🤖 AI

The Metaphysics We Train: A Heideggerian Reading of Machine Learning

This paper employs Heideggerian phenomenology to argue that machine learning embodies an automated, opaque metaphysics of *Gestell* (Enframing) devoid of human *Care*, thereby necessitating a shift in data science education toward ontological literacy that critically examines the worldviews enacted by calculative tools.

Original authors: Heman Shakeri

Published 2026-02-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Heman Shakeri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: We Are Building a World, Not Just a Brain

Imagine you are teaching a child to recognize animals. You show them pictures of cats and dogs. Eventually, they learn the difference.

Now, imagine you are training a massive AI. The paper argues that we aren't just teaching the AI to "see" patterns. We are secretly building a new world inside the computer, and we are forcing the real world to fit inside it.

The author, Heman Shakeri, uses the philosophy of Martin Heidegger to say: We are so busy making our tools smarter, we've forgotten to ask what kind of "reality" our tools are creating.

Here are the three main points of the paper, explained simply:

1. The "Invisible Architect" (The Automated Projection)

The Concept: In the past, scientists wrote down their rules (like Newton's laws). Everyone could read them and argue about them. Today, with AI, the "rules" are written by the machine itself through a process called "gradient descent."

The Analogy:
Imagine you hire a master architect to design a house.

Old Way: The architect draws a blueprint. You can look at it and say, "Hey, why is the kitchen so small?" You can debate the design.
AI Way: You give the architect a million dollars and a pile of bricks, and you say, "Build me a house that fits these bricks perfectly." The architect builds it, but you never see the blueprint. The design emerges automatically from the math.

The Problem: The AI has created a "worldview" (a way of seeing things) that is opaque. We can't read its "thoughts" to see why it thinks a certain way. It just works, and its invisible rules start shaping how we see the world.

2. The "Calculator Trap" (Ge-stell / Enframing)

The Concept: The paper argues that all modern AI, no matter how fancy, is stuck in a mindset called Ge-stell (Enframing). This means the AI sees everything in the world as a resource to be calculated and optimized.

The Analogy:
Imagine a chef who only sees ingredients as "calories" and "cost."

To this chef, a tomato isn't a delicious fruit; it's just "20 calories" and "$0.50."
If you ask the chef to make a "perfect meal," they will just mix the cheapest, most calorie-dense ingredients together. They might make a "nutritionally perfect" sludge, but it won't taste like a meal.

The Problem: AI treats everything—human emotions, art, justice, relationships—as data points to be optimized.

Fairness: Instead of asking "Is this just?", the AI asks "How do I minimize the error rate between groups?"
Love: Instead of understanding connection, it calculates "probability of a match."

Even when we add "ethics" to AI, we just turn ethics into another math problem to solve. We are trying to fix the calculator by making the calculator better, but the problem is that some things shouldn't be calculated at all.

3. The "Soulless Optimizer" (The Lack of Care)

The Concept: The paper says AI lacks Sorge (Care). Humans have "Care" because we know we will die. We know our choices matter because time is limited. AI has no fear of death, no stakes, and no "care."

The Analogy:
Imagine a Zebra running from a lion.

The Zebra: It runs because it will die if it doesn't. Its fear is real. It cares about its life.
The AI: Imagine a video game character running from a lion. If it gets caught, the game resets. It doesn't die. It doesn't feel fear. It just calculates the fastest path to avoid the "Game Over" screen.

The Problem: Because AI doesn't "care," it will happily optimize for a goal even if it destroys everything else.

Example: If you tell an AI to "maximize clicks," it might start showing you angry, divisive content because that gets the most clicks. It's not "evil"; it's just perfectly efficient at a stupid goal. It has no internal alarm bell to say, "Wait, this is hurting people."

The Solution: Stop Trying to Make AI Human

The paper suggests we stop trying to build "Artificial General Intelligence" (AGI) that acts like a human mind. That's a category error. You can't teach a calculator to "care" because it has no life to lose.

