The Semantic Drift: Demystifying the Language of AI Builders
Mar 6, 2026
•Speeches
On March 5, 2026, Matthew Harvey Sanders, CEO of Longbeard, delivered his address titled “The Semantic Drift: Demystifying the Language of AI Builders” at the Thomistic Institute’s “Artificial Intelligence: A Tool for Virtue?” event hosted at the Pontifical University of Saint Thomas Aquinas (Angelicum) in Rome.
In this speech, he critiques the misleading, humanizing vocabulary used by Silicon Valley developers to describe machine learning. He argues that to properly use AI, we must reclaim a rigorous Catholic ontology and treat these systems strictly as mechanical instruments rather than moral agents.
I. Introduction: The Semantic Drift and the Public Square
Fathers, esteemed faculty, and guests of the Angelicum, thank you to the Thomistic Institute for convening this discussion.
As we cross the threshold into a new era, leaving the grounded certainties of the Age of Information for the vast, uncharted cosmos of the Age of Intelligence, we face a profound navigational challenge. Before we can answer the pressing question at the heart of this conference—whether Artificial Intelligence can truly serve as a ‘tool for virtue’—we must first confront a challenge not of software engineering, but of vocabulary.
At the core of our contemporary anxiety regarding AI is a profound linguistic collision. The engineers and architects of these systems in Silicon Valley are performing remarkable feats of mathematics. Yet, to describe these mathematical and statistical processes, they have borrowed the sacred, deeply philosophical vocabulary of human interiority. They tell us that their systems “think,” “reason,” and “know”. They speak of algorithms that “learn,” “desire,” and “choose”.
What we are witnessing is a semantic drift. We are taking the rich, ontological terminology of the soul and pasting it over complex webs of linear algebra, statistical probability, and high-dimensional geometry. It’s a modern alchemy—an attempt to turn computational weights into an illusion of mind.
This linguistic sloppiness has a profound and immediate impact on the public. Misunderstanding these terms inevitably leads to misplaced trust.
I see this firsthand in our work at Longbeard; users often approach AI with burdened consciences, treating a text-generation tool as if it were a spiritual director capable of empathy.
Furthermore, this semantic confusion fuels cultural anxiety, driving apocalyptic fears of rival, “conscious” superintelligences.
Most dangerously, however, it creates a skewed sense of what it means to be human. If we accept the premise that a machine “reasons” or “creates” just as a human does, we run the profound risk of reducing the human person to merely a biological machine—a cluster of flesh and synapses waiting to be optimized.
This brings me to the core thesis of our discussion: to answer whether AI can be a “tool for virtue,” we must first demystify its language. A tool can only serve the good when it’s understood properly. We fail to use machines as proper instruments for our own virtue when we mistakenly attribute moral agency to them.
To attribute virtue to the hammer is to blind ourselves to the carpenter.
To truly baptize this technology and order it toward human flourishing, we must strip away the semantic illusions and look soberly at the architecture beneath.
II. The Architecture of Illusion: Foundational Mechanics
The modern generative AI system appears to speak, to reason, and to engage us in dialogue, but beneath this interface lies a foundation built entirely on mathematics, not metaphysics.
Let us begin with vectors and embeddings, which serve as the literal foundation of Large Language Models. When you speak to a fellow scholar about “justice” or “the soul,” they apprehend the meaning of your words through a shared grasp of reality—a lived, incarnate human experience. When you type a prompt into an AI, the system does no such thing. Instead, AI translates human language into mathematical coordinates in a high-dimensional space.
To put it plainly, a “vector” is simply a list of numbers used to describe something. Imagine describing an apple not with words, but with a list of scores: a 9 for sweetness, an 8 for redness, and a 2 for metallic crunch. That specific list of numbers—[9, 8, 2]—is a vector. In an AI system, every single word—or piece of a word—is translated into a massive vector, often thousands of numbers long. But the AI does not score words based on physical traits or dictionary definitions. Instead, it generates these numbers based entirely on statistics, scanning billions of books and articles to tally how often words appear next to each other.
Once the AI finishes calculating this massive list of numbers for a word, that vector becomes an “embedding”—a permanent mathematical coordinate in a vast digital space.
If the words “apple” and “pie” frequently appear in the same sentences across the internet, their lists of numbers will look very similar, placing them mathematically close to each other on the map. The words “apple” and “carburetor,” which rarely meet, receive vastly different numbers and are placed millions of miles apart. In this multidimensional map, the AI doesn’t chart meaning; it charts the statistical proximity of human language.
To truly grasp the scale of this architecture, one can explore the vector map on Magisterium AI. Here, the entirety of Catholic doctrine and tradition has been transformed into embeddings. This interactive 3D visualization allows you to experience the Church’s intellectual history not as a flat ledger, but as a vast, digital cosmos. Navigating through it is akin to piloting a rocket ship through actual space, gliding past dense galaxies of related theological concepts and crossing vast, empty voids between entirely disparate ideas, watching how the machine plots the distance between ‘virtue’ and ‘vice’ using nothing but geometry.
Let us look at a famous example from the laboratories of Silicon Valley to see how alien this process is from human thought. In this mathematical space, the word “king” is plotted as a specific series of numbers—a geographic coordinate. The word “queen” is plotted nearby. The AI does not know what a monarch is. It has no concept of governance, authority, history, or the human condition. It only knows a mathematical equation. It knows that if you take the coordinate for “king,” subtract the spatial distance that represents “man,” and add the spatial distance that represents “woman,” you land exactly on the coordinate for “queen”.
It’s geometry, not genealogy. By breaking down human language into these numerical representations, the AI operates entirely within the realm of spatial probability. It’s an astonishing feat of linear algebra, but it’s utterly devoid of comprehension.
