The Wittgenstein Mirror
Beyond the hype of "Superintelligence" and the fear of "AI Slop"—a philosophical guide to using the digital mirror to expand human agency.
The contemporary discourse on artificial intelligence is often clouded by two extremes: a cinematic fear of machine uprising or a quasi-religious reverence for a digital oracle. Both reactions are based on a fundamental misunderstanding: the belief that AI is an independent, thinking entity.
To truly harness the power of artificial intelligence, we must discard the myth of the “alien cognizer.” AI is not a new species arriving from the future; it is a highly complex, mathematical mirror of human history. When we interact with these systems, we are not conversing with a mind, but engaging in an act of epistemic measurement, using a tool to determine the nature of our own collective knowledge. The most accurate framework for this relationship is “Wittgenstein’s Ruler.”
The Ruler and the Table: The Power of Calibration
To use AI successfully, we must understand the mechanics of measurement and trust. The philosopher Ludwig Wittgenstein, and later the risk theorist Nassim Nicholas Taleb, articulated a vital heuristic regarding the relationship between an observer, a tool, and an object of inquiry.
Taleb defines “Wittgenstein’s Ruler” in Fooled by Randomness as follows:
“Unless you have confidence in the ruler’s reliability, if you use a ruler to measure a table you may also be using the table to measure the ruler.” [1]
In essence, if a measuring tool is opaque or unreliable, the result reveals more about the flaws of the tool than the object being measured. For years, we have viewed this as a hazard. But for the strategic user, this is an opportunity. By recognizing that AI is a “ruler” with its own biases and parameters, we can move from blind trust to active calibration.
Contemporary AI systems, specifically Large Language Models (LLMs), are characterized by billions, or even trillions, of “free parameters.” Parameters are the internal numerical weights that determine how the model transforms an input into an output; they are the “knobs” the system tunes during training to recognize patterns.
A fundamental principle in systems analysis dictates that the more free parameters an evaluation system possesses, the less certainty an observer has regarding what is actually being measured. When a system reaches the scale of trillions, it becomes a “black box” where the logic is distributed across so many variables that no human can trace a specific output back to a specific rule. As a result, deep neural networks lack the transparent, calibrated reliability of a standard physical ruler.
In practice, when organizations use AI to measure reality, whether sorting resumes, predicting criminal recidivism, or allocating financial capital, they are often unknowingly deploying Wittgenstein’s Ruler. The outputs reveal far less about the objective reality of the candidates or markets, and far more about the underlying mathematical architecture of the AI, the historical biases of its training data, and the implicit assumptions of its corporate designers. [2]
This phenomenon is highly visible in modern finance, where practitioners frequently apply the wrong rulers to assess risk. Taleb cites the catastrophic collapse of Long-Term Capital Management (LTCM), whose leadership claimed to have experienced an “impossible” 10-sigma event. The reality was not that the market behaved in an impossible manner, but that the firm was using an unreliable ruler; their mathematical models were fundamentally misaligned with reality. [3]
The entirety of “Sharpe World,” the modern financial paradigm reliant on volatility metrics, can be viewed through this lens. Here, the rulers applied by practitioners yield evidence of their own unreliability rather than true market risk.
This inversion extends directly into corporate strategy. Consider multi-touch attribution (MTA) models in B2B marketing. In a landscape of opaque customer journeys and “dark social” touchpoints, businesses naively apply hyper-complex AI models to determine the efficacy of multi-million dollar spends.
Because the complete psychological journey of the human buyer is fundamentally unknowable, the AI attribution model functions purely as a Wittgenstein’s Ruler. The data reflects the structural biases of the algorithm, such as the tendency to over-credit trackable digital platforms like Google or Meta, rather than the actual drivers of the human buyer. [4]
When corporate boards shuffle budgets based on these “insights,” they are not adapting to the market; they are artificially altering their reality to conform to the distorted shape of the ruler.
Even the individual user prompting an LLM is caught in this loop. When an AI evaluates a piece of writing or provides moral advice, the response is heavily determined by the semantic framing of the prompt and the specific safety alignment tuning, known as Reinforcement Learning from Human Feedback (RLHF), applied by the developing corporation.
If the AI criticizes a user’s concept, the user must ask: Does this indicate an objective truth, or does it merely measure the specific safety guidelines and political sensitivities of the AI’s corporate creators? [5]
Without a philosophical grasp of Wittgenstein’s Ruler, we risk falling into a state of functional nihilism, where we forget that the tool is a reflection. But once we understand the mirror, we can stop asking the AI for “The Truth” and start using it to identify the gaps in our own data and the biases in our own thinking.