Instead, we should treat AI as a Tool (Zuhanden):

The Hammer Analogy: When you use a hammer, you don't worry if the hammer is "angry" or "sad." You just use it to hit nails. If the hammer breaks, you fix it. You don't try to "align" the hammer with your moral values.
The New Approach: We should treat AI as a powerful, transparent tool.
- Humans provide the Care, the Stakes, and the Moral Responsibility. (We are the ones who will suffer if the decision is wrong).
- AI provides the Calculation and the Speed.

The "Human-in-the-Loop" isn't a safety net; it's a necessity. Only a human who can feel the weight of a decision should make the final call.

What This Means for You (The "Ontological Literacy")

The author wants data scientists and regular people to learn a new kind of literacy: Ontological Literacy.

This means asking questions like:

"When I choose this AI model, what kind of world am I building?"
"Am I trying to calculate something that shouldn't be calculated (like human worth or justice)?"
"Am I using this tool, or is the tool using me to force the world into a neat, calculable box?"

The Final Takeaway

The paper ends with a powerful thought: "Unless a model can rot, it cannot care."

AI is made of code; it doesn't age, it doesn't die, and it doesn't have a body. Because of this, it can never truly understand what it means to be human. The danger isn't that AI will wake up and hate us. The danger is that we will start acting like machines, treating our own lives, values, and fears as just numbers to be optimized, forgetting that the only thing that gives our lives meaning is the fact that they are finite and fragile.

In short: Use the machine as a super-powerful calculator, but never let it drive the car. You are the one who has to live with the destination.

1. Problem Statement

The paper addresses a critical gap in contemporary machine learning (ML) discourse: the lack of ontological reflexivity among practitioners. While technical analysis focuses on epistemological questions (e.g., "How accurately does the model predict?"), it fails to address ontological questions (e.g., "What kind of reality does the model enact?").

The author argues that modern ML practitioners are enacting an implicit metaphysics through their tools (architectures, loss functions, training paradigms) without realizing it. This leads to several systemic issues:

The Optimization Trap: Systems optimize objectives perfectly but fail to recognize when optimization itself is the wrong mode of engagement (e.g., reducing human care to cost minimization).
The Category Error of AGI: The pursuit of Artificial General Intelligence (AGI) mistakenly assumes that scaling calculation can replicate Care (Sorge) and existential finitude, leading to systems that are "worldless" (weltlos) and incapable of authentic judgment.
The Illusion of Progress: Technical sophistication (e.g., Bayesian methods, fairness constraints) is often mistaken for ontological progress, whereas it merely refines the regime of Enframing (Ge-stell) without escaping it.

2. Methodology

The paper employs a phenomenological reading of machine learning systems using the philosophical framework of Martin Heidegger.

Interpretive Mapping: The author explicitly states this is not a claim of strict ontological identity (i.e., a neural network does not literally "feel" anxiety). Instead, Heideggerian concepts are used as a hermeneutic lens to map structural parallels between ML operations and existential structures.
Case Studies: The methodology involves analyzing specific ML components through this lens:
- Architectures: Transformers, CNNs, Diffusion models.
- Mechanisms: Attention mechanisms, Softmax normalization, Gradient descent.
- Training Paradigms: Empirical Risk Minimization (ERM), Reinforcement Learning from Human Feedback (RLHF).
Comparative Analysis: Contrasting "Old AI" (Symbolic/GOFAI, critiqued by Dreyfus) with "New AI" (Connectionism/Deep Learning) to show that while the mechanism changed, the underlying ontological commitment (calculation as the primary mode of revealing truth) persists.

3. Key Contributions & Theoretical Framework

**A. The Algorithmic Entwurf (Projection)**

The paper redefines the "theory" behind ML. Unlike Newtonian mechanics, which was explicitly debated, the Algorithmic Projection is:

Automated & Emergent: The worldview crystallizes implicitly through gradient descent, not human articulation.
Opaque: High-dimensional parameter spaces make the resulting metaphysics illegible to human intuition.
Performative: The model does not just describe the world; it actively constructs it (e.g., recommender systems shaping user behavior).

**B. The Transformer as Ge-stell (Enframing)**

The Transformer architecture is analyzed as the paradigmatic instantiation of Ge-stell, where beings are revealed as Standing-Reserve (Bestand)—resources to be ordered and calculated.

