating evidence evaluation systems currently used in several areas of forensic science, such as fingermark examination and automated human-supervised forensic voice comparison. The requirement for known ground truth data is also emphasized for AI-based methods for use in legal systems more generally. In a rejoinder to our article, Asonov et al. make a number of claims to which we respond in this letter to the Editor.