• Safwan AMM
• 13 September, 2025

AI Reasoning: A New Species of Thinking or Just a Cage?

We Sri Lankans know this well: whether it’s solving a tough math problem for O/Ls, figuring out a cricket strategy, or even running a small business, the way we think step by step matters. Teachers always say: “Show your working.” Because the steps reveal the logic, not just the final answer.

That’s why AI’s popular method called Chain of Thought (CoT) feels so powerful. It looks like a student carefully writing step-by-step answers. But a recent study (preprint, not yet peer-reviewed) shows something surprising: those steps might not be “thinking” at all… but more like a mirage inside a cage.

A Model Built From Scratch

Instead of using big names like GPT-5 or Grok, researchers built their own AI from zero. Why? To remove noise, to avoid “cheating” by pretraining.

They wanted a clean test ground. Imagine a tuition teacher creating brand-new sample papers so students can’t copy old answers. Only then you see the real ability.

The Cage Test

Researchers tried three “nudges” to push AI outside its comfort zone:

• Changing the task – mixing familiar skills in a new way.
• Changing the length – making problems longer or shorter.
• Changing the format – asking the same question in a different style.

The result? AI failed. Even small changes made the step-by-step reasoning collapse. It could only walk inside the training cage. Outside, it lost direction.

Like a bright student who memorizes past papers, but panics when the exam gives a slightly different twist.

Humans vs. AI

Here’s the real difference.

• Humans carry lessons across contexts. A farmer in Kurunegala learns from planting paddy and applies similar tricks when planting bananas.
• AI does the opposite. It performs only when the pattern is familiar. Take it outside the pattern, it stumbles.

That’s why the study calls CoT reasoning a “brittle mirage.” I’d call it reasoning in captivity.

Bigger Models, Bigger Cage

Yes, bigger AIs like GPT-5 have more power. But the cage just becomes bigger and more comfortable. The bars are still there. The smooth fluency of AI answers makes us forget the limits—but they are real.

It’s like upgrading from a small trishaw to a luxury bus. You travel more comfortably, but you’re still on the same road.

Takeaway for Entrepreneurs & Students

For Sri Lankan entrepreneurs, teachers, or students using AI daily—remember this:

• AI is not a human copy.
• It’s not better or worse, just different.
• It thrives inside patterns, but struggles outside.

The lesson? Use AI for what it does best—pattern-based tasks, data crunching, drafting. But don’t expect it to “think” like us. Our adaptability is still unmatched.

Final Thought

We’re not dealing with a failed human brain. We’re looking at a new species of reasoning. Still confined, still caged, but unique in its own way.

The challenge is not to force AI to become human—but to understand what it is, and how it can help us without losing sight of its limits.

Just like any good entrepreneur knows: use the right tool for the right job. A hammer can’t do the work of a pen, and AI can’t yet do the work of true human reasoning.