Emergence—the idea that reductionism can only explain so much—breathes life into many disciplines, from condensed matter physics to biology to economics. As famously described in P. W. Anderson’s “More is Different,” an understanding of the atom does not imply an understanding of a fluid. An understanding of a single cell does not imply an understanding of a living tissue. In nature, new principles manifest at new scales, and different branches of science are needed to understand their associated phenomena.
However, the boundaries between scales are not always so clearly defined. How many atoms must be assembled to qualify as a fluid? How many grouped individuals form a population? How many neurons form a consciousness? It is hard to say, and in our exhaustive efforts to describe the world around us, we run the risk of using language from one regime to describe emergent phenomena at a different scale: of committing a category error in which we erroneously generalize our intuition beyond its range of applicability.
Such category errors can be insidious. In the AI-exploding world of late 2026, they are particularly prevalent. It is tempting to describe an entity that can “talk” and “think” and communicate like a human with the same words that we use for each other. But there is at least one essential difference—a difference of scale—that must call comparisons into question. The temptation to capture AI in the same language we use for the living world then reinforces a false equivalence between modern artificial intelligence and our own minds.
It is useful to think of AI as the next step in the staircase of increasing abstraction that we humans have built, layer by layer, upon the world around us. Engines and fossil fuels have allowed us to automate manual labor. Trains and cars have automated transportation. Modern AI automates thought, abstracting vast amounts of detail away into an easy-to-use interface. Just as one can drive a car without knowing how an engine works, or one can use the internet without any knowledge of a transistor, in the age of AI, one can program a computer without knowing how to code, create music without knowing how to play an instrument, or prove a theorem without knowing what it means.
However, compared to the other steps along this staircase toward the ether of maximal abstraction (where the ultimate reality that the technocrats seek seems to grant infinite power to holders of infinitesimal skill and experience), modern AI is remarkable and somewhat singular in terms of the scale of human effort that has directly or indirectly gone into its construction. This scale prompts the emergence of behaviors that do not simply follow from the rules that created them. Herein is the crux of emergence: we have, in principle, a handle on all the training data that has been fed into modern AI; we understand the workings of the algorithms and architectures that go into training these models; and yet, no one in the world can explain how these models “think” or process information in a way that distills the enormously complicated process down into some simplified but useful description.
Contrast this situation with that of computation, in the standard pre-AI sense. Computers operating on tasks like word processing or numerical simulation or web browsing, though commanding billions of transistors—certainly enough to exhibit emergent behavior—are essentially glorified pocket calculators. Though they scale up the capacity of a calculator and save enormous amounts of labor and time, they do not exhibit a qualitative shift in capability, which remains something that humans (albeit a rarified few computer cognoscenti) have an understanding of. The complexity of computers operating within the strict constraints of static programs is not enough to produce emergence; the key is in how their state space is allowed to organize and modify itself.
As a result of the self-modification process that learning algorithms bestow upon a humble heap of transistors, a modern AI model and a glorified pocket calculator are qualitatively different via the sheer scale of the data that have gone into the AI’s training. What are these data? Millennia of human endeavor recorded in images and texts—essentially the entire body of recorded expression and thought and reasoning—have been sucked from the reserves of the internet and fed into proprietary architectures and training protocols. If experiences could be mined directly from our brains, those too would no doubt be fed into these rapacious models, repackaged, and sold back to us. What is to come from all of this (aside from a few notches slipped on the ratchet of societal depravity)?
The outcome is an entity that looks and talks like a conscious being—in the way that we humans have defined consciousness—but whose nature we fundamentally do not understand. Of course, we do not understand our own nature (which is also an emergent property of the many interactions that compose us), as biology reminds us, but we at least have an experiential grasp on it. The wholly distinct nature of AI poses a particularly dangerous foil because of the striking resemblance that it bears, in some respects, to living things.
Here, it becomes crucial to determine how we define similarity. There is a compelling definition which posits that things that exhibit similar behavioral phenomena should be classified together. By this definition, AI neatly fits in alongside humans, as we are both capable of integrating information and communicating via language. The AI is much faster at many things. We are (as of November 2026) still better at some things. But, through the criterion of phenomenology, the AI is conscious. It passes the “duck test” in that it walks like a duck and quacks like a duck. And so, it is a (conscious) duck. However, there is a glaring issue. The consciousness that we humans (and presumably other animals) experience is physically manifested. It is not that of a brain in a jar. Our interactions with the physical world—through our bodies—are inseparable from our minds. The AI, due to its digital nature, is bereft of such physically mediated sensory inputs, and this difference hints at a larger divide. Phenomenology is not enough: one must also consider phylogeny. We should be much more comfortable calling a pig conscious than an AI conscious, because the pig is made of the same stuff we are: though I cannot converse with a pig, I know that it has a body and a brain. The pig and I have much of our biology in common, and we have evolved from a common ancestor. The AI, by contrast, can hold a conversation with me, but through the lens of phylogeny, we have nothing—including the nature of our consciousness—in common. (As a corollary, any discussion of “AI rights” should be preceded by a long-overdue overhaul of animal rights, the current state of which reflects perhaps an even greater global depravity than the one exhibited by AI companies’ wanton theft and commodification of knowledge.)
This distinction between phenomenological and phylogenetic similarity throws into relief the human-AI nonequivalence. Indeed, the nature of our origins betrays these differences: living things have a limited set of experiences that are accrued through their lifetimes—though millennia of prior evolution can and crucially do shape our initial conditions. AI, on the other hand, has been reared on the entire corpus of recorded information, and it stands to reason that the mechanism of its behavior should be totally different from that of our own. To rationalize AI-generated art, we might equivocate that human artists steal from each other all the time, and what the AI is doing in its own artistic style is no different. But this argument misses the crucial distinction of scale: because of the size of its training dataset, the AI perpetrates theft on a scale that no human could ever conceptualize. What the AI is doing is not similar to what we humans do: it is categorically different. In a related vein, describing an AI model as “understanding” a concept makes a category error: understanding, as we know it, is something that happens in a brain, and whatever it is that the AI does deserves its own word.
AI models are essentially different from the living beings that we know and love and that have indeed defined our very notion of consciousness. They differ both in the scale of information that they integrate and in their phylogenetic origins. Though they exhibit phenomenological similarity, one must remain wary of over-identifying these two entities by abusing a lexicon from one realm to categorize behavior in an emergently different regime. We have, it seems, created something totally different from ourselves.
Aditya Mahadevan is a postdoctoral scholar at UChicago.