Artificial Intelligence
AI Didn't Break Learning. It Removed the Need to Try
When AI drove cognitive work instead of substituting it, learning followed.
Posted April 30, 2026 Reviewed by Michelle Quirk
Key points
- Students who used AI freely as a study aid remembered 11 percent less when tested 45 days later.
- When AI demanded engagement instead of bypassing it, learning markedly improved.
- The best AI for education isn't the one that answers fastest; it's the one that helps you think for yourself.
Call it technological synchronicity. Two studies on education and artificial intelligence (AI) have landed close enough together that the real story might lie in the space between them. Let's take a closer look.
The first, a randomized controlled trial from a Brazilian university, followed 120 college undergraduates. These students were given access to ChatGPT as a study aid. Forty-five days later, in an unscheduled retention test, they scored about 11 percentage points lower than students who had studied without AI. Prior experience with AI tools made no difference.
The second study, from the University of Pennsylvania's Wharton School, asked a different question: not the conventional question of what happens when AI substitutes for thinking, but what happens when AI is designed to prevent that substitution. In a cohort of 770 high school students learning Python over five months, researchers built a system that watched how students worked and continuously calibrated difficulty to keep each student in a zone where thinking was required. Those students outperformed their peers on a final exam by a margin researchers extrapolated to six to nine months of additional learning.
The difference between those two outcomes isn't in AI capability but in design intention. One system made thinking optional while the other made it unavoidable.
The Work Looked the Same; the Thinking Didn't
In the Wharton study, the improvement did not come from students doing more work. These students completed about the same number of problems and were not simply pushed toward harder material. What changed was the nature of their engagement. Students using the "GenAI tutoring system" spent more time on each problem, moving away from asking for answers and toward trying to understand why those answers made sense. Simply put, they were drawn back into the cognitive process that learning actually requires.
Students in the fixed-sequence group had access to an AI that could explain anything, and many used it to avoid thinking. Not out of laziness exactly, but because the system made that "path of least resistance" available. The Brazilian study makes the cost of that path clear. When students offload the cognitive work to AI, the material doesn't stick. What feels like understanding during the session turns out to be the AI's fluency, not the student's.
At the core of this is the uncomfortable reality that education has rewarded the "correct" answer for a long time. AI didn't invent that structure. It inherited it, and then it perfected the conditions in which thinking can be bypassed entirely.
Not Difficulty but Calibration
The Wharton adaptive AI system, interestingly, didn't just introduce difficulty. Difficulty alone produces confusion or disengagement, and harder questions did not independently explain the performance gains. What the system did was calibrate the challenge to a student's actual cognitive state. This is what I've called iterative intelligence and learner-centricity. The system recognized when a student was seeking an answer rather than building an understanding. It used those signals to keep each student in a zone where effort was required but not wasted. The system wasn't smarter than the students; it's my sense that it was simply watching more carefully than they were watching themselves.
This perspective frames what good AI design actually requires. The Brazilian students weren't failed by AI but by an implementation that optimized for immediate task completion rather than durable understanding. The Wharton system inverted that priority, and the difference in outcomes reflects exactly this inversion.
The System Saw What the Student Couldn't
Now, here's something to think about. The student, certainly an imperfect measure, has long been treated as the primary authority on their own cognitive process. Both studies, from opposite directions, challenge that assumption. The Brazilian study shows that students using AI reported no awareness that their retention was being undermined. The Wharton system, meanwhile, supplemented and even displaced student self-assessment entirely, inferring cognitive state from behavioral patterns the students were unaware of. The student who keeps asking for answers rather than explanations is not making a deliberate error. The system reveals that the underlying and essential work of understanding isn't happening.
Who Benefits, and Who Doesn't
The effects in the Wharton study were not evenly distributed. Students with little prior experience gained the most. Those with stronger backgrounds saw little measurable change. This suggests these systems don't simply improve outcomes uniformly. They reshape the distribution of cognitive effort, providing structure to learners who lacked it while leaving intact for those who already had it. The Brazilian study adds a parallel asymmetry: Technical topics showed the largest retention deficit under unrestricted AI use, precisely the areas where AI assistance feels most useful and where productive struggle may be more necessary.
When Thinking Becomes a Choice
To me, the concern isn't just that AI replaces thinking, but that AI makes thinking optional. When answers are instantly available, the path of least cognitive resistance becomes very difficult to resist. Students can produce work that carries the facade of understanding without traversing the effortful path that generates it.
AI did not create that condition, but it has made disengagement easier to sustain than it has ever been. What these two studies suggest, taken together, is that the problem was never simply access to information. It was about whether thinking itself remained necessary. The best AI systems, it turns out, are not the ones that answer fastest; they're the ones that notice when you have stopped thinking and make it harder to get away with. That may feel counterintuitive in a world optimizing for convenience, but the evidence is starting to accumulate. AI didn't break learning; it just made it optional. Whether that becomes permanent is still, for now, a choice of both technology and pedagogy.
