The AI Learning Trap |
AI boosted student performance 40 percent, but scores dropped 17 percent once the tool was removed.
Effortful learning builds critical brain connections that frictionless AI delivery may bypasses.
AI isn't just changing what students know; it's changing how they come to know it.
Something rather curious is happening in education. Artificial intelligence (AI) seems to help students produce better work more quickly. And in the process, perhaps even with less struggle. But there's a critical difference between arriving at an answer and building the mind that understands it. And that's worth a closer look.
A recent study from The University of Pennsylvania brings this into focus. Let's start with the data. Students using large language models (LLMs) improved their performance by about 40 percent. So far so good. But when the LLM was removed, performance dropped by about 17 percent. So, does AI exact some sort of cognitive toll on students, leaving them worse off than if they never used this technology in the first place?
This pattern may reflect something even more subtle than simple dependence. In an earlier post, I described what I called the AI rebound—a dynamic where performance rises in the presence of AI, but then falls once that support is removed. What’s important is that the drop doesn’t just reflect the absence of the tool. It may reflect a shift in how much cognitive work was done in the first place. The mind adapts to the availability of external reasoning, and when that support is taken away, the internal cognitive pathways or structure may not be as fully formed.
What I believe this study reveals isn't just the overreliance on a tool but a shift in where cognition takes place. The LLM is doing exactly what it is designed to do by reducing effort and providing an answer. In doing so, it relocates part of the thinking process outside the learner, allowing the task to be completed while altering how much of the work is internally constructed.
This creates a tricky scenario that can be easy to overlook. The student succeeds, often more efficiently than before, but less of the cognitive structure that supports that success is built. The outcome reflects competence, but the underlying cognitive architecture may be less well-formed.
When Performance Feels Like Understanding
The difference between performance and understanding is at the heart of this distinction. AI does more than provide answers. It "offloads" the thinking by providing the mechanics of thought that is served up with the touch of a button. It all comes cascading forward—from the outside in—as the typical dynamics of human thought. This output has the "texture" of understanding, and it signals that meaningful cognition has taken place. But has it?
The difficulty is that understanding formed over time leaves traces that extend beyond the immediate task at hand. It shapes how new problems are approached in the future. When that path is decreased or even eliminated, it becomes more visible when the support is no longer present. I think that's what we're seeing in this study.
What Friction Actually Does
Learning depends on a certain kind of resistance, but that idea is often left abstract. Friction isn't just difficulty for its own sake; it's the processes of neurogenesis and neuroplasticity that build connections in the brain. The time spent with a problem in mind recruits memory and builds connections in ways that smooth delivery does not.
When answers arrive too quickly, that generative process is shortened, if not completely eliminated. The learner moves forward with less opportunity to, simply put, learn. The experience feels efficient, but it's also less cognitive effort. Over time, that difference accumulates, and the student becomes more practiced at following solutions than generating them.
A Harder Way to Use AI
There is another way to engage with AI in education, but it does not emerge naturally from how these systems are designed. It requires resisting the efficiency that makes them so appealing. And it requires staying within the process of thought rather than leapfrogging past it. In this process, follow a dynamic path that can extend thinking rather than replace it.
I'll suggest that this approach has two key elements. The first is iterative engagement, where a back-and-forth exchange with AI builds understanding rather than shortcuts it. The second is learner-centricity, where the material isn't generated, but created, on the spot, in ways that align with a student’s interests and needs.
Now, there's a bit of a red flag here, too. This isn't simply slower; it is more demanding. It asks for sustained attention in an environment designed to reduce it, and it requires effort at the very moment when effort can be avoided. Most learners won't default to this mode, and most systems do not encourage it. Yet when it does occur, the nature of learning may change. The student becomes more involved in shaping the process that leads to the result. And importantly, the process begins to shift from completing a task to discovering the joy of learning.
The Choice That Isn’t Obvious
It's important that this study highlights a real and measurable risk. When AI is used primarily as an engine of completion, it can bypass the conditions that make learning durable. At the same time, it points to something broader. LLMs aren't just tools we pick up and put down. LLMs may function more like environments we enter that can shape the process of learning.
What's beginning to emerge is a pattern where cognition is partially externalized. And that's allowing the experience of knowing to remain intact even as the process evolves and changes. While this shift is subtle, effects can become clear and even alarming when technological support is removed.
For better or worse, AI isn't just changing what students know; it's changing how they come to know it.
There was a problem adding your email address. Please try again.
By submitting your information you agree to the Psychology Today Terms & Conditions and Privacy Policy