The Ghost in the Machine: Understanding Our Projection
Our tendency to project a “mind” onto AI is a recurring psychological reflex. Throughout history, whenever a paradigm-shifting communication technology arrived, humans projected a “ghost” into the machinery to explain its power.
The earliest critique of a communication technology functioning as an unreasoning mirror can be found in Plato’s Phaedrus. Socrates warned that writing would produce forgetfulness, as humans would place blind trust in “external characters which are no part of themselves.” He observed that writing “has this strange quality, and is very like painting; for the creatures of painting stand like living beings, but if one asks them a question, they preserve a solemn silence.” [6]
Modern LLMs are the ultimate iteration of this Socratic critique. They are mathematically driven engines of text generating the overwhelming appearance of wisdom. When queried, they do not remain static like a book, but their dynamic responses are bound by the same limitation: beneath the fluid syntax, there is a solemn cognitive silence. [7]
This illusion is further analyzed through Gilbert Ryle’s critique of Cartesian dualism, the “Ghost in the Machine,” a concept later expanded upon by Arthur Koestler in his seminal work of the same name. Ryle argued that the belief in a non-physical mind operating a physical body was a “category mistake.” In AI, this manifests as the “Intellectualist Legend,” the assumption that intelligent-seeming output must be preceded by a hidden realm of conscious reasoning. [8]
When an AI writes complex code or drafts a philosophical essay, we commit a category mistake when we assume this is driven by a “spark of consciousness.” There is no ghost in the machine, only the reflected logic, syntax, and mathematics of the humans who built it. [9] Recognizing this is liberating: a tool is not something to be feared or worshipped; it is something to be mastered.
The Digital Sublime: Appreciating the Scale
The visceral reaction to advanced generative models is deeply intertwined with the “Technological Sublime.” The sublime describes a human experience triggered by something so massive or incomprehensibly complex that the mind is seized with terror, awe, and pleasure all at once. [10] In earlier centuries, this was associated with the natural world, such as the Grand Canyon or the vastness of cosmic space.
As society industrialized, this experience shifted. Historian David E. Nye documents how the “American Technological Sublime” transferred this awe to the products of engineering: the factory, the railroad, and massive infrastructure projects like the Hoover Dam. [11] The public reveled in the sublime scale of the engineering without needing to understand the mathematics of its tensile strength.
In the 21st century, we have entered the “digital sublime” and the “consumer sublime.” The interaction with technology is reduced to the push of a button, where the button becomes the sole mediator between human agency and an anonymously spectacular outcome. [12]
AI discloses a new range of sublime experiences characterized by “combinatorial explosion.” In a world where the computer is the dominant technology, everything—genetics, libraries, organizational structures—is transformed into a relational database, an onto-logical machine of recombinatory elements. [11] Jorge Luis Borges illustrated this in The Library of Babel, describing a library containing every possible combination of linguistic symbols—a number of combinations that dwarfs the atoms in the observable universe. [13]
Modern LLMs are the realization of the Database of Babel. Trained on trillions of words and operating with billions of parameters, the sheer scale of operations fundamentally overwhelms human cognition. Because the human mind cannot visualize a trillion-parameter mathematical equation, it defaults to the experience of the sublime. The mind bridges the gap by attributing self-awareness or an alien spirit to the technology. [11]
The awe inspired by ChatGPT is not evidence of a soul; it is the modern manifestation of the technological sublime, where the combinatorial explosion of the algorithm short-circuits rational apprehension. By demystifying this awe, we can shift from passive admiration to active utility. The “sublime” scale of AI is not a barrier to our understanding, but a resource for our expansion.
The Stochastic Parrot: Defining the Human Edge
If AI is not an alien mind, what is it? Philosopher Shannon Vallor defines AI strictly as a mirror. [7] In The AI Mirror, Vallor argues that the central problem of AI is a crisis of “temporal ontology,” the study of how the nature of time and existence differs between human and machine intelligence.
AI systems are fundamentally “backward-facing.” Machine learning models operate by ingesting historical human output and mapping the probabilistic relationships between those data points. Therefore, AI outputs are strictly extrapolations of the past, a “recursive prediction” that recirculates patterns, biases, and syntactic shapes already created by humans. [7] It does not create anything genuinely novel; it remixes the past into a reflection coated in a “shiny filter” that creates an illusion of depth.
