The Machine That Thinks for Students
Skinner built a machine that required students to think. We built one that thinks for them.
In the 1950s, B. F. Skinner built what he called a teaching machine. It presented material in small steps. It required the student to produce a response before moving forward. It gave immediate feedback on every attempt. A student could not advance without demonstrating mastery of what came before.
The machine was primitive. The design started from an understanding of what learning requires. A student who is engaged. A response before advancing. Feedback right away. Steps that build on each other.
Seventy years later, we have built a far more powerful machine. It can write a paragraph, solve a problem, construct an argument, and explain the concept, instantly.
It is the exact inversion of Skinner’s design. His machine could not proceed until the student responded. Ours responds so the student doesn’t have to.
In the science of behavior, we always ask a simple question about any environment: What does it make more likely?
Learning requires a student to respond. A student who watches, listens, or receives information passively is not learning. Learning requires effort, practice, error, feedback, and adjustment. These are not preferences. They are the conditions under which skills develop, and they have been established across decades of research.
Generative AI can perform all of them on a student’s behalf.
When a tool completes the task, the student’s opportunity to respond is eliminated. Researchers call this cognitive offloading, relying on an external system to perform cognitive work, which leaves the person less able to perform that work independently. Offloading is ordinary behavior under new contingencies. For all of human history, effort was the only path to the outcome. Now there is a cheaper path. Behavior follows cost.
Students learn from what schools reward. Students learn from consequences. When they can consistently reach the outcome without doing the thinking, they learn that the thinking was never essential in the first place.
Unrestricted AI sends a similar message. It does not teach a subject, but teaches that effort is optional. It suggests that struggling to remember, reason, and solve problems is just a hassle to be avoided, instead of the way we actually learn to think.
The contingencies, though, can be changed. Instead of asking what AI can do, ask what students at each stage still need to do on their own, and protect those opportunities.
In the early grades, children are building the foundation for all future learning, so AI should be used by teachers, not students. In upper elementary, students do their own work first, then critique AI output under a teacher's direction. In middle school, AI becomes something students study, how it works, where it fails, while their own work still comes first. By high school, students who have built independent skills can engage with AI as critical users, in that order.
The main idea is the same as in Skinner’s machine: students must respond. Any tool added to a child’s learning environment should first be checked to see if it keeps this requirement.
The founders understood that you do not get good outcomes by expecting people to be better than they are. You get them by designing institutions that channel ordinary behavior toward the outcomes you want. The same is true in classrooms. We should not expect children to resist, through willpower, a machine engineered to remove effort. We have to design environments where effort remains the path.
Skinner’s teaching machine never became the future of education. But he started from the right question: not what the machine could do, but what it required the student to do.
That is the question I argue schools should be asking about AI. In my new Manhattan Institute issue brief, One Size Does Not Fit All: A Developmentally Appropriate Framework for AI in K–12 Education, I propose evaluating AI not by what it can do, but by what students at different stages of development still need to do for themselves, and building AI policy around protecting those opportunities.
Seventy years later, the question has not changed. Only the machine has.
Correct! AI most certainly needs to be addressed by states as they work to reform education standards.
It’s hardly a coincidence that much of the foundational psychology for AI, cognitive psychology, is predicated on an express _rejection_ of behavioral science and, of course, Skinner’s findings.