Ideophones, like English bang or thud, are interactive expressions used as vivid depictions of sensory imagery of states, events, objects, or qualities (cf. Dingemanse 2011, 2012, 2018; Dingemanse & Akita 2017; Andrason 2020, 2021). They are claimed to represent a universal class of linguistic forms – that is, any given language can be expected to have a set of them. That set may be highly limited, as is the case in many European languages, but it may as well be almost as large as that of lexical categories like nouns and verbs. There are thousands of ideophones in the spoken usage of languages like Korean, Japanese and Basque, which have 4500 or more of them (see Dingemanse 2018; Haiman 2018). Ideophones exhibit an ambivalent structural behavior. On the one hand, they have been described as grammatical forms that are syntactically unattached and prosodically set off from surrounding text material. On the other hand, they have also been described as morphosyntactically integrated adverbials, adjectivals, verbals, or nominals in a number of languages. Building on some earlier work (especially Dwyer & Moshi 2003), the goal of the present paper is to look at ideophones from the perspective of grammaticalization theory with a view to accounting for this ambivalent behavior. It is argued in the paper that we are dealing here with a process that differs from ‘canonical’ grammaticalization in that the end-product of the process is a lexical rather than a grammatical form.