Who Educates the AI That Will Educate Our Students? Universities, Post-Training, and the Difference Between Declared Theory and Revealed Preference

Progress in large language models has shifted decisively from pre-training to post-training: supervised fine-tuning, preference optimisation, process supervision, and reinforcement learning with verifiable rewards. This paper argues that the shift carries an overlooked theoretical consequence. If pre-training can be described as acquisition through exposure, post-training is formation — curriculum design, guided practice, corrective feedback, and the assessment of acquired competence. The claim that artificial intelligence is constituted by educational processes, rather than being an object to which education is externally applied, thus ceases to be interpretive and becomes descriptive. The paper then addresses an empirical question: are universities already participating in post-training, and can their participation be distinguished from that of frontier laboratories? The paper identifies four modes across existing work: post-training as open methodology (Tülu), as institutional and linguistic sovereignty (Apertus, Minerva), as disciplinary grounding (Meditron), and as the translation of learning theory into model behaviour (ConvoLearn). It proposes a four-criterion framework to distinguish academic from industrial practice: the provenance of the training objective, the epistemic status of domain experts, the disposition of negative results, and the locus of validation. The underlying distinction is between objectives derived from declared theory and objectives derived from revealed preference — between a curriculum and a habitus. Three counterarguments are examined: the absence of measurable capability gains from theory-grounded academic post-training; the rising computational cost of reinforcement learning, which is closing the window of academic accessibility; and the structural character of misalignment, which originates in a pre-training stage universities do not design. The paper concludes that the value of institutional participation is epistemic and governance-related rather than performance-related, that the university's claim to this role is genealogical rather than opportunistic, and that student participation in post-training work constitutes a pedagogical opportunity in its own right. Part 5 of the series: AI is Education.