Educational responses to generative AI tend to be polarised: refusers who see it threatening academic rigour, values, and the very project of education itself, versus evangelists promising technological salvation. But what if we’re missing orientations between hand-wringing panic and techno-utopian euphoria?

This paper identifies three distinct orientations to pedagogical AI use. First, the instrumental/transactional: AI as vending machine (homework generator, text producer, academic integrity threat). Second, the meta-relational: drawing on Andreotti’s work, genuine subject-subject engagement emphasizing emergence, dialogue, and relational becoming. Third, semi-bounded pedagogical infrastructure: tools grounded in curated, discipline-specific content that enable both dialogue and structured knowledge work within a disciplinary frame.

I illustrate this third orientation through a proof-of-concept intervention: introducing NotebookLM to Business Economics students (ICT engineers) as they started preparing for final assessment. A small-scale exploratory study revealed distinctive patterns. While the tool offered both chat functionality and structured features, participants valued it for being “grounded in the slides” with “no hallucinations,” and primarily used the structured revision tools (quiz generation, flashcards, targeted Q&A) over open-ended conversation. Neither vending machine nor meta-relational critical friend, but something else.

This approach suggests design principles emphasizing epistemic grounding, disciplinary content, and scaffolded engagement. The patterns that emerged indicate a viable space worth further exploration.

The paper offers both a conceptual framework for thinking beyond the refusal-versus-adoption binary and a concrete example of semi-bounded AI design and use that educators can adapt to their own teaching contexts.