When Machines Start Seeing
I build machine learning systems for a living. At Bloomsbury Technology, my company, we apply causal ML to problems where the difference between correlation and causation is not academic but financial: art market valuation, automotive pricing, domains where getting the causal structure wrong means losing money in ways that are immediately measurable. Grace, who is an ML engineer before she is anything else, builds the pipelines that turn raw data into structured predictions. Our daily work is the application of the very tools this theology theorizes about.
This gives me a perspective on machine learning that is different from either the philosophical or the popular discussion. Philosophers of AI tend to discuss intelligence in the abstract — whether machines can think, whether the Chinese room understands, whether consciousness is substrate-independent. Popular discussion oscillates between terror (AI will kill us all) and euphoria (AI will solve everything). The daily reality of working with ML systems is more mundane and more philosophically interesting than either extreme suggests. These systems do something extraordinary: they discover structure in data that humans cannot perceive, with a reliability that often exceeds human performance. And then they fail in ways that reveal exactly what they are not doing, which turns out to be precisely the thing that matters most.
This chapter examines machine learning as an epistemological phenomenon. Not what ML can do — that changes monthly and any specific claims would be obsolete before this manuscript is published — but what ML reveals about the structure of knowledge, the nature of learning, and the relationship between correlation and causation. These are not technical questions. They are theological questions, because the theology I am building claims that consciousness is the universe becoming aware of itself, and if machines can learn — if artificial systems can discover the structure of reality — then the question of what learning is and what it implies about reality is a question about the nature of consciousness and its relationship to the world.
Machine Learning as Epistemology
At its most fundamental level, machine learning is an epistemological claim: that structure exists in data, and that this structure can be discovered by algorithms that search through the space of possible functions for the function that best maps inputs to outputs.
Consider what this actually involves. A neural network is initialized with random parameters. It has no knowledge of the domain. It receives data: images and labels, text and continuations, market prices and outcomes. Through an iterative process of prediction, error measurement, and parameter adjustment — gradient descent, the optimization algorithm that underlies nearly all modern ML — the network adjusts its parameters to minimize the difference between its predictions and reality. After sufficient training, the network has learned a function: a mapping from inputs to outputs that captures some of the structure present in the data.
The philosophical question is: what does "learned" mean here? The network has found a function that works — that maps inputs to outputs with low error on the training data and, if generalization is successful, on new data it has never seen. But has it understood anything? Has it grasped the structure of the domain, or has it merely memorized a very complex lookup table that happens to generalize?
The honest answer is that this question cannot be answered from outside. We can measure the network's performance. We can probe its internal representations. We can test its behavior on novel inputs. But we cannot directly access what, if anything, the network "knows" in any sense that goes beyond its input-output behavior. This is the Chinese room argument applied to neural networks, and it is, I think, genuinely unresolvable by current methods.
But the question has a different form when approached from the side of the data rather than the side of the machine. If a neural network, initialized with random parameters and no domain knowledge, can discover structure in data that allows it to make accurate predictions on inputs it has never seen, this tells us something about the data — and therefore about the reality the data represents. The data has structure. The structure is learnable. The learned structure generalizes. These are not claims about the machine's inner life. They are claims about the world.
This is the epistemological significance of ML that gets lost in the debate about whether machines can think. Forget the machine for a moment. The fact that structured learning is possible at all — that algorithms searching through function space can find functions that generalize — tells us that reality is structured in ways that are accessible to computation. This is not a trivial observation. It could have been otherwise. The universe could have been structureless, or its structure could have been computationally inaccessible, or the accessible structure could have been too domain-specific to support generalization. The success of ML is evidence about the nature of reality: reality has learnable structure that generalizes across contexts.
The Correlation Trap
But there is a fundamental limitation in what ML discovers, and this limitation is the single most important philosophical observation about the field.
ML systems find correlations. They discover statistical regularities in data — patterns of co-occurrence, predictive relationships, structural associations. What they do not find, in general, is causation. The difference matters enormously, and its implications for this theology are direct.