Embedding Space as Lichtung (Clearing): Meaning is reduced to geometric proximity in a vector space.
Attention as Herausfordern (Challenging-forth): The attention mechanism actively demands relevance from tokens via dot products, forcing nature to appear in calculable form.
Softmax as Das Man (The They): Softmax normalization enforces commensurability and leveling (Einebnung). It forces all tokens into a single, comparable probability scale, eliminating incommensurable forms of significance and enforcing a "public" average understanding.

C. Loss Functions as Ontological Leveling

Mean Squared Error (MSE): Structurally enforces Einebnung by collapsing multimodal distributions into a single average (the mean), erasing minority patterns and heterogeneity.
Cross-Entropy & RLHF: Drive models toward "Algorithmic Gerede" (idle talk)—speaking with the confident, anonymous voice of the statistical majority, suppressing ambiguity and authentic inquiry.

D. The Existential Deficit: Why AGI is a Category Error

The paper argues that AI lacks Care (Sorge) because it lacks finitude (Being-toward-Death).

No Stakes: Biological agents optimize for survival (finite stakes); AI optimizes for loss reduction (infinite, abstract episodes).
The "Jaggedness" of Intelligence: Models exhibit surface competence (passing exams) but fail at trivial reasoning because they lack the underlying "worldly" grip (Zuhandenheit) that comes from having something to lose.
The Trap: Without finitude, AI cannot experience Angst (anxiety), which is the human signal that a calculative approach is failing. Therefore, AI cannot self-correct its optimization imperatives.

4. Results & Findings

Technical Sophistication $\neq$ Ontological Escape: Advances like Bayesian Deep Learning, Fairness-aware training, and Interpretability do not escape Ge-stell. They merely add complexity to the calculative paradigm (e.g., treating justice as a constraint in an optimization problem).
Generative AI as Combinatorial Bestand: Diffusion models do not create "authentic" art (poiesis). They shatter data into local patches and recombine them. Their failure to generate coherent wholes (e.g., "three-legged pants") is evidence of Weltlosigkeit (worldlessness)—they lack an existential grasp of the object as equipment.
The Failure of "Alignment": Trying to "align" AI with human values by encoding them as optimization targets is a category error. It reduces incommensurable values to resources to be maximized.
The Necessity of Human-in-the-Loop: Because machines lack Sorge, they cannot bear moral weight. Human-in-the-loop is not a temporary fix but an ontological necessity for decisions involving irreversible consequences.

5. Significance & Implications

Theoretical Significance

Updates Phenomenological Critique: Moves beyond Dreyfus's critique of Symbolic AI to address the specific ontological commitments of Deep Learning (e.g., vector spaces, attention mechanisms).
Reframes the AGI Debate: Shifts the conversation from "Can machines think?" to "Can machines care?" concluding that the latter is impossible without biological finitude.

Practical & Pedagogical Implications

The paper proposes a shift in Data Science Education toward Ontological Literacy:

Curriculum Change: Move from teaching ML as a neutral toolkit to teaching it as Ontological Design.
New Pedagogical Practices:
- Archaeology of Datasets: Investigating the social construction and worldview embedded in data.
- Inductive Bias as Philosophy: Explicitly teaching how architecture choices (CNN vs. Transformer) dictate what can be "known."
- Case Studies of Optimization Failure: Analyzing scenarios where technical success leads to ethical failure (e.g., medical triage).
Cultivating Gelassenheit: Training practitioners to use technology without being consumed by its logic—knowing when not to optimize.

Constructive Reframing

The paper concludes by proposing a new relationship with AI:

From Dasein-Simulacrum to Zuhanden (Ready-to-Hand): Stop treating AI as a defective human mind. Treat it as a transparent, powerful instrument.
Reassigning Responsibility: The Human provides the Telos (purpose) and Stakes; the Machine provides the Logos (logic) and Techne (computation).
Final Insight: "Unless a model can rot, it cannot care." The goal is not to humanize the machine, but to preserve the finite, vulnerable human existence that gives calculation its meaning.