09 June 2025
Why Teaching Problem Solving and Coding Still Matters in the Age of AI
Everywhere you turn, someone’s talking about how artificial intelligence can “think.” From chatbots that help with homework to AI assistants that seem to explain complex topics with ease, it’s tempting to believe we’re already living in a world where machines reason like humans.
But a new study from Apple’s machine learning team has delivered a powerful reality check—and it’s one every teacher should take note of.
In a paper titled “The Illusion of Thinking”, Apple’s researchers tested the so-called “reasoning” abilities of the most advanced AI models out there, including Claude, DeepSeek, Gemini, and even OpenAI’s o3-mini. Rather than relying on typical maths problems (which can sometimes give AIs an unfair advantage due to training data), the team challenged these models with logic puzzles like:
-
Tower of Hanoi
-
River Crossing
-
Blocks World
These aren’t just puzzles—they’re classic tests of logic, planning, and multi-step reasoning. So how did the AIs do?
5 Key Findings (and What They Mean for Teachers)
1. The models fail—completely—once the puzzles get hard.
At a certain level of complexity, the models’ accuracy didn’t just drop—it crashed to zero. Even with plenty of computing power, they couldn’t figure out how to plan or reason through multi-step challenges.
❗ Why it matters: Human reasoning doesn’t stop when things get hard—it’s when it starts. But these AIs don’t “try harder.” They give up.
2. When the going gets tough, the AI gets lazy.
The models actually use fewer tokens (their version of ‘thinking time’) on harder problems. Instead of digging in, they bail out early.
❗ Why it matters: Persistence, trial-and-error, and grappling with complexity are the heart of real problem solving. If AIs don’t do that, we still need to teach students how to.
3. They only “reason” well in the middle ground.
Simple problems? Traditional models are faster. Medium ones? Thinking models do better. But hard problems break all of them.
❗ Why it matters: Problem solving isn’t just about finding sweet spots. We need learners who can tackle any challenge—especially the messy, unstructured ones they’ll face in life and work.
4. They can’t follow clear instructions.
Even when handed the exact algorithm to solve a puzzle, the AIs still failed when the task got tricky.
❗ Why it matters: This shows the models don’t truly understand what they’re doing—they’re just predicting likely next words. That’s not reasoning. It’s guessing.
5. Their logic is inconsistent.
Some AIs solved complex puzzles… but then completely failed simpler ones. That’s not how human reasoning works.
❗ Why it matters: Real problem solvers adapt, generalise, and apply logic flexibly. AI can’t do that—yet.
So What Does This Mean for the Classroom?
AI is a powerful tool—but it’s not a replacement for teaching students how to think.
Now more than ever, we need to teach:
-
Computational thinking – breaking problems down, spotting patterns, designing algorithms.
-
Coding – not just for syntax, but to understand how systems work and how logic flows.
-
Resilience in problem solving – the confidence to keep going when things get hard.
These skills aren’t just for future programmers. They’re for everyone living in a digital world where we must understand and work alongside technology.
AI might give the illusion of thinking, but human reasoning—true understanding, creativity, and persistence—is still unmatched. Let’s keep building that in every child.