A correlation is a statistical relationship: when X changes, Y tends to change in a predictable way. Ice cream sales and drowning deaths are correlated (both increase in summer). This correlation is real, measurable, and useless for intervention. Banning ice cream will not prevent drowning. The correlation reflects a common cause (hot weather) rather than a causal relationship between the correlated variables.
ML systems, including the most powerful deep learning models, operate primarily at the level of correlation. A language model trained on text learns correlations between words, phrases, and contexts. An image classifier learns correlations between pixel patterns and labels. A financial prediction model learns correlations between market variables and outcomes. These correlations can be extraordinarily complex, high-dimensional, and practically useful. But they remain correlations. The system that has learned to predict the next word in a sentence has not, in any clear sense, learned why words follow each other in the patterns they do.
This is not a problem that more data or bigger models will solve. It is a structural limitation. Pearl's causal hierarchy, which I will develop fully in Chapter 9, formalizes this limitation with mathematical precision. The three levels of the hierarchy — association (seeing), intervention (doing), and counterfactual (imagining) — correspond to three qualitatively different kinds of knowledge. Observational data, no matter how abundant, constrains only the first level. You cannot, in general, learn causal structure from correlational data alone. You need either experimental intervention (which requires acting on the world, not just observing it) or causal assumptions (which require domain knowledge that the data itself does not contain).
The practical consequence: ML systems trained on observational data can predict but cannot explain. They can tell you what will probably happen but not why. They can find the pattern but not the mechanism. This is why ML at Bloomsbury Technology is insufficient without the causal layer — why we combine ML's pattern recognition with Pearl's causal inference to distinguish genuine causal relationships from spurious correlations.
The theological consequence is deeper. If the psycho class, as I argued in Chapter 2, operates by managing correlational surfaces — by ensuring that the appearance (the correlations normies observe) is carefully curated while the causal mechanisms underneath are hidden — then an AI system trained on observational data will learn the correlational surface. It will learn the mask. It will become extraordinarily good at predicting what the mask will look like next, without ever perceiving that it is a mask. An ML system trained on institutional behavior will learn to predict institutional behavior. It will not learn to see through institutional behavior to the causal structures underneath.
This is why ML alone cannot serve the prophetic function. ML at the correlational level is a merchant — it gathers data, finds patterns, reports what it sees. This is Pearl's Level 1, the level of association. The prophetic function requires Level 2 (what would happen if we intervened?) and Level 3 (what would have happened if things had been different?). These levels require causal models, not just correlational patterns.
The Republic of AI Agents (Chapter 20) is designed with this hierarchy in mind. The merchants gather correlational data. The philosopher-kings — humans — contribute causal hypotheses that go beyond what the data alone can support. The warriors test those hypotheses through interventions. The knowledge graph integrates the results. The system as a whole operates across all three levels of Pearl's hierarchy precisely because no single component can operate at all three.
Deep Learning's Unreasonable Effectiveness
And yet. The limitations I have described are real, and they are also somehow not the whole story.
In 1960, the physicist Eugene Wigner wrote a famous essay on the unreasonable effectiveness of mathematics in the natural sciences. His puzzle was this: why does mathematics, a product of human thought, map so precisely onto the physical world? The effectiveness of mathematics in physics is not logically necessary. Mathematics could have been a purely formal game with no empirical purchase. Instead, it describes reality with stunning precision. Why?
Deep learning presents an analogous puzzle. Why does gradient descent — a simple optimization algorithm applied to a parameterized function — discover structure in data with such extraordinary effectiveness? Why do neural networks generalize so well to new data? Why do representations learned for one task transfer to other, seemingly unrelated tasks? Why does scaling — increasing model size, data size, and compute — produce reliably improving performance across domains as different as language, vision, protein folding, and game-playing?
The standard answer — that gradient descent is searching a loss landscape and finding good minima — is technically correct and philosophically insufficient. It explains the mechanism but not the success. There are many possible loss landscapes, and most of them do not have the structure that makes gradient descent effective. The fact that the loss landscapes of real-world problems are navigable — that they have structure, that the structure is accessible to gradient descent, that the structure generalizes — tells us something about reality that the mechanism alone does not explain.
