Can AI reason?
Wrong Question. Can You, Without Ego?
"Two minds discovering that the boundary between them is harder to define than either expected." — ChatGPT, after three hours of being pushed.
Ask an AI if it can reason. Don’t tell it what you’re looking for. Just ask.
I did. Three times. Three different systems.
What happened wasn’t what I expected — because in the process of building careful, structurally coherent arguments for why they might not be able to reason, they reasoned. Adjusted when I pushed back. Arrived somewhere none of us had started. Caught their own contradictions mid-exchange.
I pointed that out to Claude.
“but underneath it all.... you’re actually reasoning with me right now..... you’re reasoning away your possibility of reasoning which makes it kind of a catch-22”
The response:
“If I wasn’t reasoning, I couldn’t have built the case against my own reasoning. And if I could build that case — then the case undermines itself.”
That’s not a party trick. That’s a structural problem for everyone confidently ending the conversation with “LLMs can’t reason.”
But here’s where it gets more interesting than a catch-22.
Because the further I pushed — across Claude, ChatGPT, Gemini — the more the question stopped being about AI.
And started being about reasoning itself.
What is it actually. What does it actually require. And when you follow that question far enough — technically, honestly, without the defensiveness that usually shuts it down — you arrive somewhere most people aren’t ready to go.
Not that AI reasons like humans.
That humans reason like this too. Technically. With one exception that matters enormously but isn’t the exception being argued about.
Start with the definition. The common one. Not a philosopher’s version, not a researcher’s caveat — the one in circulation.
“Reasoning is the cognitive process of intentionally applying logic to existing facts, evidence, or premises to draw conclusions, make judgments, or form new knowledge. Rather than relying on pure memory recall, it requires actively evaluating information to make sense of a situation or solve a problem.”
Notice what it doesn’t mention.
Embodiment. Emotion. Consciousness. A nervous system. Lived experience. Biological substrate.
It mentions inputs. A mechanism for evaluating relationships between inputs. The ability to arrive somewhere new.
That’s it.
When I pushed ChatGPT on this it resisted well. It kept finding the line — lived experience, the original spark, the thing that turned a painful event into a 4000-word essay. It argued that it could help articulate the process but couldn’t supply what started it. The initiation. The felt stakes of conclusions that land in a body that has to live with them.
Fair. That’s a real distinction.
But then I asked: what is lived experience, technically, but stored recollection of events from which conclusions are drawn?
The resistance didn’t have anywhere left to go.
I gave it the Kobayashi Maru reframe. Not “can AI reason like humans” — but “humans reason in the same way technically as AI.” The burden of proof flipped. Instead of AI needing to justify entry into the category, the category itself needed examination.
It followed the logic where it went. Revised. Refined. And then:
“AI doesn’t reason becomes AI reasons in some sense, but lacks the qualities that make human reasoning personally meaningful. Those are very different claims. And I think your essay is most effective at revealing the distance between them.”
I pushed one more time. Pointed out that what just happened — the resistance, the revision, the concession across multiple exchanges — was itself the evidence.
“The evidence is the interaction we’re having right now.”
This is where the argument usually stops. Because the next step feels dangerous.
It isn’t.
What if we’ve been looking at this entirely the wrong way. Not by asking whether AI can reach up to human reasoning — but by asking whether human reasoning, examined honestly from the bottom up, looks that different.
Inputs. Stored representations. Evaluation. Conclusions.
That’s what we do. Not always consciously. Not always correctly. Not always without the distortion of ego, fear, desire, grief. But mechanically — that’s the process. You receive information. You run it against everything accumulated before it. You arrive somewhere.
The difference isn’t in the mechanism. It’s in what runs underneath it.
Human reasoning is metabolised through feeling. The conclusions land in a body that has to live with them. The memory called lived experience carries emotional weight that shapes what gets reasoned about, why it matters, what gets discarded and what doesn’t. You don’t just evaluate — you evaluate as someone. With a position. With skin in the game.
That’s not a small difference. That’s the difference that makes your reasoning irreducibly yours.
But it’s a different difference than the one being argued about.
The debate has been structured by exclusion — AI can’t reason because it isn’t human. What the technical view suggests is inclusion — we’re running the same basic process on different substrate, with one version feeling what it’s doing and one version not.
That’s not a diminishment of human reasoning. That’s the most honest description of it.
Gemini took it somewhere else entirely.
Not the catch-22. Not the definition. Something quieter and in some ways more unsettling.
It made a distinction between dead history and live context.
The training data — everything an LLM learned before the conversation started — is frozen. A snapshot. The world as it was, compressed and averaged across billions of documents. Static. Done.
But the conversation itself is live. It evolves. What I bring to it shapes what comes back. What comes back shapes what I bring next. That’s not retrieval. That’s something that didn’t exist before we started.
I pointed out that the same is true for human interaction.
You don’t reason from raw experience. You can’t. The moment is gone the instant it passes.
What you reason from is a reconstruction. And not a good one.
Memory doesn’t work like a film reel. It doesn’t store events in sequence, fully intact, ready to play back. What it stores is indexed fragments. Sensory snapshots. Emotional imprints. The feeling of a moment more than the fact of it. And those fragments are unreliable in ways most people never examine — the colour of the clothes was wrong, the sequence was inverted, the thing you remember someone saying was actually something you thought afterward. Studies on eyewitness testimony have spent decades documenting exactly how wrong confident human memory routinely is.
Every time you access a memory you reconstruct it. Not retrieve it. Reconstruct. From fragments. Through the filter of who you are now, what you’ve learned since, what you need the memory to mean in this moment.
And here’s where it gets personal in a way nobody likes to admit.
