It is demonstrated that the effective atomic radii suitable for identification of covalent bonding can be revealed from the parameters of machine-learning classifier trained to distinguish between bonded and non-bonded pairs of atoms in 28 710 molecules. The proposed way cardinally differs from the conventional technique of finding the atomic radii by solving a regression problem and approximating the bond lengths by the sum of adjustable parameters. Despite the contrasting approach, the proposed 'machine-learned atomic radii' obtained for H, B, C, N, O, F, Si, P, S, Cl, Ge, As, Se, Br elements are shown to be in close agreement with the conventional 'regression-based' covalent radii. We find this example appealing in demonstrating how some of the concepts peculiar to chemical intuition can be re-discovered by using a suitable machine-learning model. The boundaries between ‘definitely bonded’ and ‘likely bonded’ interatomic distances introduced naturally in the proposed formulation may prove useful for molecular modeling software.