The Illusion of Thinking: What Apple’s Puzzle Study Means for Classroom Learning

If you’ve ever watched a student wrestle with a tough math problem—backtracking, second-guessing, trying again—you’ve seen reasoning in action. It’s messy, nonlinear, and deeply human. That’s why the recent Apple paper “The Illusion of Thinking” on Large Reasoning Models (LRMs) caught my attention. As someone who’s spent years teaching in rural classrooms and designing future-ready education systems, I found the study’s implications both sobering and strangely affirming.

In essence, Apple researchers examined how the newest “thinking” AI models, like Claude 3.7 Thinking and DeepSeek-R1, perform when solving puzzles that increase in complexity—tasks that require multi-step, rule-bound planning like Tower of Hanoi or River Crossing. These puzzles, unlike many benchmark datasets, can be adjusted in complexity while keeping the underlying logic intact. What they found should give every AI optimist pause.

Despite all the Chain-of-Thought prompting and reinforcement learning polish, these models collapse under pressure. As task complexity increases, their reasoning effort—measured by the number of tokens used to “think”—starts to decrease. Even when given the exact algorithm for solving a puzzle like Tower of Hanoi, they can’t reliably follow it. They might overthink a simple problem or give up halfway on a harder one. In the classroom, we’d call this cognitive overload—or sometimes just being unprepared.

Here’s what hit me hardest as an educator: these models don’t generalize well. They’re brittle. They memorize surface patterns, but when it comes to adapting logic across tasks—something even a middle schooler can eventually learn with scaffolding—they fall apart.

And this has massive implications for education.

We’re now in a moment where schools across the country—especially rural ones like those I serve in West Virginia—are being told AI will be the great equalizer. But if AI models can’t reason through problems the way students must, then we have to be careful about how we frame their use in classrooms. They're tools, not tutors.

This research reminds me that real thinking—real learning—isn’t about speeding through a problem to get the right answer. It’s about struggle. Reflection. Trying and failing and trying again. It’s about knowing why you’re making a choice, not just what choice to make. Our students don’t just need answers. They need models of how to think clearly when things get hard.

So where does that leave us?

I’d argue that we educators need to double down on teaching metacognition, persistence, and transferability—the very things these LRMs lack. AI might help us generate content or support low-level tutoring tasks, but it’s not ready to model what complex, adaptive reasoning really looks like.

As we integrate AI into classrooms, we must remain vigilant not to outsource our most human capacities. “Thinking,” as this study shows, is still more illusion than reality in machines. That means the best teaching remains deeply personal, deeply human—and as relevant as ever.