Intelligence Is a Relationship, Not a Thing

Intelligence Is a Relationship, Not a Thing

ai blog
July 6, 2026

Joel Kowalewski, PhD

The word we changed in 1956

For seven years beginning in 1946, a singular interdisciplinary group convened in New York under the Josiah Macy Jr. Foundation: the mathematician Norbert Wiener, the neurophysiologist Warren McCulloch, the logician Walter Pitts, the information theorist Claude Shannon, the anthropologists Gregory Bateson and Margaret Mead, the psychiatrist W. Ross Ashby. Their project was to build a science of what Wiener, in his 1948 book, called Cybernetics: Or Control and Communication in the Animal and the Machine. The phrase in the subtitle carried the whole argument. What was interesting about a mind, an organism, or a machine was not a substance any of them contained but a pattern of relations each maintained with its surroundings — feedback, regulation, control. Intelligence, on this picture, was something a system did in relation to a world, not something it had.

Then, in the summer of 1956, a smaller group gathered at Dartmouth College under a different banner. John McCarthy had coined the term the previous year, in an August 1955 funding proposal written with Marvin Minsky, Nathaniel Rochester, and Shannon: “artificial intelligence.” The rebranding was not cosmetic. “Cybernetics” had located intelligence in a relationship; “artificial intelligence” relocated it inside the machine — a property the device would one day possess the way a battery possesses charge. We have been searching the machine for that charge ever since. This essay is about the error introduced by that relocation, and about how much of the present confusion over AI is the predictable consequence of a category we quietly changed seventy years ago.

A category mistake, diagnosed in 1949

The flaw had in fact been named before Dartmouth ever met. In The Concept of Mind (1949), the philosopher Gilbert Ryle described a recurring confusion he called a “category mistake.” A visitor is shown Oxford’s colleges, libraries, and laboratories and then asks to be shown “the University,” as though it were one more building rather than the organized relationship among the ones already seen. Ryle aimed the diagnosis at what he mockingly called “the ghost in the machine” — the Cartesian mind imagined as a thing concealed inside the body. The same blade cuts the question that now dominates public argument: is the system intelligent? To hunt for intelligence inside a virus, a crow, or a language model is to look for the University among the buildings.

What we refuse to call intelligence

Consider a virus that perfectly hijacks its host’s immune defenses. It explores a vast genetic sequence space and learns, in effect, a function that anticipates the host’s response; the host counters, the virus adapts, and the exchange plays out in real time. By any operational standard that is intelligence — an evolved, optimized calculation. We withhold the word only because we cannot find the intelligence anywhere in the virus. We cannot find it because it is not there. It is in the relation between virus and host.

The same anthropocentrism demotes the crow. We rank it beneath us because it neither reads nor writes, while it masters its ecological niche completely. On this point the science is recent and unambiguous. Nathan Emery and Nicola Clayton, in “The Mentality of Crows” (Science, 2004), documented tool manufacture, planning for the future, and social cognition in corvids that rival the great apes — convergent intelligence assembled on a brain whose architecture shares almost nothing with a primate’s. Frans de Waal named the deeper trouble in the title of his 2016 book, Are We Smart Enough to Know How Smart Animals Are?, arguing that researchers persist in testing animals against a human yardstick and so fail to see what those animals actually do. To deny that mastery of a niche is intelligence is to mistake the yardstick for the territory — and, not incidentally, to disguise the most consequential form of intelligence, the kind that produces control, as luck, instinct, or chance.

The cybernetic thesis, restated

The claim, then, is that intelligence is relational. Human perception is built to register objects — things, individuals, models, a language model sitting in apparent isolation — but relationships are not “in” things. This is not a new heresy so much as the cybernetic thesis the field abandoned when it changed its name. Bateson, who had sat in those Macy rooms, gave it its most durable formulation in Steps to an Ecology of Mind (1972), where he defined information as a difference, and mind as the pattern of such differences rather than a substance in a skull.

The elementary unit of information is a difference which makes a difference.