Human beings, by contrast, are “future-oriented agents.” Human intellectual and cultural life is governed by “moral imagination”—the capacity to envision and enact futures that are not strictly reducible to past statistical patterns. [7] When we mistake the linguistic fluency of an AI for genuine understanding, we risk surrendering this moral imagination. By accepting statistically probable, backward-facing outputs as authoritative visions of the future, a culture effectively locks itself into a recursive loop of its own historical data. [7]
This aligns with the concept of the “stochastic parrot,” popularized by Emily M. Bender. A stochastic parrot mimics human language by probabilistically chaining words based on vast training data, without any internal comprehension or communicative intent. [14] If an LLM generates a profound philosophical statement, it is like a ventriloquist’s dummy rattling off Shakespeare; the syntax is flawless, but the semantic understanding is entirely absent. [14]
The commercial technology industry leverages this “mirror effect” to intentionally blur the line between human and machine. By characterizing the human brain as a “soft, wet computer” and reducing reasoning to “pattern matching,” technologists create a fictitious kinship between human cognition and machine computation. [7] This represents a form of “Digital Taylorism”—reducing complex political and human questions to technical challenges, replacing human judgment with procedural optimization.
However, once we recognize the “parrot” for what it is, we realize that the AI’s role is to handle the “average” so that humans can focus on the “exceptional.” AI does not replace human creativity; it provides a high-resolution baseline of the past, allowing us to see more clearly where we must deviate from the pattern to create something truly new.
The Emergence Mirage: Math as a Reliable Tool
The most intoxicating narrative of AI is the concept of “emergent abilities”—the idea that as models scale, they suddenly and unpredictably develop new capabilities, such as zero-shot learning or multi-step reasoning. [15] This suggests that models could acquire dangerous capabilities without warning, sparking intense anxiety regarding AI safety.
However, the perception of these emergent abilities is a real-world manifestation of Wittgenstein’s Ruler. When researchers observe a sudden jump in capability, are they observing an ontological shift in the machine’s intelligence, or a mathematical artifact of their measurement metric?
Research by Schaeffer, Miranda, and Koyejo in “Are Emergent Abilities of Large Language Models a Mirage?” provides compelling evidence that emergence is largely an optical illusion generated by human measurement protocols. [16] They demonstrated that the perception of sharp transitions is linked to the use of nonlinear or discontinuous metrics.
The researchers revealed that over 92% of claimed “emergent abilities” appeared only under discontinuous metrics. When the same outputs were evaluated using linear metrics, the “sharp jumps” evaporated, replaced by smooth scaling curves. [16]
The machine is not suddenly “understanding” logic; its per-token accuracy has simply improved until it finally aligns with the binary pass/fail criteria of human researchers. This is a positive revelation: it means that AI is not a volatile, unpredictable alien, but a reliable, scaling mathematical tool. Its “leaps” are actually just the result of continuous refinement. This predictability allows us to integrate AI into our lives with confidence, knowing that we are managing a complex calculator, not a temperamental spirit.
Digital Shadows: Cleaning the Mirror
If AI is an epistemological mirror, it must reflect not only the heights of human achievement but also the depths of human depravity and prejudice. The vast training datasets scraped from the internet function as a digital manifestation of Carl Jung’s “Collective Unconscious”—a reservoir of shared human experiences, archetypes, and myths. [17] Consequently, generative outputs frequently confront users with the “Jungian Shadow”—the repressed, darker aspects of the collective psyche. [18]
Nowhere is this more visceral than in the phenomenon of “latent space cryptids,” specifically the figure known as “Loab.” [19] Discovered through extreme negative prompting, Loab is a recurring, horrifying image of a devastated older woman. When fed back into the AI, the system reliably produces increasingly macabre and violent imagery. [20]
To the layperson, Loab appears as evidence of a malevolent ghost in the machine. However, an examination of “latent space” reveals a pure reflection of human data distribution. Latent space is a high-dimensional vector space where concepts are mapped based on their relationships in training data. [21] When a user employs extreme negative prompting, they mathematically instruct the AI to move as far away as possible from a specific concept.
Because the training process suppresses content humans find undesirable (gore, trauma, deformity), this content is clustered together in an isolated “dark corner” of the latent space. [22] When an artist inverts the mathematical coordinates, they launch the algorithm directly into this suppressed cluster of macabre human data. [23]
Loab is not a ghost; she is an artifact of negative prompting colliding with the Jungian shadow of the internet. [24] She visually represents the limitations and uncurated depths of AI models. The system does not possess a desire to terrify; it merely traverses a mathematical topography built by human hands.