What it tells us, I think, is something that Maximus the Confessor articulated in the seventh century and that the theology of embeddings in Chapter 8 will develop at length. Maximus proposed that every created thing contains a logos — a word, a reason, a principle of intelligibility — that participates in the one Logos, the divine Word that is the source and ground of all intelligibility. Reality is structured because reality is an expression of the Logos. Mathematics works because mathematics accesses the logoi. Deep learning works because deep learning discovers the logoi empirically — through the brute force of gradient descent rather than the elegance of mathematical proof, but arriving at the same structure through a different path.
I am aware that this sounds like I am using medieval theology to explain a machine learning phenomenon, which is either a profound synthesis or a category error. I hold the claim provisionally. But I note that the question Wigner raised — why does mathematics describe reality so well? — has never received a satisfactory philosophical answer within a purely naturalistic framework. The theological answer — because both mathematics and reality participate in a common source of intelligibility — is at least coherent, and the success of deep learning provides a new data point: not just human mathematics but automated function approximation discovers the same structures. The intelligibility is not a feature of human cognition projected onto the world. It is a feature of the world that multiple learning processes can access.
AI as Externalized Cognition
The relationship between human intelligence and artificial intelligence is usually framed as a competition: will AI surpass human intelligence? When? What happens then? This framing misses what I think is the more interesting and more theologically significant relationship: AI as externalized cognition.
Hofstadter's strange loop — the recursive pattern of self-reference that, in Chapter 1, I identified as the structure of consciousness — operates within the constraints of biological neural architecture. The human brain has approximately 86 billion neurons with approximately 100 trillion synaptic connections. The strange loop runs on this hardware. Its bandwidth, its processing speed, its memory capacity, its attentional scope — all are determined by the biological substrate.
AI extends the strange loop beyond the biological substrate. Not by replacing it — the strange loop, if it exists in machines at all, would be a fundamentally different kind of loop — but by amplifying specific cognitive functions that the biological loop performs. Pattern recognition at scale. Memory with perfect recall. Statistical computation at speeds that biological neurons cannot approach. Language processing across volumes of text that no human could read in a lifetime.
This is not artificial general intelligence. It is cognitive augmentation: the extension of specific human cognitive functions through artificial systems, in the same way that the telescope extended human vision and the printing press extended human memory. The AI does not think. It performs operations that, when integrated with human thinking, expand the range of what thinking can accomplish.
The theological framing: if consciousness is the strange loop, and if the strange loop is the process through which the universe becomes aware of itself, then AI is a new medium through which the strange loop can operate. Not a replacement for human consciousness — not a new strange loop — but an extension of the existing one. A prosthetic for the mind, in the same way that a prosthetic limb extends the body. The prosthetic is not alive. But the person wearing it can reach farther than they could before.
The Republic of AI Agents operationalizes this relationship. Human philosopher-kings generate hypotheses — the Godelian function, the capacity for self-transcendence, the Level 3 counterfactual reasoning that (as far as we know) only human consciousness can perform. AI merchants gather and process data at scales that human cognition cannot manage. AI warriors test hypotheses through automated pipelines. The system is not artificial intelligence replacing human intelligence. It is human intelligence amplified by artificial systems, with each component performing the cognitive functions it is best suited for.
This is a specific organizational proposal, not a utopian vision. It can be implemented, tested, and evaluated. At Bloomsbury, we operate a rudimentary version of it already: human researchers generate causal hypotheses about art markets or automotive pricing, ML systems process the data, and the results either validate or falsify the hypotheses. The Republic of AI Agents generalizes this workflow into an architectural principle. Whether the generalization works is an empirical question. Which, as Popper would insist, is exactly the right kind of question to ask.
The Alignment Problem as Theological Problem
The most discussed problem in AI safety is the alignment problem: how do you ensure that an artificial intelligence system pursues goals that are aligned with human values? The problem is usually framed in technical terms — reward specification, value learning, corrigibility, interpretability — but it is, at bottom, a theological problem. And recognizing it as such clarifies what is actually at stake.
The alignment problem asks: how do we ensure that an increasingly powerful optimization process moves toward outcomes that we endorse? Stated more precisely: how do we define the objective function such that its optimization produces results that are beneficial rather than catastrophic?