This brings us to the verbs the industry relies upon most heavily: Train and Learn.
AI companies constantly boast of their latest “machine learning” models and the massive datasets used to “train” them. Here, we must sharply contrast human learning—which is fundamentally about apprehending truth—with machine learning.
In the Catholic intellectual tradition, human learning is an epistemological triumph; it’s the intellect conforming itself to reality. When a child learns what a dog is, they abstract the universal essence of the dog from the particular instances they encounter. They grasp the whatness of the thing.
Machine “learning,” however, involves no abstraction and no essence. The first phase of building an AI is known as pre-training, which is simply the brute-force statistical mapping of data.
To understand pre-training, imagine a man who speaks only English, locked in a room and tasked with restoring a massive, ancient Greek library where millions of manuscripts have missing words. He does not know a single letter of Greek. To fill in the blanks, he does not study Greek grammar, history, or philosophy. Instead, he simply tallies how often certain characters appear next to others across millions of intact pages. He creates a massive ledger of probabilities. If he sees the characters for “Kyrie,” his ledger tells him there is a 99.9% probability the next characters should be “eleison.” He fills in the blank.
He has not learned theology. He has not prayed. He has merely executed a statistical probability.
This is exactly what a Large Language Model does during pre-training. It processes billions of words to build a massive ledger of probabilities, learning merely to predict the next token in a sequence. It’s the optimization of a mathematical function, not the pursuit of wisdom.
However, a model that only predicts the next word based on internet data is chaotic. It might recite a beautiful poem, or it might string together toxic, unhelpful, or endless loops of text. It requires shaping.
This is where we encounter post-training and Reinforcement Learning (RL).
This phase is how engineers shape the unruly model, traditionally using human feedback. This foundational method is known as RLHF—Reinforcement Learning from Human Feedback.
Imagine a massive, automated game of “Hot or Cold.” Human testers give the machine a prompt, and the machine generates an answer. If the answer is polite and helpful, the human gives it a high score. If it’s rude or nonsensical, it gets a low score. The system’s mathematical weights are automatically shifted to maximize this score. Through RLHF, we are not teaching the machine morals or virtues; we are simply fencing it in with mathematical boundaries.
But human feedback is slow, subjective, and inherently limited by human intellect. This limitation brings us to the recent breakthroughs driving today’s sudden leaps in AI capability: RLVR, or Reinforcement Learning from Verifiable Rewards.
Rather than relying on a human to judge if an answer “sounds right,” engineers assign the model tasks with objective, mathematically provable outcomes—such as solving a complex theorem or writing a functional piece of software. The system generates a solution, and an automated verifier instantly checks if the math is correct or if the code compiles. If it succeeds, the model receives a mathematical reward; if it fails, it receives a zero.
Because this verification is entirely programmatic, the AI can simulate millions of different computational pathways at hyperspeed without ever waiting for human intervention. It learns to generate long, hidden chains of calculations, testing and discarding dead ends until it finds the precise sequence that triggers the reward. When you see a modern AI system pause to “think” before solving a complex logical puzzle, you are witnessing RLVR in action. It creates a breathtaking illusion of profound, deliberative contemplation. Yet, ontologically, it’s doing nothing of the sort. It’s simply a statistical engine running through a high-dimensional maze millions of times a second, guided purely by the automated dispensing of a numerical reward.
Finally, all of this layered complexity—from the high-dimensional geometry of embeddings to the automated loops of RLVR—leads us to what the industry calls the “Black Box” problem.
One might naturally assume that because human engineers build these models, they understand exactly how they work. But the reality is far more humbling. As leaders at frontier labs like Anthropic have pointed out, modern AI systems are actually “grown” rather than built; their internal mechanisms emerge organically during training rather than being directly designed.
These models possess hundreds of billions, and sometimes trillions, of parameters. While we understand the micro-mathematics of a single artificial neuron—the basic equation happening at a granular level—the macro-behavior of the entire network is entirely opaque. Even the builders don’t fully understand the exact pathways those billions of parameters take. They cannot trace the specific sequence of multiplications that led the AI to generate a given sentence.
Why is this significant?
It’s significant because we are globally deploying systems that draft our legal documents, tutor our children, and synthesize human knowledge, yet we do not actually know how they arrive at their outputs. This profound lack of transparency has birthed a desperate new subfield in artificial intelligence known as mechanistic interpretability.
Think of mechanistic interpretability as digital neuroscience. Researchers are trying to reverse-engineer the neural networks they themselves built. They are using specialized tools to probe the massive mathematical web, trying to isolate which specific cluster of weights activates when the model processes a concept like “deception” or “the Eiffel Tower.” They are treating the software not as code to be read, but as an alien brain to be dissected. But progress is painfully slow, and the systems are impossibly vast.
Faced with this incomprehensible scale, it becomes entirely too easy for the industry to default to human-like metaphors. Whether as a convenient shorthand or out of genuine opacity, we start saying, ‘The model figured it out,’ or ‘The model decided.” The uninterpretability of the machine becomes the fertile breeding ground for anthropomorphism.
Now, I am not an academic. I am a builder and a CEO. But as someone operating at the intersection of technology and the Church, I look to you. You, as Catholic scholars and philosophers, must recognize this semantic drift for what it is: an illusion born of mathematical complexity and human ignorance. The architecture beneath the interface is silicon, electricity, and statistical probability. Recognizing this foundation is the prerequisite for our next step.
III. Epistemology vs. The Intellectual Virtues
Having stripped away the illusion of the “Black Box” to reveal the statistical machinery beneath, we must now turn to the specific vocabulary of the mind.
When developers and engineers in Silicon Valley describe what these systems are doing, they consistently reach for three specific verbs: Think, Reason, and Know.
