On 6 October, the University of Pennsylvania’s Graduate School of Education announced a Living Laboratory for Human Agency and AI, bringing educators, researchers, school systems and industry into a network to design and test educational AI. This establishes an initiative, not a learning breakthrough. Its value will depend on whether it helps schools distinguish a better-looking answer from a more capable student.

That distinction should govern the money. Education buyers should pay for understanding that students retain when assistance is withdrawn, while funding the teaching needed to build it.

Consider a teenager who submits a flawless solution but cannot explain the first step. The assignment records success; the learner has acquired a dependency. Now imagine a tool that offers a hint, asks the student to try again, and helps a teacher identify the misunderstanding. Both systems might report an active user. Their educational value could be opposite.

A 2025 randomized study in a Turkish high school makes this concern concrete. Among nearly a thousand mathematics students, those given a standard AI chat interface did better during supported practice but scored 17% lower than the control group when subsequently working without it. A version designed with teacher-informed safeguards largely removed that harm, although it did not establish an improvement in unaided performance.

This is one setting and an earlier generation of tools, not a verdict on every AI tutor. It does establish why supported performance is an inadequate measure of learning.

The strongest argument for adoption deserves equal attention. For a child without affordable tutoring, patient explanations and repeated practice could be valuable. Delaying access until every uncertainty disappears could protect the advantages of families already able to purchase help.

There is encouraging evidence. World Bank researchers reported improved assessment results from a six-week, randomized after-school programme in Edo State, Nigeria. But it combined AI with structured prompts, curriculum alignment, teacher support and peer learning. The authors explicitly caution that they cannot isolate the chatbot’s contribution, and that the durability of the gains remains uncertain.

The sensible choice is therefore supervised experimentation, with expansion tied to evidence. A successful educational service includes more than a software licence.

Teachers must prepare activities, notice errors and decide when a learner needs a different explanation. The Turkish experiment’s safeguards themselves incorporated labour-intensive, problem-specific input. Treat that work as overhead to eliminate and a buyer may remove the very complement that makes the technology useful.

This changes the economics. Cheap feedback could stretch scarce teaching capacity and reduce families’ need for paid coaching. Those are possible gains, not established savings from the new Penn initiative. They must be assessed against devices, connectivity, preparation, supervision and independent evaluation. Cost per enrolled account tells us little if students stop attending or cannot solve unfamiliar problems.

The distribution matters as much as the average. In ASER 2024’s rural Indian survey, almost 90% of 14–16-year-olds reported a smartphone at home. Yet only 70.2% of boys and 62.2% of girls could bring a phone for the survey’s digital tasks. Household connectivity did not mean equal practical access.

An Indian programme should therefore test whether a student can actually use it at the required time, in a language they understand, with accessible options for disabilities and weak connectivity. Otherwise, apparent low engagement may reflect exclusion rather than lack of motivation.

Three futures follow from these choices. In a constructive one, schools combine affordable practice with teacher attention, and students demonstrate understanding months later. In an uneven one, affluent schools purchase both technology and expert support while poorer schools receive automation as a substitute for attention. In a harmful one, polished submissions conceal weaker foundations, leaving learners reliant on services they must keep paying to use.

These are scenarios, not forecasts. Contracts can make them more or less likely. Buyers should require delayed assessments, unfamiliar problems, results broken down by access and disadvantage, and publication of disappointing findings. Schools also need usable teaching materials when a vendor changes prices or exits. Otherwise, public investment can build dependence on a provider rather than durable educational capacity.

Can students explain and apply what they learned when the tool is unavailable? Does the business model reward that progress? Have costs included teachers’ preparation and the support needed by excluded learners?

Fund pilots that answer those questions, give teachers paid time to shape them, and expand what demonstrably works. The meaningful return is what a student carries away.