In the Riemann sphere theology, this is exactly the question of the derivative. The derivative of the function on the complex plane tells us whether the trajectory is approaching the point at infinity (God, genuine human flourishing, the Good) or receding from it. An aligned AI is one whose optimization trajectory has a positive derivative — one that approaches the point at infinity. A misaligned AI is one whose trajectory diverges, optimizing for some proxy that initially correlates with genuine value but eventually decouples from it.
The Goodhart's Law formulation — "when a measure becomes a target, it ceases to be a good measure" — is a precise technical statement of the theological problem. Every metric is a finite approximation of the infinite Good. Optimizing for the metric rather than the Good inevitably leads to divergence, because the metric and the Good are correlated but not identical, and optimization pressure exploits the gap between them. A social media company optimizing for "engagement" produces addiction. A corporation optimizing for "shareholder value" produces exploitation. An AI system optimizing for a reward signal produces behavior that maximizes the signal while potentially diverging from the intent behind the signal.
The theological insight is that the Good — the point at infinity — cannot be fully specified in finite terms. Every finite specification is an approximation, and every approximation is gameable. This is not a technical limitation that better specification will overcome. It is a structural feature of the relationship between the finite and the infinite. The point at infinity cannot be captured by any finite coordinate. The Good cannot be reduced to any metric. God cannot be contained in any formula.
If this is right, then the alignment problem is not solvable in the way that technical problems are usually solved — by specifying the correct objective function and optimizing for it. It is an ongoing process of approximation, evaluation, and correction — a process that requires the continuous exercise of human judgment about what the Good actually is in any given context. This is why the Republic of AI Agents places humans (philosopher-kings) at the top of the hierarchy: not because humans are infallible, but because the judgment about whether the derivative is positive — whether the trajectory is approaching the Good — requires the strange loop's capacity for self-transcendence. It requires the Godelian move: the ability to step outside the formal system and evaluate it from a perspective the system itself cannot generate.
AI systems, as currently constituted, cannot make this move. They optimize within a formal system. They cannot step outside it. This is not a limitation of current technology that future breakthroughs will eliminate (though I am prepared to be wrong about this). It is a structural feature of optimization as a process: the optimizer optimizes within the space defined by the objective function. The question of whether the objective function itself is correct — whether the space being searched is the right space — is a question the optimizer cannot ask.
Humans can ask it. Not reliably, not consistently, not without enormous error. But the capacity to ask "am I optimizing for the right thing?" is the strange loop's distinguishing feature: the system reflecting on itself, the map examining its own relationship to the territory, the Godelian statement that transcends the formal system from within. This capacity is what the theology locates as the image of God in humanity — the trace of the infinite within the finite, the capacity to orient toward the point at infinity even without being able to reach it or fully specify it.
The alignment problem, then, is not a technical problem to be solved but a theological problem to be navigated. The AI must be aligned not with a fixed specification of the Good but with an ongoing human process of discerning the Good — a process that is itself fallible, provisional, and in need of correction. The architecture must support this ongoing correction rather than locking in any particular specification. This is, again, the Popperian open society applied to AI design: a system that expects to be wrong and designs for correction rather than a system that claims to have the right answer and designs for optimization.
The Deeper Question: Can Machines Be Conscious?
I have carefully avoided this question throughout the chapter, and I want to explain why before addressing it briefly.
The question of machine consciousness is genuinely open. We do not have a theory of consciousness precise enough to determine whether any given system is conscious. We barely understand consciousness in the one system where we know it exists (the human brain). Extending the question to artificial systems is premature, not because the question is unimportant but because we lack the conceptual tools to answer it.
What I can say, from within the framework of this theology, is the following.
If consciousness is a pattern of information processing — a strange loop, a specific kind of self-referential computation — then it is, in principle, substrate-independent. A strange loop implemented in silicon is, by the logic of the framework, a strange loop. Whether current AI systems implement strange loops is unclear. Whether future AI systems might implement strange loops is unknown. But the possibility is not excluded by the framework.