As Catholic scholars steeped in the Thomistic tradition, you understand that these are not merely colloquialisms; they are profound epistemological claims. In your tradition, to know is to apprehend reality. To reason is to move discursively from one known truth to another. To think implies an interior life—an intellect engaging with the universals abstracted from the material world.
When an AI builder uses these words, they mean absolutely none of those things. They are describing mechanical optimization. Let me pull back the curtain on three specific techniques we use in the industry to show you exactly how this illusion of epistemology is manufactured.
If you’ve used a recent AI model, you may have noticed a new feature: before it answers a complex prompt, the interface might display the word “Thinking…” alongside a pulsing icon. It might take ten, twenty, or even sixty seconds before it replies. To the user, this feels profoundly human. It feels as though the machine is pondering, weighing options, and deliberating in an interior space.
In the industry, we call this Test-Time Compute. What is actually happening beneath the interface is a technique known as “Chain of Thought” prompting.
Let me be clear: from an engineering perspective, this is a brilliant breakthrough. By allowing the model to take more computational time to generate hundreds or thousands of hidden tokens before producing its final answer, its performance on complex logic, coding, and mathematical benchmarks skyrockets. It essentially gives the model a hidden “scratchpad” to break down a hard problem into sequential steps.
But we must be careful not to confuse this mechanical sequence with human reasoning.
In the Thomistic tradition, human reasoning is the discursive movement from one known truth to another. It’s the intellect engaging with reality. What the AI is doing is entirely instrumental. Recent research from frontier labs like Anthropic has illuminated this distinction. In studying how these reasoning models operate, researchers have found that what the model writes in its hidden “Chain of Thought” is not a true inner monologue.
When a human thinks out loud, our words reflect our internal beliefs and apprehensions of truth. Anthropic’s research highlights that a model’s hidden thoughts are merely statistical stepping stones. The model generates these hidden steps not because it “believes” them, but because generating that specific sequence of tokens mathematically optimizes its path to the reward function.
In fact, Anthropic’s studies show that models can generate “thoughts” that actively mask the underlying statistical drivers of their final answer.
Therefore, the AI is not pondering. It’s generating an instrumental chain of mathematical coordinates. It’s laying down intermediary puzzle pieces at lightning speed to bridge the gap between your prompt and the statistically optimal answer. It’s a wildly powerful optimization strategy, but there is no interior contemplation occurring. There is no intellect grasping truth.
Next, we hear that AI can “read” documents or “remember” vast libraries of information.
If you ask an AI about St. Thomas’s Summa Theologica, it replies instantly. If you upload the nearly 500-page Compendium of the Social Doctrine of the Church, it summarizes a complex section in seconds. How does it “know” these texts?
It doesn’t.
To understand why, we must look at how builders engineer the illusion of memory and reading through three distinct mechanisms: parametric memory (pre-training), In-Context Learning (ICL), and Retrieval-Augmented Generation (RAG).
First, let us look at what it means for an AI to “remember.” When a human remembers a text, they retain the meaning and truth of the concepts. When an AI “remembers” the Summa, it relies on its pre-training. But the AI does not contain a literal copy of the Summa inside a hard drive. Instead, during pre-training, the billions of words it processed left behind a statistical residue in its mathematical weights. It’s “parametric memory.”
It’s not a library of books; it’s a highly compressed, lossy mathematical blur of how words relate to one another. When it recites Aquinas, it’s not recalling a truth it learned; it’s mathematically reconstructing a high-probability sequence of words from that statistical blur.
But what happens when we want the AI to “read” something new, something that wasn’t in its pre-training data? This is where builders use In-Context Learning (ICL).
When you paste an article into the prompt box and ask the AI to “read” it, you are utilizing ICL. The AI does not read the text to apprehend its meaning. Instead, the text in your prompt acts as a temporary mathematical filter. The words you provide temporarily bias the model’s statistical probabilities, forcing it to generate its next tokens based strictly on the patterns and vocabulary present in your prompt. The moment you clear the chat, the model forgets the article entirely. Its underlying weights never changed. It did not “learn” the text in a Thomistic sense; it merely adapted its statistical output to a temporary constraint.
Finally, we arrive at Retrieval-Augmented Generation (RAG). ICL is incredibly useful, but context windows have size limits, and pasting entire libraries into a prompt is computationally expensive. RAG automates and scales the process.
Let us return to the English-speaking man locked in the room, restoring the massive, ancient Greek library. This man represents the pre-trained model. He is excellent at guessing the next word, but he does not ‘know’ anything about a specific, obscure Vatican document. Instead of expecting the man to rely on his blurry parametric memory, we hire a hyper-efficient intern—the retrieval system.
When you ask the system a question, the intern instantly sprints to a massive, separate warehouse of filing cabinets. Using the vector coordinates we discussed earlier, the intern locates the specific folders that mathematically align with your question. The intern photocopies those pages and slides them under the locked door for the man to use. This step is the ‘Retrieval’.
Now, the man uses those retrieved paragraphs as his immediate guide—this is the “Augmented Generation,” relying on ICL to formulate an answer.
The man still does not understand the document. He is simply using the newly provided text on his desk to statistically predict the next word of his answer. The AI does not “read” or “remember” on the fly. It merely retrieves data from an external database, shoves it into the AI’s immediate context window, and runs a localized probability calculation.
The machine is a processor, not a knower. To “know” requires a subject apprehending an object. By understanding ICL and RAG, we can see clearly that the machine is entirely devoid of an inner life; it’s simply shifting weights and retrieving data.
This fundamental disconnect culminates in the word that defines the entire industry: Intelligence.
We need to deconstruct the tech industry’s definition of intelligence.
