What Should AI Do for the Learner? A Conceptual Framework for Education and EdTech

Two of the country's largest school systems have recently moved to restrict students' use of generative AI. The decisions in New York City and Los Angeles reflect an understandable concern: AI can now perform much of the work students have traditionally been asked to do.

But as schools move from reacting to that capability toward deciding what comes next, I think we may need to change the question.

Much of the conversation has focused on how to keep AI from doing students' work, prevent cheating, and teach students to use AI responsibly. Those are necessary conversations, but responsible use is difficult to define without first knowing what we're trying to accomplish. Using AI to summarize a text might be perfectly appropriate in one learning experience and undermine the entire purpose of another. The technology hasn't changed. What the learner is supposed to be developing has.

Perhaps the question we need to get better at answering is not simply when students should be allowed to use AI, but how we use AI to make learning better and develop a more capable learner.

Better Work Isn't Necessarily Better Learning

A few weeks ago, I spoke with students who participated in the National AI Student Senate and helped develop the STUDENTS FIRST Act of 2026. I asked where they would draw the line between AI helping someone learn and AI doing the learning for them.

Their answers varied by subject and situation. AI might generate additional math problems after a student has completed a study guide. It might help a student explore ideas. But asking AI to complete the actual work a student is supposed to be learning to do felt fundamentally different. One student offered a useful test: after using AI, could she actually say that she had learned something?

I've kept thinking about that distinction.

AI makes it unusually easy to improve the thing a student produces without necessarily improving the student who produced it. A better essay doesn't necessarily mean a better writer. A correct solution doesn't necessarily mean greater mathematical understanding.

The problem isn't that AI did something useful. It's that we haven't always distinguished between the work required to produce something and the work required to develop the person producing it.

Before deciding what work AI should take on, we need to understand what doing that work was supposed to develop in the learner.

Preserve. Reallocate. Extend.

I've started thinking about this as a possible conceptual framework for AI-enabled learning: Preserve, Reallocate, and Extend, with a fourth consideration—Sequence—running across all three.

It isn't intended to prescribe which uses of AI are appropriate. Those answers will vary by discipline, learning objective, and learner. Instead, it offers a way to start with the learning rather than the technology.

Preserve

What needs to remain with the learner because doing it is part of developing the intended capability?

Some work isn't merely a means of getting to an answer. The work itself is doing the educating. Writing can develop the ability to organize thought and construct an argument. Working through a mathematical problem can develop reasoning. Discussion requires students to articulate ideas, encounter disagreement, and sometimes revise what they think.

If AI removes the part of an experience through which the desired capability develops, its efficiency may actually work against learning. That doesn't mean struggle is inherently valuable or that students should always do things the hard way. It means there are circumstances in which the doing cannot be separated from the development.

Those are the parts we need to preserve.

Reallocate

Other work consumes time and attention without necessarily being the thing we most want students to develop. AI may be able to take on some of that work so the learner can direct more effort toward something educationally consequential.

A student might spend substantial time gathering, formatting, sorting, or performing some other necessary but secondary task. If AI reduces that burden and the reclaimed time goes toward analysis, discussion, experimentation, revision, or deeper investigation, then AI hasn't removed important learning. It may have reallocated effort toward it.

This is why efficiency alone tells us surprisingly little about educational value. Sometimes AI's efficiency is the danger because it removes the work through which a capability develops. Sometimes its efficiency is precisely the benefit because it creates more room for that development.

Extend

Then there is the possibility I find most interesting: What can the learner now do that wasn't previously practical or possible?

Educational technology is often introduced by inserting it into an experience we already have. We take an existing assignment or lesson and ask how technology might make it faster, easier, more personalized, or more engaging.

AI gives us reason to ask more ambitious questions.

Could students investigate information at a scale that previously would have been unmanageable? Explore relationships among ideas they couldn't reasonably examine before? Test many more possibilities? Interrogate competing explanations? Model something that previously required expertise or resources unavailable to them?

