The third adult in every classroom app

The pitch for AI tutors is that content is no longer the bottleneck: a model can explain a fraction in Hindi to one child and a quadratic in Tamil to another, patiently, at midnight, for a fraction of a rupee. The pitch is correct about content. It is silent about the input that every study of software tutoring in India has quietly depended on, which is an adult who makes the child sit down with the device and keeps them there. That adult is not the teacher and not the parent. They are a third adult, and they are the unpriced line in every edtech budget, and whether anyone pays for them decides whether the tutor reaches the children it was built for.

Where the evidence came from

The strongest Indian evidence for computer-assisted learning is the evaluation of Mindspark by Muralidharan, Singh and Ganimian, published in the American Economic Review in 2019. Children in Delhi won lottery access to the programme, and after four and a half months the winners scored 0.37 standard deviations higher in maths and 0.23 higher in Hindi than the losers; the authors' estimate for ninety days of actual attendance is 0.6 and 0.39. Those are among the largest effects ever measured for an education intervention at that cost. And the children got them at a centre: a physical location with staff, where they came after school, sat at a machine, and worked through the adaptive software under supervision.

The "per ninety days of attendance" estimate is the tell. It is the effect on children who actually showed up, and the reason the intention-to-treat number is smaller is that many did not. Attendance was the variable, the centre and its staff were what raised it, and the effect sizes that get quoted in pitch decks are effects of software plus an adult, measured in the one setting where the adult was guaranteed.

Three adults, three costs

Around every child using a learning product there are three adults, and each costs something different. The teacher is paid by the school and has thirty other children. The parent is unpaid, often working, and, in the households where the product would matter most, may not read the script the product is in. The third adult is the one who makes the product actually get used: the centre staffer, the volunteer, the older sibling, the community tutor, whoever is in the room when the child would otherwise put the device down. In the Mindspark centres that adult was on the payroll. In a phone at home there is nobody, unless someone is paid or persuaded to be.

The failure modes are different too, and they compound. A teacher fails on time: there is no way to supervise thirty children on thirty devices while teaching. A parent fails on availability, and on the specific fact, visible in the ASER data, that a phone in a rural household is usually the parent's and usually needed elsewhere; only 57 percent of rural teenagers who could use a phone had used one for anything educational in the survey's reference week, with no third adult in the picture at all. The third adult fails on budget, because the moment a product is deployed at home rather than at a centre, the line for that adult goes to zero and nobody notices, since the effect sizes in the deck were measured with the line fully funded.

The third-adult line

The rule I apply to any child-facing learning product, including the ones people bring to me as a funder, is that the budget must carry a third-adult line. It can be a salary for centre staff, a stipend for community volunteers, a payment to a local tutor for hours of supervised use, or a documented plan for how an existing adult will be made to play the role. What it cannot be is zero, unless the product's claimed outcomes have been measured with it at zero. If the line is zero and the deck cites Mindspark, the projected outcomes belong to the centre study, not to the app, and the funding decision should be made on whatever evidence exists for unsupervised use, which is thinner and smaller.

What a funder can ask

The line turns into three questions that fit on the first page of any proposal, and that I now ask before anything else. Who is the third adult in your deployment, by name of role: centre staff, a school's own teachers during a timetabled lab period, a paid community tutor, a volunteer, an older sibling? What is that adult's cost per child per month, including the cost of recruiting and keeping them, since a volunteer programme with high turnover has a cost that does not appear on a payroll? And what happens to your outcomes when the adult is absent, which is the question the deployment will answer whether or not the proposal does, because the adult will be absent some of the time?

The answers sort proposals quickly. A programme deployed in school labs during timetabled periods has a third adult, the teacher, whose cost is teacher time, and whose failure mode is that a timetabled period is thirty children on fifteen machines; that is a real answer with a real cost. A programme sent home on a phone with a nudge campaign has answered the first question with "the notification", and the honest response is to ask for attendance data from a pilot before believing any effect size, because the effect sizes in the literature were bought with a person in the room.

Where a model can shrink the line, and where it cannot

The honest question for AI tutors is whether the model can do some of the third adult's job, and the answer is some of it. The parts of supervision that are about the session itself can be partly automated: short sessions that finish before attention does, nudges that bring a child back when they drift, and a design that makes the next step obvious so that the child does not need someone to say what to do. My own reading of what large models and retrieval do well is that they can hold a child's attention on a task better than static software could, and that is a real reduction in how much adult time each child needs.

What the model cannot do is get the child into the room. It cannot make a phone available at seven in the evening in a house where the phone is a parent's and the parent is out. It cannot notice that a child has not opened the app for a week and walk to their house. It cannot sit with a child who is embarrassed to be behind and make it safe to be behind. Those are the third adult's job in their entirety, and they are the part of the job that the centre studies were really measuring. A budget that funds the model and not the adult has bought the smaller half.

The rule, then, is not against AI tutors. It is for pricing them honestly. Put the third-adult line in, decide who fills it and what it costs, and let the model shrink the line where it genuinely can. If the line is zero, say so, and claim the outcomes that unsupervised software has earned rather than the ones a staffed centre did. The children who most need the tutor are the ones for whom the third adult is hardest to find, and pretending the line is free is how they end up with the device and without the learning.