Minds Imitating Machines: What AI Reveals About Human Intelligence
In the wake of recent AI breakthroughs — namely, LLMs and reasoning models — I can't help but ponder an old adage: "At first, art imitates life. Then life will imitate art. Then life will find its very existence from the arts." Its creator — Fyodor Dostoevsky, a Russian novelist from the nineteenth century — probably didn't anticipate just how relevant this insight would be two and a quarter centuries later, but it's surprisingly applicable in the context of artificial intelligence.
What Dostoevsky was getting at was essentially the interplay of the natural and the artificial. At first, humans draw inspiration from the natural — elements that come from the environment, typically unaltered by human intervention, ie. life — to invent the artificial — elements that are man-made, often involving synthetic processes, ie. art. Art's power lies in its capacity to imitate life, thereby inducing reflection within its recipients. More interestingly, and less intuitively, the natural begins to resemble the artificial. We find ourselves gazing through a sea of fog, admiring its beauty and wonder because, to quote Oscar Wilde, "poets and painters have taught the loveliness of such effects." Our perception of things we see in the world become altered by art and what it teaches us to see. Art interprets and describes life; over time, our very experience of life becomes the product of art's interpretations and descriptions. Art (re)invents what we see in the very act of describing it.
In AI, there is a surprising parallel. In 1956, when John McCarthy coined "Artificial Intelligence," the intent was to replicate human/natural intelligence — which for the remainder of this essay, will be referred to as biological intelligence. Foundational figures in AI, such as Alan Turing, Geoffrey Hinton, Yann LeCun unite in the shared effort to improve AI systems, ultimately allowing AI to imitate human cognitive abilities with growing accuracy. While AI is by no means a perfect replica of biological intelligence at this point in time, the pursuit of imitation has historically led to breakthroughs in Natural Language Processing (aka. NLP, which underlies models used to drive today's chatbots), Computer Vision, and Autonomous systems.
Had we failed to draw inspiration from the natural world in order to construct artifacts — tools and technology — we wouldn't have airplanes (inspired by birds), power harvesting tech (turbines were inspired by whale fins, and solar panels inspired by sunflowers), medical syringes (inspired by bee stings), neural networks (inspired by the human brain), among other inventions. If Dostoevsky was right about life imitating art, could biological intelligence (the natural) imitate artificial intelligence? If biological intelligence were to mimic the very systems it inspired, what does that say about the nature of cognition itself? Would it suggest that intelligence is more mechanistic, predictable, and reproducible than we typically assume — or would it instead expose the limits of artificial constructs in ever truly mirroring the natural? This uncertainty lies at the heart of understanding not only how we build intelligent systems, but also how those systems might, in turn, reshape how we understand ourselves.
Functional Isomorphism
This uncertainty — whether biological intelligence could be predicted or mirrored by artificial constructs — brings us to a key philosophical framework for thinking about the relationship between minds and machines. In his essay "Can an AI System Think? Functionalism and the Nature of Mentality," American philosopher Nino B. Cocchiarella builds on a theory called functional isomorphism first introduced by Hilary Putnam. Functional isomorphism rests on two ideas: functionalism and isomorphism. Functionalism, a philosophical — not scientific — theory, holds that "the essence of thought and mentality is its functionality" (pp. 4). Rather than focusing solely on outward behavior, as behaviorism — the dominant framework in mid-20th century psychology — did, functionalists aim to model the internal, invisible processes that drive cognition. They view the mind as a black box: a system whose internal operations, like software governed by logic and algorithms, are essential to understanding thought itself. In doing so, functionalism marks a departure from behaviorism's insistence that only observable outputs matter.
Isomorphism, meaning "having the same form," relates two systems with identical structures: a representing system and a represented system. The representing system maps the structural relations of the represented one, allowing valid inferences about it. For example, mathematics functions as a representing system isomorphic to natural laws like gravity, mass, and distance; algebraic solutions predict real-world phenomena because of this structural isomorphism.
Hence, two systems are functionally isomorphic to the extent that the representing system successfully captures the internal mechanisms (the algorithms) of the represented system. Cocchiarella posits that AI can be functionally isomorphic to biological intelligence. If this is true, "an AI system [could] be structurally duplicating, and not merely simulating, the mental states and processes that humans have." (pp. 1-2) The outcome is that "an AI system can think and be self-conscious… in just the way that humans can." (pp. 2) From a functionalist standpoint, where the essence of mentality lies in how a system operates rather than in what it is made of, this would mean that an AI replicating the functional architecture of the human mind would not merely mimic human intelligence — it would, in functional terms, be human intelligence.
