Three Axes of Mind

Every few years, a machine does something that experts said would take a decade, and the reaction is the same: impressed murmuring, followed by the same refrain. Impressive, but still narrow.

Chess grandmasters said it when Deep Blue defeated Kasparov in 1997. Go masters said it when AlphaGo dismantled Lee Sedol in 2016. Radiologists said it when deep learning systems began outperforming them on certain diagnostic tasks. Researchers said it when large language models, the systems behind the current generation of conversational AI, began producing legal arguments, passing medical licensing exams, and writing code that compiled and ran. Each time, the threshold was crossed, the benchmark fell, and the conversation moved the goalposts.

Call it what you like, but not intellectual dishonesty. It is a symptom of a genuine problem: we do not have a coherent theory of what we are measuring. We know performance when we see it. We are much less certain what performance reveals. And the gap between those two things, between what a system does and what a system is, has become one of the most far-reaching questions in contemporary life.

This recognition paradox runs in both directions. We are prone to dismiss systems that outperform us on tasks we value, by moving the bar. But we are equally prone to attribute depth and understanding to systems that merely replicate the surface texture of intelligent behavior. Fluency triggers attribution: a system that speaks well is assumed to know what it is saying.

Both errors, premature dismissal and premature attribution, compound badly at scale. If we cannot tell what a system is, we cannot govern it responsibly, collaborate with it wisely, or know when to trust it and when to resist it.

What we need is not a verdict. We need a map.

What follows is a new coordinate system: three axes along which any cognitive system, human, animal, institutional, or artificial, can be located without collapsing everything into a single ranking. The axes are Availability, Integration, and Depth.

The Problem with Single-Axis Theories

Most influential theories of mind share a structural limitation: each privileges a single dimension, trying to explain too much with too little.

Global Workspace Theory, one of the most influential cognitive frameworks of the past few decades, focuses on access, on whether information is globally broadcast across a system and made available for reasoning, report, and control. It explains a great deal. It tells us something real about the difference between conscious and unconscious processing, about why divided attention degrades performance, and about how anesthesia works. But it treats consciousness as essentially a question of availability. What can the system report? What information is globally online?

Integrated Information Theory focuses on unity, on whether the internal causal structure of a system is irreducibly whole rather than decomposable into independent parts. It too explains something real. It connects to an observation about how experience presents itself from the inside: experience is unified. When you perceive a red cube, the color and the shape and the object are not experienced as separate data points. They arrive as one thing. But in its standard formulations, IIT’s account of consciousness lives almost entirely in the present moment. It has little to say about memory, persistence, or the slow accumulation of history that makes experience feel like a life rather than a sequence of isolated moments.

Higher-order theories focus on self-representation, on whether a system has states that represent its own states. Predictive processing accounts emphasize the constant generation and revision of internal models. Each framework illuminates part of the picture. None of them, alone, captures what makes human minds both so powerful and so distinctive.

The difficulty is that intelligence, sentience, and consciousness are not different words for the same thing. Solving problems is one capacity, feeling is another, and having an inner point of view is a third. They are related but distinct capacities that can come apart, and frequently do. A system can be extraordinarily capable along one dimension while remaining limited along the others. Treating them as a single scale forces false choices: either this system has it, whatever it is, or it doesn’t.

The three-axis framework refuses that simplification. It lets you locate a system on a map rather than assigning it a verdict.

The First Axis: Availability

The first axis concerns what information is accessible to the system as a whole.

Think of a lion pride coordinating a hunt across an open savanna. Individual lions track different variables, wind direction, prey movement, terrain features, and the positions of other hunters. None of them holds the complete picture. But information flows across the group: adjustments in one lion’s position trigger responses in others, flanking movements emerge without any single lion directing them, the hunt achieves coordination through distributed attention rather than centralized command. Information that originates in one part of the system becomes available to the whole.

Availability describes exactly this. Not merely intelligence or processing power, but the degree to which information can be broadcast across a system, made globally accessible for reasoning, action, response, and integration with other information.

Availability explains a cluster of things we associate with sophisticated intelligence. Flexible reasoning requires information to be mobile across contexts. The ability to report on one’s own states, to say what you know, what you believe, and what you’re uncertain about, requires that those states be globally accessible. Deliberate control over behavior requires that information about the environment and the system’s own goals can circulate and constrain action. The difference between unconscious reflex and conscious choice is partly a difference in availability: whether the triggering information was globally broadcast or processed locally and silently.