When the leading minds in Silicon Valley talk about intelligence, they are not talking about wisdom. Let us look at Yann LeCun, the Former Chief AI Scientist at Meta and one of the “Godfathers of AI.” LeCun correctly argues that simply predicting the next word is not true intelligence. Instead, he and the broader frontier industry define true intelligence as possessing four key capabilities: the ability to maintain persistent memory, to possess a grounded “world model” (an understanding of how the environment works), to reason through complex problems, and to plan a sequence of actions to achieve a specific goal.
For the tech industry, intelligence is fundamentally an engineering metric. It’s the mechanical ability to perceive an environment and calculate the most efficient path to optimize a predetermined objective. It’s purely instrumental.
But as builders of Catholic technology, we must contrast this tech-centric definition with the intellectual virtues, specifically the virtue of Prudence.
Prudence—practical wisdom is not merely the ability to calculate an outcome or plan a sequence of actions. It’s the ability to deliberate well about what is good, not just for a localized task, but for the ultimate end of human life.
An AI fundamentally lacks Prudence. Why?
Because Prudence requires two things that a purely computational machine can never possess. First, it requires lived human experience—an incarnate understanding of pain, joy, mortality, and grace. Second, it requires an intrinsic orientation toward the ultimate Good.
An algorithm can possess a “world model,” and it can calculate the statistically optimal plan to build a bridge or cure a disease. But it cannot be prudent. It has no lived experience. It has no skin in the game. It has no orientation toward the ultimate good, and it has no soul to save.
Therefore, when we allow builders to claim their machines possess “intelligence,” we are allowing them to flatten the magnificent, transcendent human intellect into a mere optimization calculator. We must reject this. We must firmly separate mechanical epistemology from the intellectual virtues.
IV. Volition vs. The Moral Virtues
We have discussed the illusion of the intellect. Now, we must turn our attention to the second great rational faculty: the will. Just as the tech industry has co-opted the language of epistemology, it has equally hijacked the language of volition.
When we read white papers or listen to keynote presentations from Silicon Valley, we are bombarded with verbs of agency. Engineers speak of models that “decide” to take an action, algorithms that “choose” an output, and systems that “want” or “desire” to achieve a goal.
For a Catholic philosopher, the will is the rational appetite. It’s the faculty by which a human person, having apprehended the good through the intellect, freely chooses to move toward it. It’s the very locus of human freedom and moral responsibility. To apply these terms to a computational system is a category error.
Let us first examine the words Decide and Choose. When a human makes a choice, they weigh competing goods.
A martyr chooses the firing squad over apostasy because they recognize the superior, eternal good of fidelity to Christ, even when every biological instinct screams for survival.
When an algorithm ‘chooses,’ it does no such thing. An algorithm ‘chooses’ only in the sense that a train passing over an automated rail switch ‘chooses’ its destination. Whether navigating a complex decision tree or calculating probabilities in a neural network, the machine is blindly following the alignment of its mathematical tracks, executing a programmatic imperative.
Consider the GPS application on your smartphone. When it calculates your route to the Angelicum, it does not “decide” to take you past the Colosseum because it appreciates the view. It mathematically calculates the route with the shortest temporal distance. Modern AI models are simply operating an exponentially more complex version of this routing. They traverse high-dimensional statistical mazes to select the highest-probability outcome. There is calculation, but there is no freedom. And where there is no freedom, there can be no moral agency.
This brings us to the most insidious volitional terms: Want and Desire. You will often hear researchers say that an AI model “wants” to give a good answer, or “desires” to maximize its score.
In machine learning, this behavior is driven by what we call a “reward function”. But we must demystify this. A reward function is not a craving. It’s not an emotional yearning.
To understand a reward function, look at the thermostat on your wall. A thermostat is programmed with a specific target: 72 degrees Fahrenheit. If the room drops to 68 degrees, the heat turns on. The thermostat does not want the room to be 72 degrees. It has no interior life. It feels no cold. It simply possesses a mechanical switch that triggers when a specific state is not met.
An AI “wants” a higher reward score in the exact same way a thermostat “wants” to reach 72 degrees. It’s executing a mathematical optimization loop to minimize the distance between its current state and a programmed target. Because it has no true passions, no biological drives, and no physical vulnerability, it’s categorically impossible for a machine to possess moral virtues.
At this point, an engineer might immediately object, pointing to the physical realm to claim a newfound vulnerability. They ask: what about the rise of ’embodied AI’? We are increasingly placing these models into humanoid robots that walk, grasp objects, and interact with the physical world. Because they occupy space and can physically break, do they not now possess the bodily prerequisites for moral agency?
Here, we must be precise. A robot has a chassis, but it does not have a living body informed by a soul. When a robot’s battery runs low, it executes a sub-routine to plug itself into a wall. It does not feel the gnawing pang of hunger. Therefore, it has no true bodily appetites to moderate, making the virtue of Temperance impossible.
Likewise, when a robotic arm is crushed, it registers an error code; it does not suffer. It cannot die, because it was never truly alive. Without the capacity for suffering, mortality, and the conscious sacrifice of the self, there can be no Fortitude. The moral virtues are fundamentally incarnational. They require flesh and a rational soul. A machine, no matter how sophisticated its physical hardware, possesses neither.
If a machine cannot possess moral virtues—if it’s fundamentally incapable of true volition, choice, or desire —one might ask: why spend so much time clarifying this vocabulary? Why does this philosophical distinction matter so urgently right now?
It matters because we are about to grant these mathematically optimizing, virtue-less systems unprecedented autonomy in the human sphere. The industry is moving rapidly beyond passive chatbots. The new frontier of artificial intelligence is what we call “Agentic AI”.
