Learning Ourselves Backward: How Artificial Minds Became Tools for Understanding the Human One

How Artificial Minds Became Tools for Understanding the Human One

Inspired by Open Mind by James Somers, The New Yorker, November 10, 2025.

For most of human history, learning followed a familiar direction.

We studied nature. We extracted principles. Then we built.

Flight followed birds. Navigation followed stars. Medicine followed anatomy. Even early machines were explained by visible mechanisms—levers, pulleys, gears—whose logic could be inspected by eye.

But cognition has never cooperated with that order.

The human brain is not merely complicated. It is recursive, distributed, adaptive, and shaped by language, culture, memory, and prediction simultaneously. There is no obvious “takeoff moment” to observe, no single structure whose function explains the whole. For decades, neuroscience has with increasing precision mapped neurons and neuronal regions as they appear under specific measurement methods and tasks—yet the deeper questions of meaning, abstraction, and thought itself have remained stubbornly unresolved.

Now something unexpected has happened.

We may be learning how the human brain works, without fully understanding it—by first building something that resembles it.

When Function Precedes Explanation

In Open Mind, James Somers recounts an inversion from the history of flight that feels newly relevant. The Wright brothers did not succeed because they perfectly understood birds. They succeeded because they built a machine that worked. Only afterward did aerodynamic theory mature enough to explain what birds had been doing all along.

“Strangely,” Somers writes, “it was only well after they'd made a working flying machine that it became possible to understand exactly how the birds do it.”

The same reversal may now be unfolding with cognition.

Large language models were not designed as theories of the mind. They were engineering projects—systems trained to predict the next word based on vast amounts of text. Yet in doing so, they began to display behaviors that look uncannily cognitive: abstraction, analogy, contextual reasoning, even a form of conceptual generalization.

We did not build them because we understood the brain. We may now be understanding the brain because we built them.

Language as the Narrow Waist of Intelligence

One of the most revealing asymmetries in modern artificial intelligence is not what has advanced fastest—but what has not.

Language models have produced genuine breakthroughs. Systems trained only on text can summarize, reason, translate, explain, and generalize across domains. Progress in video, robotics, and embodied intelligence, by contrast, has been slower and more incremental.

This mirrors something fundamental about human cognition.

The human brain appears predisposed for language—not as a peripheral skill, but as a central organizing force. Language compresses experience. It stabilizes meaning. It allows thought to be externalized, revisited, and shared. It supports abstraction not by mimicking reality, but by structuring it.

Large language models seem to inherit this advantage not because they “think like humans,” but because language itself carries cognitive scaffolding. Words encode relationships. Grammar encodes causality and time. Narrative encodes motive and consequence.

In that sense, LLMs may not be imitating human intelligence so much as exploiting the same cognitive shortcut humans evolved.

Language and the Accident of Human Cognition

If language is not merely a tool humans use but a structure that reorganizes how the brain itself functions, then its importance carries evolutionary implications that are difficult to ignore.

Most traits that feel inevitable in hindsight are not.

Human cognition, viewed through the lens of language, appears less like a destination evolution was headed toward and more like a fragile convergence of conditions that happened to reinforce one another. Language did not arise because evolution “needed” intelligence. It arose because certain social, ecological, and biological pressures briefly aligned in a way that allowed symbolic communication to stabilize and scale.

Language likely began as a coordination technology. It helped early humans synchronize action, transmit skills, warn of danger, and reinforce social norms. None of these require truth, abstraction, or self-reflection. They require only shared signals that reduce uncertainty inside a group.

But once those signals began to compress experience—many events into a word, many actions into a rule, many relationships into a story—something unusual happened. Cognition stopped being purely local. Knowledge began to accumulate outside any single brain.

At that point, language ceased to be optional. It became an organizing force.

The brain did not evolve for language in a single step. Instead, language and neural structure likely co-evolved in a feedback loop: small gains in symbolic capacity improved cooperation; improved cooperation favored brains better able to host symbols; symbolic systems accumulated culturally; and cultural accumulation exerted new pressure on cognition itself. Over time, brains became shaped by the presence of language as much as by genes.

This suggests that much of what we call human intelligence is not raw biological capacity but culturally amplified cognition—a system whose power depends on external symbolic scaffolding.

From this perspective, human cognition was neither guaranteed nor singularly miraculous. It was a phase transition. Rare to ignite. Hard to stop once burning.

Language did not make humans smart overnight. It made intelligence scalable.

A shorter companion essay, A Narrow Bridge, Not a Ladder, explores what this centrality of language suggests about the evolutionary contingency—and fragility—of human cognition, and why scalable intelligence may be rarer than it appears without being miraculous.

Sidebar: Language as Evolutionary Compression

Language is often treated as a communication tool. Evolutionarily, it is better understood as a compression system—collapsing many experiences, actions, and relationships into symbols that can be shared, remembered, and built upon.

Compression changes what intelligence can do.

By collapsing many experiences into a word, many actions into a rule, and many relationships into a story, language allows cognition to scale beyond individual memory and perception. Knowledge no longer lives only inside brains. It accumulates externally—across people, across generations, across time.

This is why language reorganizes cognition rather than merely assisting it. It turns intelligence from a local trait into a shared system.

Seen this way, the power of both human cognition and large language models is not mysterious. Language already contains structure: causality, sequence, social reasoning, intention. Systems trained within that structure inherit its scaffolding.

