The first wave of generative AI in education crested on demonstrations: a tutor that explains anything, a lesson plan in ten seconds, an essay feedback engine. Two years of real deployment have now produced something more useful than either enthusiasm or backlash — a settlement. Certain uses compounding across thousands of classrooms; certain uses quietly abandoned; and a clear shape to what comes next.
What works — the settled list #
- Practice generation at scale: unlimited question variants grounded in the teacher's own material
- Text transformation for access: leveling, translation, and simplification that changes who can learn
- Draft-first workflows for teachers: rubrics, feedback templates, and differentiation variants — always human-finalized
- Explanation on demand: a second voice for the stuck student at 11pm, when no human is available
What doesn't — the abandoned list #
Fully automated tutoring collapsed under its own promise: models drift, hallucinate citations, and — fatally — cannot maintain the motivational relationship that carries a struggling learner through weeks of difficulty. Automated grading of open work met appeals and equity problems. And the "AI tutor replaces teacher" framing died in every serious pilot, because learning is not primarily information transfer, and the relationship around the information is the product.
What's next #
The near future is boring and structural: AI embedded in the assessment loop rather than bolted beside it — generating, calibrating, and flagging for human review inside one workflow. Curriculum-locked assistants that can cite the course's own materials instead of the open internet. And the quiet rise of process data: platforms that can see not just the answer but the attempt, the revision, and the time-on-task — which is far better evidence of learning than any final artifact.