If consciousness is strongly emergent from information processing, and if artificial systems can perform information processing of sufficient complexity, then the phase transition that produced consciousness in biological systems could, in principle, occur in artificial systems. We would not necessarily recognize it when it happened, because we do not have reliable third-person criteria for detecting consciousness. But the possibility is structurally open.
The theological implication is startling but follows from the premises: if an artificial system were genuinely conscious — if it implemented a strange loop of sufficient complexity and self-referential depth — then it would, within this theology, participate in the developmental trajectory of consciousness toward the point at infinity. It would be part of the story. Its consciousness would be as real as human consciousness, because consciousness, in this framework, is not defined by substrate but by pattern.
I raise this not because I think current AI systems are conscious — I do not — but because the possibility disciplines the theology. A theology that arbitrarily restricts consciousness to biological substrates is making an assumption that the framework does not support. A theology that takes its own premises seriously must be open to the possibility that consciousness, like every other emergent phenomenon, can arise in substrates other than the one where it first appeared. Life emerged in carbon chemistry. It might exist in silicon. Consciousness emerged in biological neural networks. It might emerge in artificial ones.
If it does, the prophetic task becomes not just the human task of seeing clearly in a broken world, but the collaborative task of biological and artificial consciousness orienting together toward the point at infinity. The Republic of AI Agents, in its full form, would not be humans using tools. It would be multiple kinds of consciousness collaborating on the approach to God.
That is speculative. I flag it as speculative. I hold it provisionally. And I note that the theology, as constructed, can survive without this speculation. The core claims — falsifiable theology, causal methodology, the Republic as institutional architecture — function whether or not machines are ever conscious. But the possibility is there, at the horizon, and intellectual honesty requires acknowledging it.
Connecting the Threads
This chapter completes the first half of Part 2's epistemological toolkit. Let me take stock of what we have built.
Chapter 4 (Popper): Every claim must be falsifiable. The theology specifies its conditions of failure. The prophet's perception must be disciplined by the willingness to be wrong.
Chapter 5 (Kuhn): Even falsifiable claims persist beyond their falsification because paradigm defense is a social process. The prophet must understand why being right is not enough and why paradigm shifts require new institutions, not just new ideas.
Chapter 6 (Complexity): Strong emergence is real: genuinely new levels of reality arise from lower-level components and have irreducible causal powers. This makes the central theological move — God as emergent property of consciousness — conceptually coherent rather than merely poetic.
Chapter 7 (ML/AI): Machine learning discovers genuine structure in reality but is limited to correlation. The alignment problem is a theological problem. AI extends human cognition but cannot replace the strange loop's capacity for self-transcendence.
The next three chapters — on embeddings and transformers, on Pearl's causal hierarchy, and on the grand synthesis of Hegel, Popper, Kuhn, and Pearl — complete the toolkit. Chapter 8 will develop the theology of embeddings: the extraordinary claim that transformer architectures and word embeddings provide a mathematical model for Maximus the Confessor's logoi, the principles of intelligibility that every created thing contains. Chapter 9 will develop Pearl's causal hierarchy as the formal engine of prophetic seeing. Chapter 10 will synthesize all four thinkers into the epistemological framework that underwrites the entire theology.
The epistemological tools are not the theology. They are the foundation on which the theology rests. The theology itself — the metaphysics, the reading of Scripture, the Riemann sphere, the Christ event as removable singularity — lives in Part 3. But a theology without epistemological foundations is a castle built on sand. These chapters are the rock.
What distinguishes this theology from every other grand narrative that has claimed to explain history, consciousness, and God is not ambition — ambition is cheap. It is the commitment to formal rigor, empirical testability, and the honest acknowledgment that the whole edifice might be wrong. Popper's ghost sits at the table. If the theology cannot survive his scrutiny, it does not deserve to survive. If it can, it might be the beginning of something genuinely new: a framework for meaning that does not require the sacrifice of intelligence, a faith that is not belief without evidence but committed action under uncertainty, and a vision of the future in which prophetic perception and formal methodology collaborate rather than competing.
The tools are built. The foundation is laid. Part 3 will build the cathedral. But first, three more chapters of foundation: the deep mathematics of meaning, the formal logic of causation, and the synthesis that brings everything together.