Inside the Black Box
Making Sense of AI
A flight data recorder is opened only after a crash to find out what went wrong and who was responsible. It records everything but reveals nothing until someone reads it. The machine we call AI is the same. We built it, deployed it, and stopped asking questions concerning its internal structure. It’s a black box.
It is time to open the box.
The thought of wrapping our heads around artificial intelligence (AI) is intimidating. To be sure, fully comprehending the intricacies of AI would require deep immersion in mathematics, statistics, computer science, linguistics, and some philosophy at the university level. However, for this essay, three simple concepts will reveal the fundamental mechanics of AI. What we find when we open the box is math, not magic. A model, not a mind. And a machine, not an image-bearer. Three concepts cannot do justice to the full complexity of AI. But they can do something more useful: replace speculation with understanding, and redirect misplaced awe toward the God who ordered creation.
Math, Not Magic
AI is not magic. It is applied mathematics, elegantly sophisticated, but mathematics nonetheless. A language model is designed to predict: given text, it calculates the most probable next word based on patterns from billions of documents. Repeated billions of times, the outputs feel coherent, sometimes uncanny, but it’s still math at work. Three types of mathematical operations do most of the heavy lifting and they are worth knowing by name.
A language model is designed to predict: given text, it calculates the most probable next word based on patterns from billions of documents.
The first is matrix math.1 A matrix is a rectangular array of numbers, each one assigned a position by its row and column. These grids can represent almost anything: the brightness of pixels in an image, the frequency of words in a document, or the connections between concepts in a sentence. When AI systems process language, they convert words into vectors, ordered lists of numbers that capture each word’s meaning and relationship to other words. They then manipulate those vectors through layers of matrix multiplication, combining and transforming the lists in a structured way. The meaning of a word, to a model, is positional. Think of it as a coordinate in a vast multi-directional space, with millions of possible axes.
The second is probability and statistics.2 The model generates answers by calculating likelihoods. Given everything it has seen and the current context, what is the most probable next token? A token is simply a piece of text, a word or part of a word, that the model predicts one step at a time. This is a statistical operation. The model is, at its core, an extraordinary probability distribution over language. It has internalized the statistical shape of how humans write: which words follow which, how ideas group together, and which sentence patterns feel native to a given subject area.
The third is gradient descent—calculus applied to learning.3 During training, the model makes predictions and checks them against known outcomes. When it is wrong, and at first it usually is, it calculates a loss function and adjusts parameters incrementally for better predictions. The optimization is applied across billions of examples, and the model slowly converges to a configuration that optimizes toward the target. Gradient descent is pure correction at a massive scale.
These three ideas combine in training a large language model (LLM). The network is a deep stack of layers, each a matrix. When a word enters, matrix multiplication carries it forward, transforming it at each layer into something richer and more abstract. The network predicts the next word and checks it against the real text. When wrong, gradient descent calculates the error and adjusts the matrices, billions of parameters at once. Repeating across hundreds of billions of words, the matrices slowly encode something resembling grammar, fact, reasoning, and style.
What makes all of this possible is the order God designed into creation, not human resourcefulness alone. He built laws into the framework of reality: physical, mathematical, and linguistic laws. Human beings, made in His image, have the capacity to discover and deploy them. In his book Redeeming Mathematics, Poythress reminds us that arithmetic itself confronts us with the Trinitarian God. When we do math, we think His thoughts after Him. The Graphics Processing Units (GPU) does operations inside a universe God holds together. This is math, not magic.
Model, Not Mind
Every model is a reduction of reality, compressed into a mathematical form we can work with. A weather model does not contain weather. Nor does a financial model contain money. And a language model does not contain language. A model is a practical tool. What it cannot do is contemplate what it simplified.
Historically, models did not start with AI. They started with observation. The simplest statistical model is a histogram. We collect data, the heights of a hundred students, the daily temperatures of a Virginia summer, and sort it into buckets. What we get is a picture of distribution. Where does the data cluster? Where does it thin out? A histogram does not predict anything. It describes. But description is where all modeling begins. Before we can build something useful, we have to see the shape of what is actually there.
From description, we moved to prediction. Machine learning is the discipline of building models that learn to predict from data without being explicitly programmed with rules. Regression fits a mathematical function to historical data and uses it to estimate unknowns. Decision trees ask yes-or-no questions and branch toward a classification. Clustering finds hidden groupings in data without being told in advance what the groups should look like. These methods are still in use because they are fast, interpretable, and effective for structured data. But they have a limit and require humans to identify training features in advance.