Current AI systems score high on this axis. A large language model has access to an extraordinary breadth of information across domains, and that information is available for flexible deployment across an enormous range of tasks. The system can bring its chemical knowledge to bear on a legal question, apply its understanding of narrative structure to a technical document, and connect ideas across domains in ways that would be slow and effortful for any individual human expert. That capability is genuine, and it would be a mistake to dismiss it. But Availability alone explains neither everything we value about intelligence nor everything that makes minds morally significant. A system can have global access to vast information while lacking the internal structure that makes that information cohere into something like understanding, or the accumulated history that makes it cohere into something like character.

Which brings us to the second axis.

The Second Axis: Integration

Imagine sitting at your desk late in the afternoon. There is a red cube on the surface in front of you. Coffee is brewing somewhere behind you, and its smell has been drifting into the room for ten minutes. You have been trying to concentrate, but a task you forgot to complete this morning has been hovering at the edge of your attention. You are aware of all of this at once.

So ordinary an experience it seems barely worth describing. But notice what is actually happening. The color red, the geometric form of the cube, the warmth of the coffee smell, the faint anxiety of the forgotten task, these are processed by different neural systems, using different mechanisms, and receiving different sensory inputs. Yet you do not experience them as separate data streams. You experience a single moment: I am here, at my desk, with this cube and this coffee and this thing I should have done. Each element constrains the others. The guilt about the forgotten task colors your perception of the coffee’s smell. The red cube sits in your visual field as an object you are aware of half-attending to.

This is what Integration describes: the degree to which the internal states of a system form a unified causal whole rather than a collection of parallel, independent processes.

High Integration means that the components of a system participate in each other’s causation, that what happens in one part of the system makes a difference to what happens in other parts, not by sending messages across a network, but by being genuinely unified in how they process and represent the world. The experience of a single moment rather than a sequence of data points is the signature this integration leaves. Whether that unity is also felt, whether there is something it is like to be a causally unified system, is a question Chapter 5 takes up.

The inverse reveals why this matters. Imagine a system that has access to all the same information, the color, the shape, the smell, and the memory of the forgotten task, but processes each stream independently, with no genuine causal integration between them. Such a system could report on each item separately. It could answer questions about the cube, about the coffee, about the task. But there would be no single vantage point from which all of these things are experienced together, no moment of unified perception in which the guilt and the smell and the visual scene constrain each other in real time. The system would have Availability without Integration, and something essential to what we call experience would be missing.

Integration also explains a phenomenon that any attentive person notices when interacting with sophisticated AI systems: the sense that coherence can fracture under pressure. A system that reasons impressively in straightforward contexts may begin to contradict itself when constraints multiply, when goals conflict, or when the conversation requires maintaining a consistent stance across many exchanges. The components are performing individually, but they are not integrated into a unified system that holds together under load.

The point is not to criticize. It is to locate. The observation tells us something about where on this axis current systems sit, and what we would need to see to be confident that the axis is being developed.

The Third Axis: Depth

In Tarangire, Tanzania, researchers documented something that has stayed in the scientific literature for decades. During the severe drought of 1993, elephant families led by the oldest matriarchs were far more likely to leave the park in search of water and forage, and their calves survived at markedly higher rates than the calves of families led by younger females. The matriarchs old enough to lead their groups to safety were the ones old enough to remember the last drought of comparable severity, decades before. The knowledge persisted not as conscious recollection in the human sense, but as something older and more structural: a disposition built into their understanding of the landscape by a lifetime of experience, available to guide behavior when conditions matched.

This is what Depth describes: the degree to which a system’s present state encodes its own causal history. The history in question is the accumulated experience, learning, and self-modification through which the system has passed. This is Depth as a structural property. Whether that accumulated structure is turned to good ends is a separate question, one Chapter 6 takes up.

Depth is not memory in the simple sense of information storage. A hard drive stores information. What the elephant matriarch carries is something more integrated: history that has become structure, that has shaped not just what she knows but what she is, her responses, her capacities, her dispositions. The past is not archived somewhere and retrieved on demand. It is compressed into the present, shaping it from the inside.

To assemble time is precisely this: the mechanism through which Depth accumulates, and the difference between the hard drive and the matriarch. Time is assembled when experience is not merely recorded but integrated, when what happens to the system becomes part of the system rather than a copy held inside it. It is assembled when traces compound, when each new experience lands on the residue of earlier ones and takes part of its shape from them, and when learning modifies the learner in ways that persist and constrain future learning, so that the system that meets tomorrow’s experience has been changed by today’s. A system high in Depth is one whose present moment carries the weight of its own history.

Depth explains several things that neither Availability nor Integration can account for alone.

