Collecting linguistic data in a diglossic speech community poses a special challenge for grammarians because of the socially sensitive variation between a High and a Low variety. In Ferguson’s (1959a) original definition only those cases counted as diglossia in which the varieties belonged to the same language. Later definitions expanded the scope to include communities where non-cognate languages serve as High and Low variety, which are much easier to keep apart for both speakers and grammarians. In this paper I explore how Latin and Arabic grammarians dealt with the linguistic variation in their speech community. They regarded all variation as part of one language; yet, in their description of the language they limited themselves to data from the High variety, which was the only one recognized by them as correct speech. Even when they referred to speech errors, they did not target the Low colloquial, but mistakes made in writing by semi-literate users of the High variety.