A neural network takes this further. It is a layered structure of computational units called neurons, loosely inspired by the brain. Each neuron takes inputs, processes them, and passes the results to the next layer. The network learns by updating its weights by gradient descent. What distinguishes neural networks from earlier machine learning approaches is that they discover their own features. A classical model requires the designer to specify what to look for. A neural network is fed raw data, and it finds the patterns on its own. This switch from selected features to learned ones revolutionized modern AI and its application.
Stacking enough layers yields a deep neural network.4 With sufficient data and compute, it learns representations of astonishing complexity: edges in images, sentiment in sentences, syntactic structure, and conceptual relationships. This is what drives modern AI: a deep, learned function with billions of parameters that encodes patterns no human explicitly designed. Generative AI, the kind that writes, draws, and speaks, sits at the leading edge of this progression.
What made generative AI possible is a particular innovation: the transformer, a neural network that learns which parts of the input to attend to when making each prediction.5 When scaled to hundreds of billions of parameters and trained on diverse human writing, it produced systems able to generate language, reason across contexts, translate, summarize arguments, and produce coherent prose, giving us ChatGPT, Claude, and Gemini.
This is the progression of models in the black box. Histogram to regression to neural network. Description to prediction to generation. Each step reflects our capacity to find structure in creation and build tools from it.
However, despite their complexity, these models have limits. They are not equivalent to the human mind. In Minds, Brains, and Science, John Searle draws the same conclusion: computers have limits and cannot think.6 They are syntactical, not semantical. Minds are more than syntactical. To illustrate this, Searle designed the Chinese Room thought experiment: a person locked in a room follows rules to manipulate Chinese symbols, producing answers to questions that are indistinguishable from those of a native speaker, yet without understanding the semantics of the responses. This is syntax without semantics, like a model that encodes the form of words without knowing their substance. It produces sentences about grief without having grieved, and describes courage without having feared anything.
Searle names the philosophical limit; Genesis names the theological one. Language was created by God for communication, the means by which he draws His creatures into covenant relationship with himself and with one another. Poythress argues in In the Beginning Was the Word that human language images the eternal Word, the second person of the Trinity, through whom all things were made (John 1:1–3).7 We speak because God speaks. And we speak because we think. Language is the outward expression of an inward mind: a soul oriented toward truth, capable of meaning, accountable for what is said. This capacity belongs to image-bearers alone. Though the model produces language-shaped outputs, it is not a mind.
Machine, Not Imago
A machine bears the fingerprint of man in the form of innovation. God’s image is reserved for those he breathed life into. A GPS navigates without knowing where it is going. A search engine retrieves without understanding what it finds. A language model answers without knowing what was asked. The machine is a second-order creation. It is advanced technology created by man to fulfill the mandate. It is derivative.
Start with the hardware. The cloud is a metaphor for data centers: enormous, climate-controlled, power-hungry buildings full of networks of specialized processors called GPUs, drawing electricity at an industrial scale, consuming more power than small cities. Though the model seems seamless in our browser, it rests on a vast system of concrete, steel, silicon, and software. The machine’s responses are compelling enough that we stop thinking about who built it, how it was trained, what it cannot do, and who is responsible when it fails. The abstraction is so animated and compelling that the notion of machinery working in the background disappears entirely.
Next, the software. Workflows that build large language models in stages.
The first is data curation. Before the model learns anything, someone decides what it reads. Trillions of words scraped from the web, digitized books, academic papers, code repositories, and legal documents, filtered, cleaned, weighted. Every inclusion and exclusion is a choice. The design choices and biases of the people who built the dataset are embedded in the model before a single parameter is trained.
The second is pre-training. The model is trained on that corpus to predict the next token, billions of times across billions of documents, using gradient descent at a massive scale. This is where the weights are formed: the numerical parameters that encode everything the model knows: its fluency, its conceptual associations, its range of reference. The process takes months, requires thousands of GPUs running without interruption, and costs tens of millions of dollars. What comes out the other side is a base model, extraordinarily capable but raw and without direction.