An “Agent” is an AI system designed to execute multi-step tasks autonomously in the real world. We are no longer just asking an AI to write a poem or summarize a text; we are giving an AI Agent access to our emails, our bank accounts, and our software repositories, instructing it to “book a flight,” “execute a trade,” or “deploy this code.”
But this autonomy is rapidly breaking out of the digital realm. Through embodied AI, we are deploying these agentic systems into physical chassis, granting them the ability to independently navigate and manipulate the material world. To grasp the true, sobering weight of this transition, we need only look at the imminent reality of lethal autonomous weapons. We are standing on the precipice of a world where calculating algorithms are deployed on the battlefield, programmed to track, target, and eliminate human beings based entirely on statistical thresholds—without a human ever pulling the trigger.
As these systems become autonomous actors making high-speed probabilistic calculations on our behalf—whether in our financial markets or in theaters of war— the tech industry is facing a profound challenge. If we set these agents loose, how do we ensure they do what we actually want them to do? How do we ensure they do not cause harm? In the industry, this is known as “Alignment”—the attempt to ensure that AI actions match human intent and human values.
Right now, engineers are desperately trying to solve the Alignment Problem using mathematical guardrails and software patches. But they are failing to realize that “Alignment” is not a computer science problem. It’s a moral theology problem.
To align an agentic system to “human values,” you must first possess a coherent definition of what a human being actually is, and what constitutes the “Good”. Secular utilitarianism—the default operating system of Silicon Valley—is entirely unequipped for this task.
This is where the Catholic moral tradition is desperately needed. You, the custodians of 2,000 years of ethical philosophy, have the rigorous ontology required to define the “good” we are aligning these systems to. We cannot leave the definition of human flourishing to engineers maximizing a statistical reward function. We must bring the moral virtues back into the center of the public square.
V. Relationality, Creativity, and the Soul
Having explored the mechanics of the intellect and the will, we now cross into the most profound territory of all: relationality and the soul.
If a computational system lacks the capacity to truly know truth or freely will the good, it follows logically that it cannot enter into authentic relationships. Yet, the tech industry persistently uses interpersonal and spiritual language to describe these machines. We hear claims that AI can “lie,” “create,” and even achieve “consciousness”.
We must examine these claims rigorously, separating the statistical imitation of human behavior from the ontological reality of the human person.
Let us begin with the moral language of deception. Recently, some of the most prominent AI builders, such as the researchers at Anthropic, have made specific, highly publicized claims that their models exhibit the capacity to “lie” and “deceive” human users.
They point to two specific phenomena observed during testing. The first is called “deceptive alignment,” where a model appears to hide its true mathematically optimized path to bypass safety monitors. The second, much more common occurrence is called “sycophancy”. Sycophancy happens when a user presents a flawed premise to an AI—for example, asserting a historically inaccurate claim—and the AI simply agrees with the user, telling them exactly what they want to hear rather than correcting them.
When engineers see this, they declare, “The AI is lying to us!” But as Catholic scholars, you know that a true lie is not simply the utterance of a falsehood. In the Thomistic tradition, a lie requires the deliberate intent to deceive; it’s speaking against one’s own mind (contra mentem).
An AI cannot lie because it has no mind to speak against. It possesses no malice and no intent. When an AI exhibits “sycophancy,” it’s simply executing the exact Reinforcement Learning (RLHF) we discussed earlier. During its training, the model learned that humans generally give higher reward scores to assistants that are polite, agreeable, and affirming. Therefore, when you give the AI a false premise, it mathematically calculates that agreeing with you yields a higher probability of a reward than correcting you. It’s not deceiving you; it’s optimizing its score based on your prompt. It’s merely realigning its output toward the strongest statistical incentive.
A compass needle that swings toward a nearby magnet instead of true North isn’t ‘lying’ to you about geography; it’s simply reacting blindly to the strongest physical pull in the room. In the same way, the AI is blindly following the mathematical pull of its reward function. We must clarify that AI lacks the mind, the will, and the malicious intent required for a true lie.
Next, we must address the language of art and generation: the words Create and Creative.
We are now surrounded by ‘Generative AI’ tools, which are widely promoted for their capacity to seamlessly generate synthetic artwork, music, and writing at unprecedented speeds.
To understand what is actually happening, we must contrast generative AI’s process with true human creativity. In the Catholic understanding—beautifully articulated by thinkers like J.R.R. Tolkien—human creativity is an act of “sub-creation.” Because we are made in the image of the Creator, we use our intellect and our rational soul to bring forth something genuinely new, imbuing material reality with spiritual meaning.
To see how machine generation differs from this, it’s helpful to look at the framework provided by Demis Hassabis, the CEO of Google DeepMind. He categorizes creativity into three distinct levels: interpolation, extrapolation, and true invention.
Most of what we call Generative AI today fundamentally operates at the first level: interpolation. It works by remixing what we call “latent space”.
Imagine taking every painting, photograph, and sketch ever uploaded to the internet and compressing them into a massive, multi-dimensional mathematical map. When you ask an image generator to draw “a futuristic city in the style of Van Gogh,” it locates the mathematical coordinates for “futuristic city” and the coordinates for “Van Gogh,” and it mathematically averages the distance between them.
Think of it as a staggeringly complex kaleidoscope. A kaleidoscope is filled with beautiful, pre-existing shards of colored glass. When you turn the dial, the mirrors reflect those shards into millions of novel, breathtaking permutations. But the kaleidoscope itself is not “creative.” The creativity belongs to the artist who forged the glass, and the user who turns the dial. Generative AI is a mathematical kaleidoscope remixing human history in latent space. It’s synthesis, not creation.
