Everything used many times by many people undergoes wear and tear, sometimes called erosion or attrition in grammaticalization research. In the long history of Indo-European languages, no modern language preserves the significans of any lexical item in the same form that it had in its ancient ancestor, let alone in the form posited for Proto-Indo-European. Erosion affects lexical and grammatical signs alike. Thus in French – to illustrate from a language notorious for its extreme degree of erosion – the Latin noun aqua /āk^wa/ became eau /o/ and Augustus ‘august’ became aout /u(t)/. On the side of grammaticalized items, Latin habet ‘has’ became ha /a/. Thus, erosion is not something exclusive to grammaticalization.

There is some cross-linguistic evidence to show that grammaticalization provokes stronger erosion than happens generally to linguistic signs in diachrony. The following table shows some examples from Yucatec Maya.

The left section of the table shows four grammaticalized items, first in their source form, then in their contemporary grammaticalized form. All of these examples display perceptible erosion. The right section of the table presents four lexical items which are phonologically maximally similar to the items recruited for grammaticalization and share their history, but show no symptoms of erosion. And these are not the only lexical items that are so stable; they are representative of entire sets which have conserved their phonological integrity over time.

From Chinese, some cases of erosion may be cited (Sun & Bisang 2020). As may be seen, the erosion typically affects the root vowel.

In a tone language, a frequent aspect of phonological erosion is loss of tone on the grammaticalized formative; its tone thus changes to neutral tone. In Mandarin Chinese, apart from the examples just quoted, the noun mén 'class' has first tone, but its grammaticalized variant -men, a plural suffix, is toneless. The same goes for the nominalizer de, which is toneless, but apparently stems from a tonic demonstrative pronoun, and for the continuative particle li, grammaticalized from lǐ ‘mile’.

Erosion must be distinguished from merger. An item which is grammaticalized may lose in phonological substance; but this does not imply that its significans melts together with the significans of its host. Grammar affords analytic access to linguistic units. This requires that the lexical and the grammatical information carried by word forms be recognizable. This is easily achieved if they have separate significantia, such as, e.g., a stem and an agglutinative suffix. Even a root and its apophony do the service, because they can still be distinguished phonologically and thus be recognized as distinct signs. It is only in the last phase of grammaticalization of an inflectional operator that it may be merged with its base into a totally irregular form, as is the case with Engl. go - gone. In the end, the paradigmatic relation between two inflected forms of a given stem may be suppletive, like go - went. Only in this case are we faced with merger in the last phase of grammaticalization. Its product has the status of a fossil. A systematic conception of such phenomena may require that we consider this last process as the lexicalization of a grammatical form.

Erosion is the phonological counterpart to desemanticization. A very general statistic principle determining the composition of linguistic signs postulates a correlation between the complexity of the significans and the complexity of the significatum of a linguistic sign (Lehmann 1978). To the extent that this principle holds, erosion is expected to accompany desemanticization. There are many cases in which this is true, including the above examples and the earlier examples of the grammaticalization of Romance and Germanic auxiliaries. However, there are also counterexamples. Thus, most of the numeral classifiers of Persian and Mandarin have conserved the phonological form of their nominal origin. It may be appropriate to eliminate phonological weight from the set of definitory parameters of grammaticalization and instead consider it a statistical effect of desemanticization.