The third is fine-tuning, and it is here that the moral stakes become most visible. The engineers who shape the model’s behavior decide which topics it will engage, which it will refuse, what counts as a helpful response, and whose standards of fairness become the default. These are ethical decisions. They draw on a combination of Aristotelian, Kantian, and utilitarian frameworks, applied by committee, but most likely without theological grounding. Those decisions are then embedded in a system that behaves accordingly, without conscience, without accountability, and without any awareness of what it is enforcing. For example, LLMs like ChatGPT can be used to generate disinformation at scale. Outputs appear authoritative but are entirely fabricated, because the model has no capacity for distinguishing truth from plausible-sounding text.
Then comes inference: the moment we engage the machine directly. We ask the question and hit send. The model executes. The neural network, the same one built during training, runs a forward pass on our input with its learned weights frozen in place. Every layer fires in sequence, applying its matrix multiplications and activation functions, until it produces a probability distribution over likely next tokens and returns a result. During training, the network updated its weights for each example. During inference, it does not. The weights are fixed. The network simply executes. What we see is fluency. What is happening underneath is computation, algorithms processing data, driven by a model, governed by man-made policies, indifferent to the result.
At every stage, hardware, data, training, tuning, inference, the black box bears the vestiges of man: his choices, his biases, his frameworks, his values, embedded at every stage of its construction. AI is a machine, designed and made by man, but not Imago Dei.
A Theological Perspective
Math, not magic. Model, not mind. Machine, not imago. Three concepts. Enough to understand AI with a little more clarity. Enough to resist the hype that substitutes speculation in place of understanding.
Creation has structure because the Creator is orderly. We can do mathematics because God built patterns into reality. We can build models because the world holds structure worth capturing. We can engineer machines because we were made to make things, to take the raw material of an ordered world and develop it. This is the creation mandate (Genesis 1:28). It is the Christian account that explains why science works. Common grace distributes this capacity broadly so that the engineer who does not confess Christ still discovers real patterns in a creation God holds together.
And yet, we are all fallen creatures. Our datasets reflect this. The engineers who shape the model, the companies who deploy it, the users who trust it without question, all of us bring the distortions introduced by the antithesis (Genesis 3:15) east of Eden. This is not a limitation of the machine since it has no mind or moral responsibility. It simply executes. Though the machine has no conscience, God gave man the law. The Decalogue is not a relic of ancient times. It is an ethical architecture for ordering human life, human work, and human tools toward justice, truth, and the glory of God.
Through Christ’s redemptive work, the antithesis has a resolution. Christ restores what the fall corrupted and completes what the mandate began. He is the image of the invisible God, the firstborn over all creation, in whom all things were created, things visible and invisible, and in whom all things hold together (Colossians 1:15-17). The mathematics, the models, and everything in the black box fall within His lordship. As creative beings, we continue to live and build, between the Fall and the restoration, sustained by grace, and oriented toward human flourishing until His return.
Conclusion
Christians with a Kingdom understanding of what is in the black box are called to shape the machine’s future. Every sphere of society, the academy, the church, the state, the marketplace, carries its own God-given authority and responsibility. AI does not belong solely to the tech industry. In the academy, it shapes what is taught. In the church, how it is understood. In the state, how it is regulated. In the marketplace, how it is built. Each sphere has a role, and each answers to God within its own domain. Our calling is clear. Build retrieval systems grounded in truth. Design guardrails that reflect a coherent moral framework. Write policies that name accountability clearly. And develop tools that promote flourishing. God has called us to engage, not to wait. The time to open the black box is now.
S.D.G.
Gilbert Strang, Linear Algebra and Learning from Data (Wellesley: Wellesley-Cambridge Press, 2019), 2–97.
Ronald E. Walpole, Raymond H. Myers, and Sharon L. Myers, Probability and Statistics for Engineers and Scientists, 8th ed. (Upper Saddle River: Pearson, 2007).
Gilbert Strang, Linear Algebra and Learning from Data (Wellesley: Wellesley-Cambridge Press, 2019), 344–357.
Andrew Glassner, Deep Learning: A Visual Approach (San Francisco: No Starch Press, 2021).
Ashish Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems 30 (2017).
John Searle, Minds, Brains, and Science (Cambridge: Harvard University Press, 1984), 28–41.
Vern S. Poythress, In the Beginning Was the Word (Wheaton: Crossway, 2009), 23–25.