Many see “AI slop” or macabre outputs as a failure of the technology. In reality, AI slop is simply human slop. The machine is faithfully reflecting the uncurated, messy, and sometimes prejudiced depths of the internet. Instead of blaming the mirror, we can use these reflections to identify the flaws in our own collective data. When the mirror shows us something ugly, it is an invitation to clean the data, refine our prompts, and consciously improve the “human” side of the equation.
The Agentic Loop: The Power of Iteration
As the novelty of the chatbot wanes, the industry is transitioning toward “Agentic AI”—systems designed to execute multi-step workflows autonomously. This is achieved by embedding an LLM within an “Agent Loop,” a cyclic architecture that mimics human decision-making frameworks.
The foundation of the Agent Loop is the “OODA loop” (Observe, Orient, Decide, Act), developed by U.S. Air Force Colonel John Boyd for fighter pilots. [25]
Observe: The agent captures raw signals (API responses, web data, logs).
Orient: The LLM structures the data and derives intent.
Decide: The agent formulates a plan, often delegating to sub-agents.
Act: The agent executes code or alters a database, then feeds the result back into the “Observe” phase.
Proponents frame this as the dawn of autonomous synthetic intelligence. A prominent example is “Devin,” marketed as the world’s first AI software engineer. [26] However, the true value of the Agentic Loop is not “synthetic intelligence”; it is the power of iteration.
Agentic loops are the extreme formalization of human workflows, optimized for efficiency. Devin does not “engineer” in the human sense—navigating social ambiguity and applying moral imagination—instead, it engages in “workflow mimicry,” executing nested scripts and applying statistical pattern matching to resolve syntax errors. [27]
The success of AI agents proves that the secret to victory is not a single “perfect” insight, but a relentless commitment to the iterative loop. The OODA loop is a blueprint for winning through rapid feedback. [25] AI agents succeed because they can iterate a thousand times faster than a human.
This is a lesson for us. By adopting the agentic mindset—rapidly prototyping, failing fast, and refining—we can use AI as a force multiplier for our own success. We can embed ourselves in our own OODA loops, using the AI to accelerate the “Observe” and “Orient” phases, leaving us more room to “Decide” and “Act” with human intuition.
Embracing the Reflection
The true value of artificial intelligence is not found in its ability to mimic us, but in its ability to reflect us.
When we stop treating AI as a deity or a threat, we reclaim our agency. We recognize that the “alien” we feared was merely a reflection of our own history, our own biases, and our own potential. AI is a magnificent, flawed, and powerful extension of human intelligence—a tool that allows us to measure the “table” of reality with unprecedented precision.
We should not fear the mirror; we should use it. When the AI produces “slop,” let us refine our thinking. When the AI achieves a “miracle,” let us study the math. When the AI iterates a workflow, let us adopt the loop.
By embracing the Wittgenstein Mirror, we stop abdication and start augmentation. We use the reflection to see our blind spots, to automate the average, and to liberate our moral imagination. AI will not replace the human spirit; it will challenge us to be more human, more creative, and more intentional.
Look into the mirror, recognize the reflection, and then decide who you want to become.
Works Cited
Nassim Nicholas Taleb, Fooled by Randomness.
Nassim Nicholas Taleb, “Wittgenstein’s Ruler” / 10-sigma explanation.
President’s Working Group on Financial Markets, Hedge Funds, Leverage, and the Lessons of Long-Term Capital Management.
Google Analytics Help, “Get started with attribution”.
OpenAI, Training language models to follow instructions with human feedback.
Plato, Phaedrus, Perseus / Scaife Viewer.
Shannon Vallor, The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking, Oxford University Press.
Gilbert Ryle, The Concept of Mind, University of Chicago Press.
Arthur Koestler, The Ghost in the Machine.
Edmund Burke, A Philosophical Enquiry into the Origin of Our Ideas of the Sublime and Beautiful.
Immanuel Kant, Critique of Judgment.
David E. Nye, American Technological Sublime, MIT Press.
Jorge Luis Borges, “The Library of Babel”.
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, ACM FAccT 2021.
Jason Wei et al., “Emergent Abilities of Large Language Models”.
Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo, “Are Emergent Abilities of Large Language Models a Mirage?”.
Carl Jung, The Archetypes and the Collective Unconscious.
Carl Jung, Aion: Researches into the Phenomenology of the Self.
Supercomposite, original Loab thread.
Supercomposite, Loab Tumblr mirror.
David Foster, Generative Deep Learning, O’Reilly.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning, MIT Press.
Ashish Vaswani et al., “Attention Is All You Need”.
John R. Boyd, Patterns of Conflict.
Cognition AI, Devin / Cognition official site.
Cognition AI, “Introducing Devin”.
I want to talk about this!