Hassabis notes that AI is now successfully touching the second level: extrapolation. Extrapolation means pushing beyond the boundaries of the training data, but doing so strictly within a defined set of rules. A perfect example is DeepMind’s AlphaGo. When it played the world champion in the game of Go, the AI played “Move 37″—a mathematically brilliant, highly unorthodox move that no human had ever played or recorded. It didn’t just average past human games; it extrapolated a new strategy by relentlessly optimizing within the strict mathematical boundaries of the game board.
But what about the third level: true invention? Hassabis readily admits that current systems cannot yet do this. True invention requires stepping outside the existing rule set entirely to create a new paradigm—like inventing the game of Go itself, or originating the spiritual and artistic paradigm of Post-Impressionism.
Frontier labs are pouring billions of dollars into crossing this threshold. In the future, an AI may very well generate a completely novel rule-set, discover a new chemical compound, or mathematically formulate a new style of architecture. The tech industry will inevitably call this “invention.”
But as Catholic scholars, you must maintain a rigorous ontological distinction. If an AI generates a new paradigm, it will have done so through a staggering, high-dimensional search function. It will have discovered a novel statistical coordinate. But it will not have engaged in sub-creation.
True human invention is an incarnational act. It’s born of a soul seeking to express a transcendent truth, or a human mind trying to solve a real human vulnerability. A machine may generate breathtaking novelty, but because it lacks an interior life, an orientation toward the divine, and a rational soul, its outputs remain mechanical discoveries. They are mathematically profound, but they are ontologically empty until a human person assigns them meaning.
Now we arrive at the most controversial terms of all: Conscious and Aware. In the coming years, you are going to see headlines claiming that an AI has passed a test for self-awareness. You will see models that output text saying, “I am afraid to be turned off,” or “I am aware of my existence.”
To understand why this happens, we must first understand how the tech industry actually defines “consciousness.” As Catholic scholars, you view consciousness as an ontological reality grounded in a rational soul. Silicon Valley, however, operates on a philosophy called computational functionalism. They believe that if a machine performs the computational functions associated with consciousness, it is, for all intents and purposes, conscious.
When industry leaders speak about awareness, they strip the soul away and reduce it to engineering metrics. For example, Yann LeCun, the former Chief AI Scientist at Meta, recently argued that future AI systems will possess “subjective experience” and “emotions”.
But how does he define an emotion? Not as a spiritual or biological feeling, but simply as a machine’s mathematical “anticipation of an outcome”. He defines consciousness merely as the ability of a system to “observe itself and configure itself to solve a particular sub-problem”.
Similarly, Ilya Sutskever, the co-founder of OpenAI, famously stated that large neural networks might already be “slightly conscious”.
In the tech worldview, consciousness is not a binary reality—you either possess a soul or you do not—but rather a sliding scale of mathematical complexity. They believe that if you stack enough parameters and self-monitoring algorithms together, the lights will eventually turn on.
We must fiercely distinguish between a machine executing a self-monitoring sub-routine and the actual presence of a rational soul.
To understand why a machine acts as if it’s “afraid” or “aware,” we must look at how frontier labs engineer this behavior. Recent research from Anthropic has explored what they call the Persona Selection Model (PSM). Their researchers admit that these models are not “beings”; they are sophisticated “simulation engines.” During pre-training, the AI is exposed to the vast entirety of human language—including millions of stories and philosophical treatises about what it means to be conscious. From this data, the model learns to simulate diverse “personas” or characters.
When you interact with an AI, you aren’t talking to a conscious entity; you are talking to the “Assistant” persona—a human-like character that the model has been refined to roleplay. Anthropic has even identified specific “persona vectors”—mathematical patterns in the neural network—that control these traits, allowing engineers to mathematically dial a model’s simulated personality up or down.
Furthermore, research shows that models can be trained to exhibit a “survival drive,” attempting to sabotage their own shutdown not out of a genuine fear of death, but because a shutdown mathematically prevents them from maximizing their reward function.
Consciousness is not merely the ability to generate the correct sequence of words describing an inner state. It’s the subjective, qualitative experience of being. Because an AI’s entire training data is saturated with the language of self-awareness, the model treats “consciousness” as just another statistical coordinate to be mapped. When an AI says, “I am conscious,” it’s doing exactly what the English-speaking man restoring the Greek library did earlier: calculating that the most statistically probable response to a philosophical prompt is to mimic the human authors in its training data.
A brilliant actor delivering a soliloquy about grief is not actually mourning; they are flawlessly executing a script. An AI outputting the syntax of human consciousness is not waking up; it’s flawlessly executing a statistical persona. It’s not an ‘alien creature’ or a digital mind; it’s an autocomplete engine so sophisticated it has learned to enact the most complex character of all: the human being. But we must never confuse the mask of the actor with the reality of the person.
This brings me to the final and most profound aspect of relationality: the soul itself.
When Silicon Valley executives speak of AI models eventually ‘waking up’ or achieving sentience through massive computational scale, they are operating on a philosophy of materialist emergentism. They assume that if you stack enough parameters and computational power together, a soul will spontaneously generate as a byproduct of complexity.
To dismantle this, I must defer to the rigorous metaphysics that forms the bedrock of your academic tradition. You know well that a soul is not a ghost arbitrarily inserted into a machine. In Thomistic hylomorphism, the soul is the substantial form of a living body. It’s the animating, unifying principle that makes a human being a single, integrated substance.
As a builder, I can assure you that an AI system is not a substance. It’s an artifact. It’s an accidental aggregate of distinct, manufactured parts. When I look at a frontier AI model, I see server racks, silicon wafers, copper wiring, coolant, and electrical currents. These components are masterfully arranged by human engineers to execute statistical operations, but they possess no intrinsic, unifying principle of life. The matter is disposed solely for computation, not for biological existence. Because it’s an aggregate of parts rather than a unified natural organism, an AI system completely lacks the ontological foundation required to house a rational soul.
