Corpus, Not Curriculum: Computational Corpus Planning in National AI Strategies

It appears in a particular register: as corpus, model and infrastructure; something to be assembled, trained and secured, rather than as a relationship between teachers and learners. We describe this as a distinct phase of language planning, computational corpus planning, in which the training dataset assumes the role once paradigmatically occupied by the dictionary and adequacy is defined positionally against other languages.

A five-part typology of justification is developed and mapped across regions. Protection commitments were concentrated in Europe (8 of 12 cases), with only one Asian strategy employing a protection rationale, while linguistic-cultural sovereignty framing co-occurred with a language commitment in all fourteen cases in which it was coded. Where foreign language does surface, it is not treated as a competence to be developed but as friction to be engineered away: one strategy states that natural language processing “can compensate for the lack of literacy or foreign language skills.” We argue that the invisibility of language education in AI policy is therefore not simply an omission, but a consequence of how these documents constitute language in the first place.

Inter-rater reliability is reported for all fields (κ = .791 for language-policy presence and .727 for foreign-language content).