Did Complexity Just Break AI’s Brain?

There’s a curious irony in the world of artificial intelligence that has really gotten me thinking. In fact, I've often discussed the nature of "thought" in the context of large language models. From "cognitive theater" to "technological architecture," I've studied these thinking machines and explored the illusion of fluency masquerading as thought. Now, Apple’s new research on reasoning models takes a close look at this very issue.

In their new report, The Illusion of Thinking, Apple researchers opened the hood on a class of large language models that specialize in “reasoning.” These aren’t just your average autocomplete engines. These are what some have called large reasoning models, or LRMs. These are AI systems trained to provide multistep, chain-of-thought responses (CoT) that mimic human logic and deliberation.

But here’s the catch, and I think it's a big deal. The Apple researchers suggest that beneath the seductive prose and logical scaffolding, these models often fail at the very thing they appear to be doing—reasoning.

A Collapse in Complexity

The report outlines a dynamic that should give AI optimists reason to pause. It seems that as the complexity of problems increases, the performance of these reasoning models does more than just decrease; it collapses.

At low complexity, simpler LLMs outperform more advanced models. At medium complexity, reasoning models shine. But as the cognitive load increases, when abstraction or multistep logic is required, they falter. Or, as the authors expressed:

“We identify three performance regimes: (1) low-complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse.”

Worse still, the models don’t seem to know they’re failing. They produce what appear to be sound, step-by-step answers. To a lay reader, or even a trained one, these outputs can seem rational. But they’re not grounded in any algorithmic or consistent method. They’re approximations of logic based on semantic coherence.

Fluency Isn't Thought

This should force us to confront a curious and troubling possibility. What if these models are only simulating the structure of thinking? What if chain-of-thought prompting is less a window into machine reasoning than a mirror reflecting our own cognitive biases and our tendency to confuse coherence with truth and verbosity with understanding?

And of course, there's some human logic behind this. We are often persuaded by narrative. A well-structured explanation carries an aura of authority. But AI exploits this bias at scale, and its fluency creates a false signal. It doesn’t just mimic reasoning, it performs it in a way that I've argued is antithetical to human thought.

And I believe that this performance has consequences. In medicine, law, education, and mental health, LLMs are being considered as decision-support tools. If we’re building systems that fail at complexity while appearing competent, we risk introducing a new kind of cognitive hazard. The concern isn't just that AI is wrong, but that AI is convincingly wrong.

The Cognitive Mismatch

What Apple’s findings seem to highlight is a growing divergence between human cognition and artificial coherence. Human cognition thrives in the friction of thought and adaptive strategies like analogy and metaphor. LLMs, in contrast, optimize for surface-level fluency where tokens are aligned with tokens, not truths aligned with truths.

The more LLMs look like they’re thinking, the more we’re fooled into believing they are. But if these systems can’t scale their reasoning with complexity, they are merely rhetorical engines, not cognitive ones.

So, Where Do We Go From Here?

We should begin by exercising more skepticism—not cynicism, but critical curiosity. AI may produce detailed reasoning traces, but without consistency, they aren't explanations; they're performances. And my sense is that many have grown far too comfortable mistaking performance for proof.

We also need better tools—not just to evaluate AI’s answers, but to understand its methods. Benchmarks that test outcomes are no longer enough. We need to interrogate the process behind the prose.

And finally, perhaps most importantly, we need to acknowledge that we are only beginning to understand what artificial “thinking” really is. Intelligence, it turns out, may not scale smoothly or mimic human cognition in any straightforward way. And in that uncertainty lies both the danger and the responsibility of this moment. There’s more to uncover. Much more.