Every educator now operates inside two loud minorities: those who promise AI will revolutionize learning by autumn, and those who insist it will dissolve education itself. Both are selling certainty the technology does not have. What follows is the settled middle — where AI demonstrably helps classrooms today, where it quietly fails, and the two places it genuinely should not be allowed.
Where it works today, boringly #
- Question generation: turning a syllabus into unlimited practice variants — the single highest-yield classroom use
- Summarization: condensing dense material into reviewable structure, with the source always linked
- First-pass feedback: flagging mechanics, completeness, and common errors so teacher time goes to judgment
- Accessibility: text leveling, translation support, transcription — quiet gains that change who can learn at all
Where it quietly fails #
It fails at novelty — generating lessons that look fresh but restate the textbook. It fails at calibration — questions that test wording rather than understanding, which is why human review of generated assessments is non-negotiable. And it fails at care: an AI cannot notice that a student's answers changed style after November, because nothing in the data model knows what November meant. Teachers read the room; models read the text.
The two hard lines #
First, AI should never be the final grader of consequential work — not because models cannot score, but because accountability for a grade must terminate in a human the student can face. Second, student data belongs to students: any tool that trains public models on your class's work has spent something that was not yours. Institutions that hold these two lines capture the productivity gains without inheriting the liabilities.