Kolmogorov–Arnold Networks (KANs) have recently been proposed as powerful alternatives to multilayer perceptrons (MLPs), motivated by their universal approximation guarantees and adaptive spline-based activations. In this paper, we present a systematic empirical comparison of KANs and MLPs for molecular property prediction tasks in chemoinformatics. Across four benchmark datasets (Tox21, ESOL, Lipophilicity, BBBP) and under controlled experimental conditions, we consistently observe that MLPs match or outperform KANs. This finding calls into question the practical utility of KANs in descriptor- and fingerprint-based chemoinformatics pipelines, tempering recent optimism. We discuss possible reasons for this performance gap and conclude by outlining the directions of research in light of the findings in this paper. In the appendix on the mathematical analysis of KAN versus MLP, we prove that two KAN layers can emulate one MLP layer and that one KAN layer cannot emulate one MLP layer.