Is the Failure to Reason Proof That AI Is Reasoning?
We expect smart systems to get better when the going gets tough. After all, that’s what intelligence is for, right? But what if the opposite happens? What if a machine that fails to reason under pressure is actually showing us that it’s reasoning, just not the way we expected?
When AI Tries (and Fails) to Think
Today’s large language models (LLMs) like Claude, DeepSeek, and OpenAI’s o-series have amazed the world with their ability to explain ideas, solve math problems, and even reason through puzzles. More specialized versions — sometimes called Large Reasoning Models (LRMs) — go even further, producing detailed, step-by-step chains of thought to tackle complex tasks.
And yet, researchers have found something strange: as you increase the complexity of a problem, these models don’t just slow down or get a bit less accurate. They collapse.
Performance falls off a cliff. They may use fewer reasoning steps. Their logic becomes inconsistent or loops. They “give up” — despite having more than enough space to keep thinking. It looks like a bug. But what if it’s a feature?
Humans Do This Too
When people try to solve hard problems — like figuring out a tricky logic puzzle or debugging a code error — we don’t always succeed. Our working memory runs out. We fixate on a bad strategy. We second-guess ourselves. Sometimes, we give up.
Cognitive scientists like Herbert Simon and Daniel Kahneman have shown that humans are bounded reasoners. We don’t reason like flawless machines; we reason within limits. We use shortcuts (heuristics). We satisfice instead of optimize. And we can overthink problems just as easily as we can underthink them.
Sound familiar?
The strange collapse behavior of AI models at high complexity echoes these very patterns. Like humans, they show signs of bounded effort: trying harder up to a point, then suddenly retreating. They don’t fail randomly. They fail predictably. Structurally. Cognitively.
Reframing the Narrative
Much of the current debate around AI’s reasoning ability assumes a binary: either the model can reason, or it can’t. Either it gets the right answer, or it’s just faking it.
But this is an outcome-based view. It’s like judging a student only by whether they got the answer right, not whether they understood the process.
What if we flipped it?
What if the attempt to reason — even if it fails — is more telling than the result?
After all, when a model breaks a problem into subgoals, reasons step by step, and only falters late in the chain, that’s not the behavior of a parrot spitting out memorized patterns. That’s a system trying to reason — and hitting its limits.
In other words: the collapse itself may be the strongest evidence that the model is reasoning at all.
A Diagnostic, Not a Defect
This perspective has major implications:
In AI development, it suggests that we should study reasoning failures not just to patch them, but to understand what kind of cognition is emerging inside these models.
In AI safety, it means we need better ways to detect when a model is reasoning well versus when it’s bluffing — and collapse indicators (like abrupt token drop or hallucinated logic) might be key signals.
In education, it invites us to see AI not just as a tool for delivering answers, but as a thinking partner whose errors — like a student’s — can be instructive, if we know how to read them.
It also reframes what success looks like. A perfectly correct answer with no reasoning trace is less trustworthy than a flawed answer with a clear, interpretable process. We should ask how a system got to its answer — not just whether it got there.
Introducing the Paper: The Collapse of Thought
These are the questions we take up in our new paper: “The Collapse of Thought: Reconceptualizing Reasoning Failure in Large Language Models as Evidence of Cognitive Authenticity.”
We argue that what’s been interpreted as a failure — the collapse of reasoning under high complexity — is actually a hallmark of reasoning itself. Drawing on research from AI, cognitive science, and philosophy of mind, we propose a process-based definition of reasoning and apply it to the behavior of frontier AI models.
Along the way, we:
Analyze empirical findings from recent studies, especially The Illusion of Thinking (Shojaee et al., 2024), which documents reasoning collapse in top-tier models.
Show how LLMs’ failure modes resemble human cognitive limitations — including overthinking, heuristic shortcuts, and early termination.
Reframe reasoning “collapse” as bounded cognition, not incompetence.
Explore how this view reshapes education, AI evaluation, and the symbolic-vs-statistical debate.
We close with a philosophical and practical vision: AI systems that struggle visibly may actually be more useful — and safer — than ones that produce confident answers with no interpretable process. The goal isn’t perfection. The goal is transparency, humility, and collaboration.
If you’re curious about the future of AI reasoning — not just where it works, but where it breaks — this paper is for you.
The paper can be accessed here.