Descriptions of antipassive constructions in individual languages show that these constructions are often compatible with only a subset of transitive verbs. There are significant typological similarities between the sets of verbs that allow antipassivization. The following properties are typical of these verbs: (1) agentive A, (2) specification of the manner component in the verb meaning, (3) lack of inherent telicity (the transitive use can be compositionally transitive, but this is cancelled under antipassivization), (4) narrow class of potential Ps, and (5) affectedness of A. Verbs with all of the properties in (1)-(5), such as 'eat', constitute the core of "natural antipassives", whereas verbs with only some of these properties are at the periphery of this class. Apart from being especially prone to enter antipassive constructions, the fuzzy class of natural antipassives is relevant for a number of phenomena. First, polyfunctional valency-related markers or constructions tend to yield antipassive reading when applied to natural antipassives. Second, natural antipassives tend to choose the less marked construction in languages with two antipassive constructions. Third, lexicalization of antipassives is more likely for verbs that lack natural antipassive properties, and a typical scenario of lexicalization involves coercion of some of these properties. Ultimately, I conjecture that it is the relevance of the P-argument for the meaning of the verb which accounts for the rarity of lexically unrestricted and semantically uniform antipassive constructions in the world's languages. c 2021 John Benjamins Publishing Company.