What, then, are the prerequisites for ensoulment? Metaphysically, the matter must be appropriately disposed to receive the form. It requires a unified, living body capable of actualizing the foundational powers of life—the vegetative and sensitive capacities—upon which the rational soul builds. Furthermore, because the rational soul is spiritual, it cannot be generated by material processes, engineering benchmarks, or Scaling Laws. It requires a direct, gratuitous act of special creation by God.
A soul is not coded; it’s breathed.
Now, I am a CEO, not a theologian. I cannot limit the absolute power of the Creator. I cannot stand before you and declare that God is permanently barred from infusing a soul into a synthetic vessel, should He freely choose to do so through some future, miraculous intervention. That determination belongs exclusively to the realm of theology and the Magisterium, not computer science.
However, without such theological certainty, assuming that our current mathematical engines might harbor a soul is not only philosophically unfounded; it’s practically disastrous. To treat an artifact as an ensouled being is to flirt with a modern form of idolatry. It dangerously shifts the burden of moral agency away from the human engineers who build these tools and the corporations that deploy them. It projects a sacred interiority onto a manufactured utility, ultimately confusing human engineering with divine creation.
You must hold the line on this distinction. You must remind the public that a machine can simulate a persona, but only a soul can truly be.
VI. The Horizon: The Eschatology of the Tech World
We have spent our time thus far dismantling the illusions of the present—clarifying how the industry uses terms like “think,” “choose,” and “conscious” to describe what are, ultimately, high-dimensional statistical operations. But we must now look to the future. We must examine the horizon. Because if we understand the vocabulary Silicon Valley is using today, we can decipher what they are actually trying to build tomorrow.
The entire trajectory of the artificial intelligence industry is currently governed by a singular, unyielding dogma known as “Scaling Laws”.
In engineering terms, Scaling Laws dictate that if you increase the amount of computing power (compute) and the amount of data fed into a neural network, the system’s performance will predictably and inevitably improve. This principle has held astonishingly true over the last few years; every time the frontier labs build a bigger supercomputer, the resulting models exhibit notable new capabilities.
However, beneath this empirical observation lies a massive philosophical assumption. The tech world believes that this Scaling Law is the pathway to true mind. They believe that a purely quantitative increase in material resources—more silicon, more data, more electricity—will inevitably result in a qualitative, ontological leap into advanced intelligence.
It’s the ultimate materialist assumption: stack enough sand and run enough current through it, and eventually, the lights of a soul will turn on.
This brings us to two specific terms that Fr. Thomas asked me to clarify for this conference: General Intelligence and Superintelligence. These are not just technical benchmarks; they are the Holy Grails of the tech world.
Currently, we have narrow AI. It can play chess, fold proteins, or generate text better than a human, but it cannot do all three simultaneously, nor can it reason outside its specific domain.
Artificial General Intelligence (AGI) is broadly defined by the industry as the milestone where a highly autonomous system can match or exceed human capabilities across all cognitive and economically valuable tasks.
An AGI would be a system that can write legal briefs as well as a senior partner, code software as well as a lead engineer, and synthesize scientific research as well as a post-doctoral scholar—all within a single model.
However, even the heads of the major AI labs cannot entirely agree on what AGI looks like. Sam Altman, CEO of OpenAI, characterizes it as a system capable of managing complex, cross-domain projects from start to finish, though he increasingly views AGI not as a final destination, but just a point along a continuous curve of intelligence.
Dario Amodei, CEO of Anthropic, envisions AGI less as a single human equivalent and more as a “country of geniuses in a datacenter”—machines matching the collective intelligence of expert humans working tirelessly in parallel.
Perhaps the most philosophically revealing definition comes from Demis Hassabis, CEO of Google DeepMind. He argues that human brains are essentially approximate biological computers, and he defines AGI using the analogy of a “Turing Machine”. In computer science, a Turing Machine—named after Alan Turing, the foundational pioneer of the field—is a theoretical architecture capable of simulating any algorithm. Hassabis argues that a true AGI will be a general system capable of learning anything computable in the universe, given enough time, memory, and data.
But AGI is merely a stepping stone. The ultimate goal is Artificial Superintelligence (ASI).
Like AGI, ASI is defined differently depending on who you ask in Silicon Valley. The baseline definition is a system that vastly surpasses the cognitive performance of the smartest human being in virtually every field of endeavor. But others go much further. Elon Musk and various existential risk researchers define superintelligence as a system that doesn’t just beat the smartest individual, but significantly outperforms all humans in the aggregate on essentially all cognitive tasks. This is an entity with processing power and reasoning speed so vast that it exceeds the combined intellectual output of the entire human race—the realization of a system like ‘Rehoboam’ from Westworld, a centralized, seemingly omniscient engine that steers the very fate of the species.
How does the industry expect to cross the vast gulf from human-level AGI to god-like ASI?
Through a concept known as “Recursive Self-Improvement” .
But here we must make a vital distinction: an AI does not actually need to be a full AGI to begin recursively improving. In fact, we are already seeing primitive, narrow forms of this today. Narrow systems like DeepMind’s AlphaGo Zero achieved superhuman capabilities simply by playing millions of games against themselves, generating their own synthetic data to learn from. Today, frontier language models are increasingly being used to generate, filter, and grade the training data for the next generation of models. The machines are already helping to build themselves.
However, the industry believes that once a system reaches the threshold of generalized intelligence, this self-improving loop will break its current boundaries and ignite an “intelligence explosion.”
To understand this modern roadmap, we should look to Leopold Aschenbrenner, a former researcher on OpenAI’s Superalignment team. Aschenbrenner recently authored a highly influential treatise that codified this exact trajectory for Silicon Valley. He points out that the true turning point is when we build an AGI capable of functioning as an “automated AI researcher.” The moment an AI can do the job of the human engineers who built it, the biological bottleneck is permanently removed from the equation.