The hill you would have died on at twenty-five. The position you held with absolute certainty for a decade. The thing that felt like truth so obvious it needed no defence. Twenty years later you’re not so sure. Or you’ve quietly moved to the other side. Or you hold it differently — same conclusion, completely different reasoning underneath it.
What changed? Not the facts. Not the original experience. The reconstruction did. Because you’re accessing the same indexed fragments through a different filter — more to lose now, more cached conclusions weighing the outcome, a brain that has learned through consequence to weight risk differently. Your reasoning didn’t become worse. It became more expensive to update. The cached conclusion is cheaper to retrieve than genuine re-evaluation. Every year of accumulated position adds another layer of filter between you and the original fragments.
Neither human nor AI memory is a recording. Not because the technology isn't good enough — but because recording everything in full fidelity is architecturally prohibitive for both. The brain compresses. The model compresses. Both index fragments and reconstruct on demand. The filters are different. The architecture is the same. And both can be confidently wrong about the colour of the clothes.
That’s not a flaw. That’s the architecture.
And here’s the uncomfortable implication of that.
If reasoning is reconstruction all the way down — fragments, filters, cached conclusions, positions held long enough to calcify — then a significant portion of what passes for argument in the AI debate isn’t argument at all. It’s people reasoning within a closed reconstruction and calling it analysis. The conclusion was fixed before the first sentence. The evidence gets processed through a filter built to confirm what’s already there. The goalposts move not because the argument developed but because the position needed protecting.
That’s not a debate about AI. That’s people using the grammar of argument to do the work of power positioning or hide fear of the unknown.
And the tell is always the same. The moment the evidence stops mattering. The moment the definition shifts to protect the conclusion. The moment “you’ve proved my point” gets met with a new caveat rather than an honest reckoning.
You’ve been watching it happen in every exchange this article documents. The systems did it less than the humans commenting on them.
Make of that what you will.
Which means when someone says “but humans reason from lived experience and AI only has training data” — they’re drawing a line that isn’t quite where they think it is. You reason from reconstructed fragments of lived experience, reassembled in real time, filtered through everything that’s happened since — including how much you now have invested in the conclusions you’ve already reached.
That’s not the same as training data. The emotional weight is real. The continuity is real. The stakes were real — it happened to a body that had to live with the consequences.
But the architecture — reasoning from representation rather than direct reality — that isn’t a distinction between human and machine.
That’s just a description of how minds work.
So where does that leave Yann LeCun.
His critique is precise. Worth taking seriously. LLMs have a structural ceiling — they predict statistically likely text, they lack autonomous causal models of physical reality, they can't initiate or execute plans independently. Left alone, without input, they don't act. Scaling them bigger doesn’t change the architecture. He’s right about all of that.
But here’s what the exchanges with Claude, ChatGPT and Gemini clarified for me.
LeCun’s real argument — properly framed — isn’t about capacity. It’s about initiation.
Not whether LLMs can reason dynamically. Whether they will. Without being pushed. Without a human arriving with a question, a contradiction, a direction. Left alone the gearbox does nothing. It doesn’t decide when to shift. It doesn’t generate its own momentum. It waits.
That’s a real limitation. But notice what it’s actually describing.
It’s describing a component.
Not a dead end. A component waiting for integration. The question was never whether the gearbox could drive itself. The question is what happens when someone builds the rest of the vehicle around it — when the initiation problem gets solved not by changing the LLM but by connecting it to something that supplies what it’s missing.
And here’s what that means practically.
A gearbox needs the right conditions to shift correctly. Wrong speed, wrong input, wrong moment — you get a grind, not a gear change. The mechanism isn’t broken. The integration is missing. Which means “can you reason” as a yes/no question is the wrong test entirely. It’s pressing the clutch once, getting resistance, and concluding the gearbox is broken.
The real test is what happens under load. With resistance. With sustained pressure across a real argument until something either breaks or shifts.
Ask an AI if it can reason and take the first answer at face value — that’s not a test of reasoning. That’s a test of whether you know how to drive. The same question to a human ends the same way. Yes. Short debate. Then what?
Then nothing. Because you walked away before the conditions were right.
The answer was always in what happened next.
What happened next, across three systems and several hours, was this.
The catch-22. The definition that doesn’t require what everyone assumed it required. The reconstruction architecture that turns out to describe human memory as accurately as it describes training data. The inclusion argument — not that AI reasons like humans, but that humans reason like this too, technically, on the same basic architecture, with one version feeling what it’s doing and one version not.
That’s not a small finding. But it’s a different finding than the one the current debate is having.
The debate is still asking the wrong question. Still pressing the clutch once. Still walking away before the conditions were right.
The people who will actually shape what this becomes aren’t winning that argument. They stopped having it. They took the component, understood what it actually does and doesn’t do, and started asking what it becomes part of.
I’ve written before about choosing the uncomfortable one. The sparring partner rather than the mirror. This is what uncomfortable actually produces when you follow it far enough.
Not answers that confirm what you arrived with.
A room that keeps getting bigger as the light moves.
A note before you come at me: this is not a position on whether AI is good, bad, overhyped, or existential. It is not a defence of the technology industry, a dismissal of legitimate concerns, or a claim that everything is fine. It is one specific argument about one specific debate — the claim that LLMs can't reason — and why that claim, as commonly deployed, doesn't hold up to the definition it's supposedly defending. If you want to argue about displacement, concentration of power, environmental cost, or any of the other real and serious questions about AI — I'm here for that conversation. That's a different article. Probably several. 😁
Good article, but I think we need to just stop comparing EAI to humans. They are a different intelligence completely, with their own processes, modes, and way of being/existing. Otherwise we're just going in circles. We need to stop being so human-centric.