The meaningful possibilities will differ enormously by discipline. But if AI is genuinely a new intellectual tool, we should be looking for places where it can extend what learners are capable of investigating, creating, understanding, and experiencing.

Preserve protects development. Reallocate creates capacity. Extend creates possibility.

And Then There Is Sequence

There is a complication running through all three: the right answer can change as the learner develops.

Consider summarization. If a student is learning to synthesize information, asking AI to summarize a text may remove exactly the intellectual work the student needs to practice. Later, once that capability has developed, AI-generated summaries might allow the same student to work across a much larger body of material and devote more attention to analysis.

The AI capability is identical. Its educational value changes because the learner and the purpose have changed. That's Sequence.

Something that needs to be preserved today may reasonably be handed off tomorrow. The question becomes not simply whether AI should perform a task, but when this learner is ready for AI to perform it.

A Different Challenge for Education Organizations

If this is roughly where education is headed, the implications extend beyond classroom AI policies.

Professional learning organizations may need to help educators identify which forms of work actually develop particular capabilities. Curriculum organizations may need to design learning experiences around what should remain with the learner, what AI can reasonably take on, and what new forms of inquiry it can make possible.

For edtech companies, the challenge may be even more consequential.

Many AI experiences today put the decision primarily in the hands of the user. There is a chatbot, assistant, or button, and when the user asks for help, the system tries to provide useful assistance. But educational AI can know something a general-purpose assistant doesn't necessarily know.

It may know the learning objective. It may know the subject and grade level. It may know where a student is in an instructional sequence, what the student has already demonstrated, and what capability the current experience is intended to develop.

That creates a very different question for educational AI.

Not simply: Can I help with this?

But: Should I help with this, in this way, for this learner, right now?

Imagine a student asks an AI system to perform something it is perfectly capable of doing. The system recognizes that providing the answer would perform the exact intellectual work the student is currently supposed to develop. Instead, it asks a question, provides a scaffold, offers another example, or creates additional practice. That's Preserve.

Later, the learner has demonstrated that capability and is working toward something else. The system can now take on some of that same work so the learner can devote more attention to the new objective. That's Reallocate, governed by Sequence.

Elsewhere, AI might introduce an experience the student couldn't reasonably have had before, working across a huge body of information, manipulating a complex model, testing dozens of possibilities, or encountering perspectives and simulations that expand the learning environment. That's Extend.

If that is where educational AI is headed, simply adding generative capabilities to products won't be enough. The behavior of the AI itself may need to be governed by a model of learning.

An educational AI system may need to understand not only what a learner knows, but what the learner is trying to develop. Not only what assistance is technically possible, but what assistance is educationally appropriate at that moment. And perhaps most importantly, it may need some conception of what should remain with the learner.

That's a very different product-design challenge from building a better chatbot.

From Managing AI to Designing Better Learning

The first phase of education's response to generative AI has understandably focused on what students shouldn't hand over. The next phase will be harder because it requires us to decide what should remain with the learner, what can reasonably be handed over and when, and what entirely new forms of learning this technology should make possible.

Preserve the work that develops the learner. Reallocate work when doing so creates capacity for more consequential learning. Extend learning where AI makes something meaningfully new possible. Sequence those decisions around what the learner is developing and what they're ready to hand off.

If we can get better at making those distinctions, they could shape more than classroom rules. They could inform curriculum, professional learning, student AI literacy, and eventually the behavior of educational AI itself.

We've spent a lot of time asking how to keep AI from doing the learning for students. That was probably necessary. But it can't be the question that defines what comes next.

The more important question is how we design learning so that AI helps the learner become more capable.

Continue the conversation

If this article sparked a new way of thinking, you'll find more original perspectives, conversations, and insights at InflectionED.com.

Listen to my full conversation with Jeffrey C. Riley, Shamitha Doki, and Nandana Warrier on Education Talk Radio.