Scaling feedback?

This maps to what I've seen on the AI-integration side: feedback scales well when the response space is constrained. Branched scenarios and knowledge checks with pre-defined paths? Scripted or AI-assisted feedback works beautifully there, because the rubric is baked into the design. Open response is where it gets shaky, and DShaw's rubric point nails why — hand a model a generic rubric and you get confident-sounding mush. The AI feedback I've found actually useful sat on top of rubrics written like a strict SME wrote them: specific examples of what good looks like, not adjectives.

The other decision I'd force early: where does the human sit in the loop? AI as a first-pass draft for a human reviewer to approve is a very different scaling story than AI as the final voice to the learner. In regulated training especially, that second one needs a paper trail and a person accountable for it. My default is AI-drafts-human-signs-off, and fully automated feedback only for low-stakes, objective checks.