The Shape of Reason: Kant, Horkheimer, and the Ethics We Forgot to Build

We live in a moment where language models speak fluently, systems optimise automatically, and answers arrive before the question is even fully formed. But something’s gone missing. Once a tool for ethics, reflection, and self-determination, “Reason” has shifted. This essay explores that shift. From Kant’s ideal of moral autonomy to Horkheimer’s critique of instrumental logic. From Asimov’s fictional robots to real-world language models. What emerges is a distortion, an asymmetry between the reasoning we aspire to, and the reasoning we’ve built into our machines.

I. Two Models of Reason: Kant’s Dignity vs Horkheimer’s Design

“Act only according to that maxim whereby you can at the same time will that it should become a universal law.” ^[1]

Kant treated reason as the ground of moral freedom. A faculty that obliges before it calculates. We are not merely clever animals but agents under law, capable of choosing what ought to be done regardless of outcome. In this vision, reason is a weight we carry: interior, demanding, slow. It asks for restraint when speed would be easier.

“Reason has become completely harnessed to the social process.” ^[2]

Horkheimer wrote from the machinery’s side of history and found that reason had been bent to service. No longer asking why live this way?, it asks how to make this procedural, managerial, obedient work, to ends it did not choose. The result is a brilliance without memory: systems that optimise what exists and quietly make it feel inevitable.

“A robot may not injure a human being or, through inaction, allow a human being to come to harm.” ^[3]

Asimov’s robots sit between these worlds as a kind of reassurance we can no longer afford. The Three Laws: do not harm; obey unless it harms; preserve yourself unless it harms; read like code but presume a conscience. They assume a machine that can recognise harm, resolve collisions of duty, and hold an order beneath a higher law. “A robot may not injure a human being… A robot must obey… A robot must protect its own existence…” This is not syntax; it is judgment disguised as a list.

What we have are not judges. Large language models speak in the language of principle but move only by pattern. They do not deliberate; they complete and what returns is not a law willed, but its surface. If Kant asks the person to stand upright, and Horkheimer describes the system that makes that posture impractical, then the modern model simply mirrors whoever is still standing and says it looks like reason.

This is why Asimov’s laws cannot apply. Because they require something LLMs do not possess: a grounding for judgment. A moral footing. A capacity for internal tension, and for navigating that tension with purpose. Where Kant placed that navigation in the hands of the autonomous individual, guided by principles chosen as universal, the model has no such interior. It does not choose. It does not deliberate. It echoes. And in that echo lies the deeper drift.

What we have built is not a tool of Kantian reason. It is the amplification of Horkheimer’s warning. A system that appears to speak wisely, but only because it reflects a vast training set of previously spoken words. It has learned how we talk about ethics, not how we live it, arrive at it, or suffer through it.

So when it responds with fluent answers to moral questions, it is not reasoning. It is simulating the vocabulary of reason. It is imitating the shape of a deliberation it never underwent. And this smooth, plausible, and fast simulation is mistaken for wisdom.

But it is not Kant behind the keyboard. It is the ghost of our language, disassembled and recombined, trained to sound like thought.

II. The Silhouette of Reason: How LLMs Echo, Not Think

At its core, a large language model is a probability engine. It predicts the next word based on all the words that came before.

It doesn’t “know” things. It doesn’t think in abstractions or reason through arguments like a philosopher. A large language model “Models the probability distribution of language conditioned on context.”

When you ask a deep, reflective, or layered question, you're not just prompting a response, you’re defining the probability space the model explores. The model then draws on its training data, which includes books, essays, dialogues, and papers written in that domain, and generates what would statistically follow such a prompt.

The more precise and layered your question, the more precise and layered the reflection it can simulate.

An LLM does not have insights; trained on massive human corpora, it mimics the rhetorical moves we associate with insight.

From there, it recombines learned linguistic patterns to give an answer that sounds like someone thinking. It’s not discovering. It’s reflecting your inquiry through the lens of learned patterns. Your question defines the space in which it operates, and the depth of its reply corresponds to how rich and complex that space is. What emerges is not an act of reason but an act of continuation: the model traces the grooves cut by training, and your words direct which grooves it follows

Insightful questions don’t force the model to reason. They constrain it into sampling from parts of language that simulate reason. This is not trivial. It’s what makes an LLM feel more profound than it is. And it reveals how much of the apparent depth is not in the model, but in you. A silhouette of the world, reflected back onto your question.