The Ecology of Intelligence
Of course, this is a radical notion, and many philosophers disagree with the fundamental assumptions of functionalism — the essence of human cognition may not be its functionality. In her book "The Atlas of AI," Kate Crawford, a researcher studying the social and political implications of AI at Microsoft Research and NYU, draws our attention to a different framework, highlighting "the fundamental ways in which humans are embodied, relational, and set within wider ecologies." (pp. 5) Crawford suggests that the essence of intelligence is its ecology — the social, historical, economic, political, and cultural environment that produced it. Intelligence is not an autonomous entity — it is indistinguishable from the wider context in which it is cultivated. Her critique that "this belief that the mind is like a computer, and vice versa, has 'infected decades of thinking in the computer and cognitive sciences,'" (pp. 7) underscores the stakes of functionalist assumptions. By reducing intelligence to computation, one risks overlooking the embodied, situated, and historically contingent dimensions of what it means to be human.
Crawford argues that AI is not some abstract software floating in a digital void. It is deeply embodied, but in a very different way than biological intelligence. AI's embodiment is not biological or sensorimotor — it is industrial: built through labor, energy, rare-earth extraction, logistical infrastructures, and capitalist systems. Because AI departs from this ecological framework of intelligence, it is not definitionally "intelligent." To hold on to the perspective that AI is intelligent ignores how systems of power shape what it is, what it does, and whom it serves. After all, AI is not autonomous, as much as it appears to be.
Crawford illustrates this point with the example of Clever Hans, a horse once deemed the "Smartest Horse in the World" in the nineteenth century. Clever Hans was nothing short of a miracle — he could solve math problems, tap out answers to addition, subtraction, multiplication, and division problems. Everyone at the time — including the New York Times — believed that Clever Hans was truly intelligent. When Pfungst, a psychologist, investigated Clever Hans' ostensibly extraordinary capabilities, he found that Clever Hans had simply learned to draw from contextual cues — the questioner's posture, breathing, facial expression — to answer math questions correctly. Indeed, he didn't actually understand the mathematical operations he would apparently perform. Crawford draws parallels between Clever Hans and AI — like Hans, AI is simply trained "to follow cues and emulate humanlike cognition." (pp. 4) Ultimately, AI is not intelligent; it is simply taught to imitate human behavior.
Marr's Three Levels
Crawford's framework complicates a functionalist's framework. While Cocchiarella adopts the paradigm that intelligence is functional — anything that is functionally isomorphic to human/biological intelligence is intelligent — Crawford insists intelligence is inherently social, historical, and ecological. A parallel distinction is that between the software and the hardware in computer science. The software of a computer is not tangible — it consists of all the algorithmic logics, instructions, and programs used to run machines. Its antithesis, the hardware, refers to a computer's external, tangible components — computer chips, motherboard, keyboard, etc. Functionalists focus on the software, approaching intelligence in terms of its algorithmic logics which can be physically instantiated in multiple ways (aka. Multiple realizability, coined by Putnam). Insofar as the algorithmic logics — the software — of AI are functionally isomorphic to biological intelligence, how they are physically realized, through brain tissues or computer chips, is irrelevant. Crawford's framework, conversely, focuses on the hardware, emphasizing the materiality of intelligence. By arguing that "AI is embodied and material," (pp. 8) she contrasts the hardware of biological intelligence — ecological — to that of AI — industrial. Hence, AI is not one and the same as biological intelligence because of how they are physically realized. At first glance, these two frameworks are irreconcilable. Lucky, upon closer look, Crawford actually builds on Cocchiarella's functional isomorphism.
Notably, Cocchiarella and Crawford operate on two different levels. The former is focused on something intangible, while the latter, tangible. David Marr, one of the originators of computational neuroscience, formalized such levels in his Three Levels of Analysis for understanding complex information processing systems, such as the brain and a computer. Marr's three levels consist of a computational, algorithmic, and implementational level, in this descending order. For any information-processing system, the computational level specifies the problem it faces in a generic manner — a goal-oriented task it seeks to accomplish. The algorithmic level describes how the identified computational problems can be solved in terms of (intangible) instructions or programs. Finally, the implementational level is the physical mechanism by which the algorithm is instantiated. Cocchiarella focuses on the computational and algorithmic levels, treating intelligence as a software defined by its algorithmic logics. This software assumes a kind of universality — intelligence, at its core, is a set of algorithms and patterns that can be translated across systems. If a system replicates the causal-functional structure of the mind, it is intelligent, and this structure could be instantiated anywhere: a human brain, a computer, maybe even alien silicon. Fundamentally, functionalists are solely concerned with the computational and algorithmic levels, assuming intelligence is immaterial and substrate-independent.
