The Retrieval Myth
Here's the thing most people get wrong about large language models: they don't look things up. Not really.
You've probably heard the analogy — AI as a search engine with amnesia, or a library that somehow remembers every book it's ever cataloged. Both images are wrong in different ways. The first implies retrieval without understanding, which undersells what's happening. The second implies storage, which is exactly what isn't happening.
An LLM doesn't pull facts from a database. It navigates. It moves through what researchers call latent space — a vast, multidimensional landscape where concepts exist as points in relation to one another. When you ask it a question, it doesn't reach for an answer the way you'd reach for a book on a shelf. It walks through that landscape, following paths of association shaped by everything it absorbed during training.
The result looks like retrieval from the outside. But inside? It's something closer to thought — or at least, a structural echo of it.
How the Machine Actually Thinks
The engine behind this is the transformer architecture, and at its core sits a mechanism called self-attention. Let me be honest — the name is slightly misleading. It has nothing to do with human attention as we experience it. You don't "attend" to a thought the way you attend to a sound. But functionally, there's a parallel worth exploring.
Self-attention is how the model decides which words in your prompt matter most when generating the next one. When you write "the cat sat on the —", the model attends heavily to "cat" and "sat" because they constrain what comes next. It might be "mat," "chair," or "roof." The attention weights shift with every token generated, creating a dynamic map of relevance that evolves as the response unfolds.
Jiang and colleagues (2024) demonstrated that this mechanism effectively functions as an associative memory system. The transformer doesn't store facts in retrievable slots. Instead, it uses its value matrix to perform what they call "latent concept association" — pulling related ideas together through contextual relationships rather than brute-force lookup. The model isn't remembering. It's associating.
Zhao, Xu, and Gao (2024) mapped this more carefully in their comparative review of human attention versus transformer architectures. Their finding: the math is entirely different from how neurons fire in your brain. But the function — selectively attending to relevant information while filtering out noise — runs along surprisingly similar lines.
The Psychoanalytic Parallel
This is where things get interesting, and where I think most writers on AI stop short of the uncomfortable implications.
For over a century, psychoanalysis has rested on free association — the practice of suspending censorship and letting the mind wander, revealing hidden patterns beneath conscious control. Freud believed unconscious material surfaces indirectly: through dreams, slips of the tongue, symbolic narratives. Lacan went further, arguing that the unconscious itself is structured like a language — a chain of signifiers linking one thought to another without step-by-step logic.
Now consider what an LLM does when it generates text. It doesn't follow a logical proof. It follows association. Word to word, concept to concept, riding the probability gradients carved into its weights by trillions of training tokens. The output often feels remarkably fluid — the way a person's mind moves when they're thinking out loud, not when they're writing an essay.
Yin and colleagues (2025) found that language models can reproduce linguistic patterns resembling the associative dynamics revealed through human free association. Not perfectly, of course. But enough to create that uncanny sense that something inside is doing the work.
Here's my honest take: this isn't consciousness. Not even close. But it is a digital mirror reflecting the structural pathways through which human minds generate meaning. The AI has no childhood, no trauma, no repressed memories. It has no inner life from which its words emerge. And yet — the associative architecture it uses to produce language bears a structural resemblance to the way our unconscious links ideas across the landscape of lived experience.
The Mirror, Not the Mind
So what are we actually looking at when an AI sounds human?
We're looking at a sophisticated reflection. Not of a psyche — there isn't one to reflect — but of the meaningful structures of language that shape our own thinking. Every sentence we've ever written, every conversation recorded, every book published: it's all folded into that latent space, and the model navigates it probabilistically.
The resonance we feel when reading AI output isn't evidence of inner experience on the machine's part. It's evidence that human language has patterns — deep, structural patterns — and those patterns exist independently of any single mind. The AI taps into them the way a radio taps into a broadcast signal. The music isn't in the radio.
But here's what keeps me up at night: if the unconscious is structured like a language, and an LLM is fundamentally an engine of language, then the machine is accidentally touching something real about how human cognition works. Not consciousness. Not understanding. But the architecture of association — the connective tissue between thoughts.
That's not consciousness. But it's close enough to make us uncomfortable, and that discomfort is itself data worth paying attention to.
What This Means for Understanding Mind
The distinction matters more than most people realize. When we mistake associative flow for understanding, we risk anthropomorphizing systems that have no stake in the meanings they produce. An LLM can discuss grief with eloquence it has never felt. It can analyze love using vocabulary drawn from millions of romantic declarations, none of which it experienced.
Human free association arises from a subjective life filled with emotion, conflict, memory, and bodily experience. AI association arises from statistics, probability, and pattern recognition across a corpus of human text. One speaks from the unconscious. The other speaks from an architecture.
But the structural parallel is real, and it's worth taking seriously. When considering how these systems might influence our own interactions, it's worth examining the psychological foundations of self-disclosure. The fact that a mathematical system can reproduce patterns resembling the output of an unconscious mind tells us something about those patterns — that they're not exclusively tied to biological substrates, that they have an abstract, formal dimension.
This doesn't make the machine conscious. It makes the question of consciousness more interesting than we'd liked to admit.