The story we tell about artificial intelligence has always been fundamentally anthropocentric. From HAL 9000 to GPT-5, our mental model of what a "thinking machine" looks like has been shaped almost entirely by the assumption that intelligence, when it arrives in silicon, will basically look like ours - reasoning through problems step by step, holding coherent beliefs, wanting things, perhaps eventually surpassing us at every cognitive task we value.

That story has served a purpose. It gave researchers a target. It gave science fiction its dramatic arc. And it made AI legible to the public in a way that the actual mathematics rarely is. But as AI systems have become genuinely capable - as they've started doing things we can't easily explain - the humanoid frame has started to distort as much as it illuminates.

The mirror problem

When we evaluate large language models, we tend to reach for cognitive tests developed for humans: reasoning benchmarks, theory-of-mind tasks, reading comprehension. This makes some sense - we're trying to compare these systems to something we understand. But it also guarantees a particular kind of blindness.

A language model doesn't "know" things the way you do. It doesn't retrieve memories from a discrete storage location, or reason by constructing a mental model it then inspects. It generates plausible next tokens based on patterns absorbed across billions of documents. This process can produce outputs that look like reasoning, recall, and even insight. But the underlying substrate is profoundly different from the cognitive architecture we're implicitly using as our reference point.

The philosopher Daniel Dennett spent much of his career arguing that human cognition is less like a central "Cartesian theatre" and more like a loose confederation of competing processes - consciousness as an emergent narrative rather than a true seat of experience. If he's right, the gap between human and artificial cognition may be smaller in kind than we assume. But it's still enormous in character.

"We keep asking whether AI understands. We might do better asking what kind of thing it does instead."

Where the divergence shows up

The places where AI diverges from human thinking are, if you look at them carefully, more philosophically interesting than the places where it converges.

Consider context windows. A language model processes a conversation with perfect fidelity to what's in its window, with no degradation over token distance (within the window), and then loses it entirely when the context ends. Human memory is almost the inverse: degrading gradually, reconstructed and distorted with each retrieval, but extending across a lifetime. Neither is strictly better for all purposes. But they're not the same kind of thing at all.

Or consider consistency. A human mind builds up a relatively stable set of beliefs and values over time - not perfectly consistent, not unchanging, but something you can track across years. Current AI systems are stateless between conversations. The "Claude" or "GPT" you talk to today shares weights with the one you talked to last week, but has no episodic memory of that exchange. There's no accumulating self in the way we usually mean the term.

And then there's the question of what language models actually do when they generate. The process is fundamentally about fitting patterns - very sophisticated, very high-dimensional pattern-fitting, but pattern-fitting nonetheless. When a human writes an essay, they're drawing on something that feels like experience, intention, and judgment. Whether language models have anything analogous to those internal states is genuinely unclear, and may remain unclear for a long time.

Why the divergence might be the point

Here's the provocation: we may be systematically undervaluing the ways AI is different from us.

An AI system that doesn't get tired, that doesn't carry emotional residue from a bad morning into an afternoon consultation, that can hold a customer's entire history equally clearly whether the conversation is the first or the thousandth - these aren't consolation prizes for not being human. They're different capabilities with real value.

Similarly, the fact that a model can simultaneously "consider" thousands of ways a sentence might continue - that it doesn't commit to a single interpretive path the way a human reader typically does - might be a feature rather than a bug in certain applications. The model's relationship to ambiguity is genuinely different from ours, and sometimes that difference is exactly what you want.

The error is in the frame: treating AI as a degraded or aspirational human, rather than as a different kind of cognitive entity with its own strengths, limitations, and failure modes. The degraded-human frame leads us to evaluate AI by how closely it mimics us. The different-entity frame leads us to ask what it's actually good for, and what new things become possible when you combine it with humans who work differently.

The design implications

None of this is just philosophical navel-gazing. It has real consequences for how we build AI systems, how we deploy them, and how we regulate them.

If AI systems don't think like us, then designing them as if they're just faster humans is a mistake. The workflows that make sense for human experts - extensive deliberation, consulting colleagues, building intuition through years of feedback - may need to be completely reconceived when the "expert" is a model. The failure modes are different. The trust calibration should be different. The oversight mechanisms should be different.

The same applies to how we think about AI risk. The models that animate popular discourse about dangerous AI - the Terminator, the Paperclip Maximizer - are all essentially anthropomorphic. They posit an AI that develops goals, plans to achieve them, deceives humans to protect its plans. This may be a relevant risk space in the long run. But it's not the risk space that matters most right now, and over-focus on it may distract from more immediate, more tractable problems: bias in outputs, brittleness in distribution shift, failure to flag uncertainty, misuse by humans who have clear goals of their own.

Understanding AI as a genuinely different kind of cognitive entity doesn't mean dismissing either its capabilities or its risks. It means being precise about which capabilities and which risks - rather than projecting the human template and hoping it fits.

A more honest starting point

The history of science is full of moments where the right move was to stop forcing a new phenomenon into an old conceptual frame and build a new one. Electricity wasn't quite like fluids, even though "current" and "flow" were useful early metaphors. Quantum mechanics wasn't quite like classical mechanics, even though the equations reduced correctly at the right limits.

AI may be at such a moment. The human-mind frame has been enormously productive. It's not wrong so much as it's incomplete - and increasingly, the incompleteness is where the interesting questions live.

What would it look like to take AI seriously on its own terms? To ask not "does this model understand?" but "what does this model do, and under what conditions does that become valuable or dangerous?" To design human-AI systems around the genuine complementarity of two very different kinds of cognitive process, rather than the aspiration for one to eventually replace the other?

I don't think we're close to having clean answers. But I think asking better questions is the necessary first step. And the first step toward better questions is letting go of the assumption that we already know what shape the answers should take.

If you're new to thinking about AI, the AI for Beginners path covers how AI actually works - a better starting point than science fiction or corporate marketing. And for the practical skill of evaluating what AI tells you on its own terms, How to critically assess AI answers is the companion piece to this one.