Conversely, Crawford focuses on the third level — the implementational level — zeroing in on the materiality and embodiment of intelligence, whether artificial or biological. Once a universal algorithm is instantiated, in any context — whether geopolitical, economic, cultural — it becomes localized, conditioned, situated. AI realized and trained in the U.S. reflects Western ideologies, assumptions, and data biases. An AI trained in China, India, or Kenya would internalize different patterns of intelligence — because its training data, labor conditions, infrastructural design, and even its goals would differ. Intelligence, once realized at the implementational level, is never neutral. It is shaped by the ecology of its origin. And — so is biological intelligence. Chomsky's Universal Grammar — the idea that all humans are born with an innate, biological capacity for language, a shared cognitive structure that underlies all possible, localized human languages — illustrates the universality of human cognition at the algorithmic/computational levels, and its localization at the implementational level. Implementation localizes a system of intelligence, introducing it the values, biases, and goals of the cultures that produced it. A simplified analogy is a universal recipe (a software containing all the instructions and algorithms) that turns out differently depending on the kitchen, ingredients, and cook (the wider ecology).
Crawford's framework does not reject functional isomorphism — it deepens it. Any complex information-processing system has three levels — two of which are addressed by functionalists. Crawford addresses the third level, the missing piece of the puzzle. In the same way that biological intelligence possesses universal algorithms that are localized to its ecology, AIs also possess functionally isomorphic algorithms localized to the materials, labor, infrastructure, social ideologies that produced it. For an AI to truly duplicate biological intelligence, it must be not only algorithmically isomorphic, but also ecologically isomorphic, with the same localized values used in its training. This is, evidently, easier said than done: one must not only duplicate the structure of intelligence itself, but the wider ecology in which the desired system of intelligence exists. Only then does AI (the representing system) truly predict biological intelligence (the represented system). To return to the original question — "could AI be used as a system to interpret and describe biological intelligence in the same way that math is used for natural phenomena?"
Mechanistic Interpretability
At the intersection of AI and cognitive science is mechanistic interpretability (or mech interp), a subfield of AI research focused on answering the question: "How does a trained neural network actually work under the hood?" Mech interp reverse-engineers complex AI systems to understand their internal, "mental" processes. With mech interp, you can open up the "black box" and try to figure out what each neuron in the neural net is doing and why, so to speak. The question is, can mech interp be to human cognition what mathematics is to natural phenomena? That is, a formal, abstract system that is functionally isomorphic to the mind and allows us to understand — open up the "black box" of — cognition? If so, it could be used as a tool (a formalized system) to interpret, describe, and predict cognition!
Crawford complicates this process by adding that simply duplicating cognitive algorithms is not sufficient. For mech interp to be truly predictive, it must also duplicate the localized data inherent to the desired human reflecting their wider ecology. Just as every person is "trained" on a unique set of data, the respective AI must also receive the same set of training data. If and only if all these conditions are met — functional and ecological isomorphism — artificial intelligence can theoretically explain and predict biological intelligence. Once again, we would achieve another instance of using the artificial to justify/understand the natural.
Though Dostoevsky is a literal legend for noticing the interplay of art and life, his quote I'd argue is missing a fundamentally important question — what's the point of describing life through art? Why do artists and poets spend their whole lives trying to capture, interpret, and describe life? Similarly, why do mathematicians dedicate whole careers creating questions, inventing formulas to describe and explain nature? I'd argue that their efforts have a positive, social utility. Through art and literature, artists and poets create "cheat codes" or "shortcuts" that help us quickly identify the artfulness of life. By teaching us the beauty of fog, for instance, artists illuminate it — making it newly visible, heightening our attention to it. Our goal, then, is to find art in life, not just life in art.
In this light, the point of using art — or AI — to describe life is not merely to imitate it, but to strive to illuminate it. Just as poets reveal the beauty of fog, AI — when properly constructed — can help us see the structure of thought itself. It illuminates cognition, making visible the invisible patterns that shape our minds. But importantly, these structures reflect the context in which they were made. If we want AI to deepen our understanding of cognition, rather than distort or misrepresent it, we must ensure it mirrors not only our mental functions (which are generalizable) but also the ecological, cultural, and ethical realities that define us (localized). Only then does the artificial not just imitate the natural, but help us truly understand it.