Ravi Gupta’s case for AI tutors collapses under the trial he cites. In his Wall Street Journal op-ed, “New York City’s School AI Ban Goes Too Far”, the former RePublic Schools CEO argues that banning generative AI for students through eighth grade is shortsighted because AI tutors are essential to differentiation, and that bureaucrats and teachers’ unions—not pedagogy—are blocking their adoption. The randomized Khanmigo study he cites found only “modest improvement.” He also says students left on their own “barely used” the tutor. Put those sentences together and the case points somewhere else.
It points to teachers.
Gupta is right about the problem. A single fifth-grade teacher can face twenty-five students spread across five levels of mathematics, from algebra to basic addition. The average student is a fiction. Differentiation is nearly impossible at that ratio, and the traditional alternative—tracked classrooms—clusters disadvantage. No child should be assigned to a permanent lower track because the adults around the system could not organize a better answer.
But the cure is not a chatbot that children barely open.
A tool that produces modest gains under supervision and little engagement when optional is not a substitute for a teacher. It is an adjunct that needs a teacher in the room to make it work. The mechanism Gupta presents as proof of scalability—the absence of a human in the loop—is the mechanism that turns the product into shelf-ware.
Gupta knows this. He concedes the answer in passing: “we should then train adults to motivate, coach and hold students accountable for their progress.” That sentence is the column he should have written. The constraint on American classrooms is not access to a chatbot. It is the slow bleeding of experienced teachers, the substitution of long-term substitutes for full staff, and the erosion of training pipelines.
You cannot tutor a child into caring about fractions if there is no adult in the room who knows the child.
The AI tutor, at that point, is an expensive reminder app. The reminders could arrive through a worksheet, a printed checklist, or a five-minute check-in from any adult in the building. None requires a large language model. None requires a Wall Street Journal editorial to bless it.
This does not mean AI has no use. Gupta is right that a tutor can adapt to individual gaps. But the cost of running that adaptation at scale is the cost of small classes: human time. Someone has to read the data, calibrate the tool, explain the mistake, and sit with the student when the tool gets stuck. Software does not eliminate the labor. It hides the labor and then calls the hiding efficiency.
His proposed 50/50 school day reveals the conceit. Half the day with technology; half without it—to read, work problems by hand, learn trades, exercise, and play with friends. That sounds lovely. It also sounds like a private school with a $40,000 tuition and a parent body that can lobby the headmaster.
New York City has 1.1 million students. The trades-and-friends half of the day needs actual shops, actual gym teachers, actual recess monitors, and actual adults who can keep a room from becoming an emergency. None of that arrives with a Khanmigo login.
The “AI tutors don’t pay union dues” line is the tell. Gupta wants to cast this as labor against progress. It is not. Teachers’ unions have plenty of problems. They have also not prevented a fifth-grader in the Bronx from reading at grade level. The relevant facts are less theatrical: we do not pay teachers enough to keep them, train them long enough to make them good, or staff schools at the ratios that make the work possible.
And yes, New York City spends more per student than most large districts, while children still leave school without what they need. That is an argument for asking where the money goes and what the staffing model demands. It is not evidence that children need more unattended software.
So build what works. Shrink class sizes at the bottom of the distribution. Put caps at eighteen students in kindergarten through third grade and twenty-two in fourth through twelfth. Pay teachers to coach through lunch, after school, and on Saturdays. Let schools run low-stakes pilots without permission slips from Tweed. Use Khanmigo honestly: supervised, narrowly scoped, and modest.
Then build the labor pipeline that Gupta’s own concession smuggles back into the room. Pay first-year teachers what second-year associates make at law firms. Pay master teachers what senior engineers make. Create paid, year-long student-teaching residencies. Give new teachers a full year of mentoring. Offer loan forgiveness that actually clears in five years, not ten.
AI can help where it helps. It can sort practice problems. It can flag a gap. It can give a teacher another view of the child in front of her. But it cannot supply the relationship, judgment, accountability, or roster flexibility that makes differentiation real.
What schools need is not a software pipeline. It is a labor pipeline: adults who know each student by name, and the time to teach the one standing in front of them.