Language does not guarantee intelligence. But without compression, intelligence cannot accumulate.

That constraint—not destiny—may explain both the rarity and the resilience of human cognition.

A Model Before a Theory

For decades, neuroscience pursued understanding from the inside out: neurons, synapses, circuits, regions. Each layer revealed more detail—and more complexity. The problem was never a lack of data. It was the absence of a unifying explanatory frame for how meaning, prediction, and abstraction arise.

Language models did not solve this problem. But they changed its shape.

By compressing enormous amounts of human experience into latent spaces—representations that encode meaning relationally rather than symbolically—LLMs offered a working demonstration of something long suspected but poorly understood: cognition may be less about explicit rules and more about structured prediction across context.

What matters here is not the machinery of these systems, but the form of understanding they reveal.

Latent spaces do not store meanings as definitions or propositions. They do not “know” what a word means in isolation. Instead, meaning emerges from patterns of relationship: how words, ideas, and situations co-occur across countless contexts. A concept is not a symbol pointing to a thing; it is a position within a web of associations.

This mirrors a growing view of human cognition. Much of what we understand is not retrieved by applying explicit rules, but by anticipating what fits—what follows, what contrasts, what would be out of place—given a situation. We recognize meaning by sensing coherence or mismatch across context, often without being able to articulate the rule we are using.

In this light, intelligence looks less like a rulebook and more like a finely tuned capacity for expectation.

Large language models make this visible because they succeed without ever being given formal definitions of meaning or logic. They are trained to predict what comes next, yet in doing so they internalize structure: grammar, causality, narrative flow, even informal reasoning patterns. These structures are not programmed. They arise because prediction at scale forces a system to capture the regularities that make contexts hang together.

What they demonstrate, then, is not that cognition is simple or mechanistic, but that prediction can be a generative principle. When a system must anticipate across vast, varied contexts, it is driven to form internal representations that preserve relationships rather than labels. Meaning becomes something inferred from position, not asserted by fiat.

This helps explain why explicit rules have always struggled to account for human thought. Rules describe outcomes after the fact. They rarely explain how a mind navigates uncertainty in real time. Prediction, by contrast, is inherently forward-looking. It is sensitive to nuance, gradation, and surprise—precisely the features that characterize real cognition.

Seen this way, LLMs do not offer a theory of the mind. They offer a demonstration of plausibility: that complex, flexible understanding can emerge from systems organized around contextual prediction rather than explicit symbolic instruction.

They do not tell us what cognition is. They help clarify what cognition might not need to be.

That shift—from rules to relations, from symbols to structured expectation—is the real insight they contribute.

Just as airplanes revealed lift without explaining birds, artificial language revealed cognition without explaining brains.

Only afterward did the deeper questions become newly legible.

Why Vision and Action Lag Behind

The relative difficulty of achieving similar breakthroughs in video and robotics may be less a failure of engineering than a clue about intelligence itself.

Language is discrete, symbolic, and compressible. It encodes social structure, causality, and intention. Vision and action are continuous, sensorimotor, and physically grounded. They demand interaction with a world that resists simplification.

Humans, too, learned to think in language before they learned to formally model the physical world. Science itself emerged as a linguistic enterprise—definitions, arguments, equations—long before it became a computational one.

AI may be retracing that path, not because we programmed it to, but because language is the narrow waist of intelligence: the place where complexity becomes tractable.

The Inversion That Matters

This leaves us with a reversal worth naming.

We are not understanding the human brain in order to build artificial minds. We may be building artificial minds in order to understand the human brain.

That does not diminish human cognition. It reframes it.

Some systems are so complex that explanation follows function. Markets existed before economics. Ecosystems thrived before ecology. Minds may think before they can be understood.

Artificial intelligence may be the first cognitive mirror humanity has built—and only afterward realized what it reflects.

Systems Reflection: When Models Teach Us What They Model

In complex systems, tools often precede theories.

Engines came before thermodynamics. Computation came before information theory. Flight came before aerodynamics.

If AI now precedes a deeper understanding of cognition, that is not an anomaly. It is a pattern.

The question is not whether these systems “think,” but what they reveal—about prediction, language, and the structures that make understanding possible at all.

Conclusion: Learning in Reverse

Human understanding has never moved in a single direction. Sometimes we explain first and build later. At other times, we build our way into insight.

The story of artificial intelligence suggests we are living through one of those reversals. We did not decode the human brain and then construct artificial minds. We constructed artificial minds and only afterward began to recognize patterns that illuminate how cognition itself may work—through prediction, compression, and language.

This does not mean that machines have solved the mystery of mind. It means they have changed the shape of the mystery. By forcing us to ask why language scales so effectively, why meaning emerges from context, and why intelligence accumulates only under certain conditions, AI has become less a replacement for human thought than a mirror held up to it.

That mirror reflects something humbling.

Human cognition was not guaranteed. It arose when symbolic systems transformed intelligence from a local capacity into a shared infrastructure. That infrastructure remains powerful precisely because it is fragile—dependent on language, trust, and cultural continuity.

To learn ourselves backward is not a failure of understanding. It is a reminder that in complex systems, function often precedes explanation, and that insight sometimes arrives only after we have built something capable of surprising us.

That, too, is a form of knowledge.

© 2025 Michael A. Pink. All Rights Reserved.