The Illusion of Thinking: Why AI Reasoning Models Aren’t Really Thinking

When you read an answer from a modern AI, it feels like the model is “thinking”. The step-by-step reasoning, the self-reflection, the way it works through puzzles, it mimics human thought almost uncannily. But a new paper from Apple researchers, “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem COmplexity”</strong></a>, makes one thing clear: models aren’t actually thinking.

Instead, they’re generating reasoning-shaped text. And when we probe them with controlled environments like Towers of Hanoi, Checker Jumping, River Crossing, and Blocks World, the illusion cracks.

The Regimes for “Reasoning”

One of the most interesting insights from the paper is that models behave in three distinct regimes:

1. Low complexity: Regular language models (without explicit reasoning mode) actually do better. They’re faster, more efficient, and more accurate, no overthinking involved.

2. Medium complexity: This is where “thinking” models shine. They explore longer chains of thought and sometimes manage to edge their simpler counterparts. The illusion looks strongest here.

3. High complexity: Both models collapse. Performance goes to zero. The reasoning traces shrink, effort decreases, and the model often latches onto an early wrong guess and never recovers. The curtain falls, and the illusion disappears.

This is powerful, because it shows us that “thinking mode” doesn’t scale into true generalizable reasoning. It only delays the inevitable collapse.

When More Thinking Becomes Less Thinking

Here’s the most counterintuitive finding: as tasks get harder, reasoning models think less.

Normally, you’d expect that if a problem is more complex, the model would use its token budget to generate more reasoning. But the opposite happens: they reduce their effort, even though they still have space left.

It’s as if a student facing a difficult exam question, wrote one line of scratch work and then gave up. Not because they were out of paper, but because they didn’t know how to proceed. The “illusion of thought” is revealed here: what looks like effort isn’t backed by actual problem-solving ability.

Overthinking, Fixation, and Collapse

When problems are easy, models do something oddly human: they overthink. They might land on the right solution quickly but then continue exploring wrong alternatives, wasting effort.

When problems get harder, the opposite happens: they fixate. They latch onto an incorrect path early on and keep digging deeper into the hole, unable to step back and correct themselves.

Humans can sometimes recognize, “Wait, this path isn’t working, let’s backtrack”, but models rarely do. The researchers call this “limited self-correction”, and it exposes another gap: the models don’t have a meta-level awareness of their own reasoning. They can’t truly evaluate their thoughts, only simulate what evaluation might sound like.

Even With the Answer, They Fail

The most striking experiment: researchers gave the models the actual algorithm to solve puzzles like Towers of Hanoi. The task was no longer to “figure out the solution”, but simply to execute the steps.

And still, the models broke down.

Think for a moment about that. EXecution should be easier than discovery. If you hand me the steps to cook a recipe, I might mess up the flavor, but I’ll still get something edible (maybe hehe). These models sometimes fail event at following the recipe.

This tells us they aren’t reasoning in a logical sense at all. They’re generating text that looks like reasoning. When the text patterns don’t align neatly with the structure of the problem, the illusion falls apart.

Why This Matters

It would be easy to dismiss this as “AI haters nitpicking”, after all, these models can still solve real-world problems, write essays, pass exams, and assist with creative tasks. Isn’t that enough?

But here’s why it matters: If we mistake the illusion of thought for real thought, we risk overestimating these systems. We might start trusting them in situations where robustness, consistency, and genuine reasoning are essential: science, safety-critical planning, or hight-stakes decision-making.

It’s one thing if an AI gets a math puzzle wrong. It’s another if we expect it to reason about drug discovery, legal arguments, or policy tradeoffs, only to discover that its reasoning collapses under true complexity.

Don’t Confuse Simulation with Thought

When I first experimented with “thinking models” I’ll admit: I was impressed. The chain-og-thought outputs feel so much more human. It’s tempting to imagine the model really grappling with the problem, almost like a student working things out on paper.

But this paper is a reminder: the thinking is a performance, not a process. These models are powerful pattern-matchers, not thinkers. THey’re not climbing the ladder of general intelligence, at least not yet.

That doesn’t make them useless. Far from it. But it does mean we should adjust our expectations. Treat their reasoning like a magic trick: enjoy the illusion as in a magic trick, but don’t mistake it for real thinking.

What do you think? Are we giving too much credit to AI models when we call what they do “thinking”? Drop your thoughts in the comments, I’d love to hear if you’ve spotted the illusion on your own use of these tools.