This study explores the ways in which the semantics of personal middle construc tions is encoded across languages. In Dutch, German and English, middles are syntactically unergative and the implicit Agent is syntactically inert. In Greek and French, middles are syntactically indistinguishable from generic passives: they exhibit a derived subject and a syntactically represented Agent. What unites the two types of middle is the interpretation they receive. The cross-linguistic variation invites the following question: what determines the choice of structure employed to convey the middle interpretation Any attempt to address this question requires a characterization of the mid dle interpretation itself. I make the following novel proposal: middles ascribe a dispositional property to the understood object. Disposition ascriptions are subject-oriented generic sentences. The core properties that middles share across languages follow: the genericity of an otherwise eventive predicate the promotion to subject position by syntactic movement or base-generation, and the interpre tation of the otherwise internal argument: the demotion and interpretation of the otherwise external argument. The crosslinguistic variation relates to the following two factors. First, the different means available to languages to encode genericity distinguishes between unergative and imaccusative middles. Unaccusative middles obtain in languages like French and Greek, which encode genericity in the morphosyntax in the form of imperfective aspect . Languages where genericity is not expressed by aspectual morphology, i.e. German. Dutch and English, employ unergative structures. An additional factor at play within Germanic is the nature of the anaphoric system. I attribute the illicit ness of zich in middles to the nature of the Dutch reflexive paradigm, which includes a complex anaphor, zichzelf. In the absence of a com plex anaphor in German, sich can function as an argument but also as a marker of valency reduction its occurence in middles is expected. The approach makes predictions for other structures besides middles and other Germanic languages, such as Afrikaans and Frisian.