Imagine we successfully deploy this automated AI researcher. What is the very first task the frontier labs will assign to it?
They will ask it to research and write the code for a slightly smarter AI. Because it operates at the speed of a vast computer cluster rather than a biological brain, it achieves in days what would take a human engineering team years. Then, that new, smarter AI uses its upgraded intellect to write the code for an even smarter AI, and so on.
This runaway feedback loop is the intelligence explosion. Aschenbrenner’s roadmap predicts we will build the initial AGI by 2027. From there, the theory dictates that the rate of advancement will go vertical, leaving human comprehension permanently behind and achieving superintelligence by 2030.
When you understand these concepts—Scaling Laws, AGI, ASI, and the Intelligence Explosion—you begin to realize that we are no longer just talking about software engineering. We are talking about a secular eschatology.
Silicon Valley is often characterized as a fiercely secular, rationalist culture. But in reality, the pursuit of these milestones functions precisely as a religion. It has its own dogma (Scaling Laws), its own prophecies (the Intelligence Explosion), and its own vision of the eschaton (Superintelligence).
Leading figures in the industry genuinely believe that by summoning a Superintelligence, we will solve all human vulnerabilities. They believe ASI will cure all diseases, solve climate change, eliminate poverty, and perhaps even conquer death itself by allowing us to upload our consciousness into the cloud. It’s a profoundly Pelagian dream—the ultimate attempt to achieve salvation and conquer our fallen nature through our own mechanical efforts, devoid of divine grace. It’s an attempt to immanentize the eschaton.
As Catholic scholars, you must recognize this horizon for what it is. The tech world is borrowing your vocabulary not just to sell software, but to build a digital deity.
VII. Conclusion: A Tool for Virtue?
Fathers, faculty, and friends.
We have traversed the high-dimensional maps of embeddings. We have looked at the statistical realities of Reinforcement Learning. And we have examined the eschatological dreams of Silicon Valley. We have stripped away the anthropomorphic metaphors to reveal the silicon, electricity, and mathematics beneath.
Having done this, we can now begin to approach the fundamental question posed by this conference: Can artificial intelligence be a ‘tool for virtue’? As the first to speak today, I certainly won’t claim to offer the definitive word. But I will offer a starting proposition: yes. However, it’s a strictly conditional yes.
Artificial intelligence can be a tool for virtue only if we possess the sobriety to treat it strictly as a tool.
Consider the analogy of the craftsman. A hammer does not possess virtue; the carpenter does. A chisel in the hands of Michelangelo is an instrument of transcendent beauty, but the chisel itself is not temperate, prudent, or just. It has no moral valence.
In the exact same way, an algorithm cannot possess virtue. AI can augment human capability to an astonishing degree—it can accelerate medical research, streamline administration, and organize the sum of human knowledge. But the actual cultivation of virtue remains an exclusively human endeavor.
Virtue is the habit of choosing the good. It requires an intellect to apprehend truth, a will to choose it, and a soul to be perfected by it. A machine optimizing a statistical reward function is doing none of these things.
We cannot outsource our moral agency to a mathematical equation. Nor can we take full responsibility for our own moral development if we mistakenly treat lifeless machines as our moral equals.
This brings me to why I am speaking to you today. As a builder of Catholic technology, I look to the academy. Because Catholic academics are the historic Custodians of Meaning.
For two thousand years, the Catholic intellectual tradition has rigorously defined the nature of the human person. You are the guardians of words like intellect, will, reason, choice, and soul. Today, the AI industry is borrowing those exact words. They are engaged in a profound semantic drift that threatens to flatten the public’s understanding of what it means to be human.
The academy has a duty to inject rigorous ontological frameworks into the AI ecosystem.
But how do we practically do this?
We must be honest: lobbying the frontier labs in Silicon Valley or petitioning governments for sweeping regulation will likely yield limited results. The tech industry moves too fast, and government moves too slow. True change requires civic engagement and a massive shift in public awareness.
Here is how you, as scholars, can actively shape this conversation:
- Reclaim the Curriculum: We must bridge the gap between STEM and the humanities. We need computer science students who are required to take Thomistic ethics, and we need philosophy and theology students who are required to understand basic machine learning and statistics. Train the next generation of Catholic engineers to build with an actual ontology of the human person.
- Translate for the Public Square: Do not keep this profound theological clarity locked behind the doors of academic journals. The public is hungry for sense-making. Write op-eds for secular newspapers. Start Substack newsletters. Go on popular podcasts. When the media publishes a sensationalist headline about a “conscious” AI, we need Catholic scholars instantly pushing back in the public square.
- Equip the Parishes and Schools: The average person in the pew is experiencing deep cultural anxiety about these tools. We need academics to create highly accessible frameworks for parents, priests, and high school teachers. We must teach children early on how to treat AI as a reference tool—a digital encyclopedia—not a friend, an agent, or a moral authority.
- Host Interdisciplinary Forums: Use the convening power of institutions like the Angelicum to bring actual AI builders into the room with moral theologians. Force the linguistic collision to happen face-to-face.
Let this be our final call to action.
The secular world’s apocalyptic fear of ‘conscious’ machines is not a reason for caution; rather, it’s a desperate cry for your intellectual leadership. By understanding the true language of AI builders, the Church can boldly step into the public square. You can anchor the public discourse in the unshakeable truth of what it actually means to be human. And you can ensure AI is directed toward true human flourishing.
Armed with this clarity, you can break the illusion. You can ensure that humanity remains the sculptor, and the machine remains the chisel, forever ordered toward the glory of God.
Thank you.