Gregory Bateson, Steps to an Ecology of Mind (1972)

Half a century later, Andy Clark and David Chalmers pressed the same intuition from the opposite direction. Their paper “The Extended Mind” (Analysis, 1998) argued that cognition “ain’t all in the head,” that mind and world form a single coupled system whose boundary at the skin is an unprincipled place to draw a line. One can point at the physical composition of a virus like a photograph, while remaining fully ignorant about the relationships between it and organisms like humans. Yet it is precisely these relationships that encode function and meaning (whether it is harmless or harmful across individual organisms, replicates fast or slow or is highly contagious to some or all organisms in a population). We might go so far as to state that its existence is purely relational, given its dependence on a host organism. Whatever it seems to be, it is not. Many will perhaps conclude from the virus case that reality is little more than an encoding of relationships (e.g. information). In fact, a similar argument is often extended to our brains. They are entirely processes, encoding relationships in the world. It might stand to reason, then, that these processes or encodings could be simulated on computers. It is the relational definition that actually gives some credibility to the claim that intelligence could be computable or can be simulated, if not now, at some point. However, if we were to entertain this position in this century (21st), biological substrates would need to be inconsequential. Even today’s bio-inspired technology is more classical (in its material composition and interactions with the world) than we might imagine. It is also clear that none of the properties of a virus are realizable using other substrates. The relationships that are instrumental in defining it do not translate to profitable and durable materials like stainless steel, aluminum, wood, or plastic. We should therefore resist the temptation to assign biological powers to current computers/machines that lack the ability to model the organizational and substrate-specific encodings that make intelligence and consciousness real and not mere illustrations from our vantage point. Ability that entails a biological machine. While a quasi-biological (“artificial” biological machine) is not impossible in principe at this time, it does imply that we would need to first answer the life question, making the realization of consciousness/intelligence outside of biology a secondary pursuit. Modern AI systems are, if far from biological benchmarks, nevertheless real and intelligent from a intelligence-as-relationships perspective. Insofar as our interactions or relational encodings include AI. AI exists within the web of human relationships. Our belief in the capacities of AI are very much real in their consequences. Even if “intelligence-as-thing” represents a logical error, it predisposes anyone who holds that belief to behave differently, encoding different relationships, between themselves and the world. That could have significant social consequences.

Consequences

We have seen that he adjective “artificial” begins to dissolve when AI is relationally defined. If the substrate of reality is in part mathematical relationships, then a system that is mathematical and relational is not fake, trivial, or “artificial.” Our relational vocabulary here — learning, understanding, intelligence — gives AI systems an appropriate grounding as extensions of us rather than external, autonomous agents

When intelligence is not relational or in things and the “artificial” in AI is not challenged as we have done here, the danger of AI is not that some object in a server quietly crosses a threshold into genius. The danger is our willful ignorance of the complex relationships that comprise reality and the place for regulatory systems to manage that complexity. This is precisely the lineage the sociologist James Beniger traced in The Control Revolution (1986): the so-called information society is not a rupture but the latest turn of an industrial process that, since the nineteenth century, has met every surge in complexity — railroads, mass production, global logistics — with a new technology of control. While sociologists might disagree, “control” is itself not an inherently problematic term. Should we view AI from Beniger’s control narrative, it is just the next technology for automating the regulation of systems that have become too complex for unaided human management. The risk of control lives in the relations we are not trained to perceive. Control can hide the truth. An ideal control system is one that is understandable and clarifies rather than obscures complexity. An autonomous control system labeled as “super intelligent” in its own right is an acceptance of our ignorance and subtracts our own intelligence. That is a risky form of AI. It is the thing one points to, the hero, lifting the burden that true intelligence poses, which is to live, express, embody, relate, and be intelligence.

The cyberneticists of 1946 kept the relation in view. The rebranding of 1956 installed a ghost in the machine, and the decades since have been spent searching the machine for it. The more exact question is not whether a system is intelligent but what it is intelligent in relation to — the host it defeats, the niche it commands, the complex world it silently regulates. That is where intelligence has always been. The call or burden of intelligence is to embed ourselves more fully and authentically in the world. But are we intelligent enough (that is, sufficiently interconnected and self-aware) to grasp this?