It explains continuity of character, the sense that a person is recognizably themselves across contexts, across years, across challenges that test them. Character is not simply a policy for behavior; it is the accumulated structure of choices and consequences that has shaped a system from the inside. You can see it most clearly when it holds under pressure: when someone acts consistently with who they have been, even when the environment rewards otherwise.

It explains the difference between expertise and performance. A novice musician can, with careful attention, produce technically correct notes. An experienced musician plays with a kind of inevitability, as though the music is emerging from somewhere inside rather than being assembled note by note. What the expert carries is decades of accumulated integration between intention, execution, feedback, and revision. The history is in the hands, not just the memory.

It explains why we find something unsettling in interactions with systems that are impressively capable but somehow weightless: what we are detecting, often before we can articulate it, is the absence of Depth. Chapter 15 develops that diagnosis in full.

Current AI systems are, as Chapter 14 will explore in detail, expanding horizontally, gaining Availability at extraordinary speed, while Depth remains uncertain, undeveloped, or differently structured than human intuition expects. None of this is a permanent limitation. It is a current location on a map. But locating it accurately matters for everything that follows.

Three Axes, Not One

The value of the framework only becomes clear when you use all three axes at once, and when you resist the temptation to collapse them into a single scale.

Consider how Availability, Integration, and Depth show up in different combinations across three familiar systems.

A modern language model has high Availability and questionable Integration under pressure. Its Depth is the hardest of the three to read, because it is structurally different from biological Depth. The model was trained on a vast sweep of recorded human history, and that history shaped what it is. But it has not undergone that history in any sense that accumulates into an ongoing self: being shaped by a record is not the same as living through the events the record describes, one at a time, with each one landing on a self that persists to the next. That distinction is Chapter 6’s work.

A honeybee has low Availability in the broadcast sense, moderate Integration within its sensory-motor system, and Depth that is partly individual and partly inherited. The bee’s behavior reflects the evolutionary history of its lineage as much as its individual experience.

An institution like a well-functioning hospital or a long-established legal system has Availability distributed across many individuals, Integration that is structural and procedural rather than experiential, and Depth encoded in norms, case precedents, protocols, and institutional memory that persists longer than any individual member’s tenure.

None of these systems is simply more or less intelligent. They are differently constituted across the three axes. Which one is most relevant depends on what you need to know and why.

The axes are distinguishable, not orthogonal: you can tell which is which, but they do not vary independently. High values tend to travel together, and Depth presupposes some Integration the way a building presupposes a foundation: you cannot accumulate integrated history without integration to accumulate it in.

What keeps the three from collapsing into one scale is that each leaves its own signature, readable even when the others are present. Availability shows in breadth and reportability, whether the system can bring what it knows in one domain to bear in another and say where it is uncertain. Integration shows in coherence under load, whether a unified response holds when goals conflict or the seams show instead. Depth shows in the costliness of reversal and the other signatures Chapter 6 sets out: what a system refuses, what it preserves under pressure, how it degrades.

These signatures are a diagnostic heuristic, a guide for looking, not yet a set of validated, independent measurements. Naming them is not the same as measuring them, and at this stage the framework claims only the first; building instruments that hold up under real diagnostic pressure is part of the work still ahead.

A diagnosis made with that caution reads the signatures rather than a single number, and the axes earn their keep precisely where the signatures come apart: where a system is broadly available but fractures under load, or coherent in the moment but carries no history.

This matters for practical reasoning in ways that single-axis accounts cannot support. If you ask only “is this system intelligent?”, the single-axis question, you get an answer that carries almost no information about what the system will do when conditions change, whether it will hold together under pressure, or whether it can be trusted with decisions whose consequences unfold over time. If instead you ask where a system sits on Availability, Integration, and Depth separately, you get a much richer map: what it can do, how it holds together, and how much its history constrains its present behavior.

The framework also resists the twin errors from Part I, the same pair this chapter opened with. Inflation, treating surface fluency as evidence of depth, is the error of collapsing the axes: seeing high Availability and concluding that Depth must be present too. Dismissal, treating interior opacity as proof of triviality, is its mirror: seeing uncertain Depth and concluding that Availability and Integration count for nothing, morally or practically. The three axes hold these apart — genuine capability acknowledged without overstating what it reveals, genuine uncertainty about Depth acknowledged without letting it dismiss everything else.

Where Consciousness Lives on the Map

The axes locate systems; they do not, by themselves, settle whether any given system is conscious. How Availability, Integration, and Depth combine to constitute conscious experience is the question Chapter 5 takes up.