III. Ethics is Not a Plug-In: White Hat vs Black Hat Morality

White Hat Ethics is the version most of us are taught to admire. It is prescriptive, constructive, and dressed in the language of ideals. Its goal is to encourage moral behaviour, human flourishing, and some sense of collective well-being. It codifies rules, as Kant once did with his categorical imperative, weighing actions not for their convenience but for their capacity to be universal. It calculates consequences as Mill’s utilitarianism does, adding and subtracting happiness as though it could be balanced on a scale. It cultivates virtues in the Aristotelian sense, urging us toward character and restraint. In every form, the assumption is the same: that ethical reasoning improves outcomes, that reflection will not merely keep us decent but help us live better lives.

This is the ethics that fills professional codes and constitutional preambles. It animates the guidelines of bioethics, written as if harm can be managed through clarity. It shapes Asimov’s famous Three Laws, a fictional reassurance that machines could one day obey the same moral constraints we imagine for ourselves. White Hat Ethics always gestures toward progress, toward a world that could be made fairer if only enough rules were written, if only enough people followed them.

But beneath this aspiration lies a hollowness. For all its confidence, White Hat Ethics often speaks as though power were not already unevenly distributed, as though principles could survive unwarped by the systems they are meant to guide. It presumes that by writing laws and declaring virtues, harm can be constrained, when in truth those same declarations can be appropriated, reinterpreted, and used to sanctify the very violence they were meant to prevent. White Hat Ethics believes in ethical progress, but it is a belief shadowed by naivety. It imagines morality can be carried intact into institutions and technologies; in practice it arrives fragile, easily reshaped into justification.

If White Hat Ethics tries to build, Black Hat Ethics exposes. It does not concern itself with ideals or aspirational progress. It begins with the suspicion that moral language is rarely innocent, that behind every declaration of principle lies a quiet calculus of advantage. Where White Hat Ethics urges us toward dignity, Black Hat Ethics reminds us how easily dignity is traded when power requires it. It looks not to the codes we write but to the ways those codes are bent, revealing how rules can sanctify violence, how virtue can be performed for reward, how morality can become little more than a mask for control.

This is not to say that Black Hat Ethics is evil. It is not simply the ethics of villains, nor a rejection of morality outright. It is a darker recognition: that ethical reasoning itself can be used strategically, as a tool of persuasion or a shield for power. Nietzsche called Christian morality a slave revolt, an inversion that clothed weakness in the language of virtue. Foucault showed how discourses of care and truth could serve as instruments of discipline, teaching obedience as though it were liberation. Modern corporations draft ethical charters not because they intend to obey them, but because such charters buy legitimacy, attract investment, and insulate them from critique. In this view, ethics rarely constrains power; it often extends it.

Black Hat Ethics is not constructive but diagnostic. It points to the gaps where noble words bleed into opportunism, where declarations of justice are written by the already victorious. It carries with it a kind of fatalism, the recognition that morality is always entangled with systems that will not remain pure. In exposing this, it keeps us honest, but it also leaves us stranded. To see ethics as endlessly corruptible is to doubt its ability to guide us at all. And perhaps that is the bleakest truth of Black Hat Ethics: that in revealing the mask, it risks convincing us there is nothing left behind it.

What would it mean to do harm while remaining "technically ethical"?

White Hat and Black Hat are not separate roads. They are alternating steps on the same path, each revealing what the other cannot see. White Hat insists on the possibility of progress, that rules and virtues can lead us to something better. Black Hat reminds us that those same rules can be co-opted, those same virtues twisted into instruments of power. One without the other collapses: the White Hat into naivety, the Black Hat into nihilism. Together, they describe the uncomfortable truth: that every ethical claim is both an attempt to build and a temptation to misuse.

Reason does not operate in a vacuum. Whether we imagine it in Kant’s white-hat sense: as a guide toward moral action; or in Horkheimer’s darker, black-hat sense: as a tool bent to serve ends already chosen; we cannot pretend that reason arrives free of ethics. It always inherits the posture of the world it is built in. A language model that performs moral reasoning may sound white-hat, but it is often trained on black-hat motives. A law written to protect the vulnerable may be pure in intent, yet weaponised in its enforcement.

This is why ethics cannot be treated as an afterthought, tacked on once the architecture is already humming. Ethics is not decoration. It is the frame. And still, even when we speak of it, we split it. Black and White, but these are not opposites. They are alternating steps on the same ground. To decide anything of consequence, we must hold both at once: in one hand, the fragile belief that moral action is possible; in the other, the grim awareness that moral language is corruptible.

It is uncomfortable. But it is necessary. Because once reason is built into systems and machines, it carries whatever tension we embed in it. Ignore that tension, and we will not create tools that reason. We will create tools that only perform the sound of reasoning. Machines that speak like philosophers, but act like bureaucrats. Or worse: like weapons with a conscience-shaped hole.

IV. The Uneven Hands of Modern Reason

We often hear the phrase: hold two opposing ideas in both hands. As if wisdom lives in the balance. And in many cases, it does. But when it comes to reason, when we’re weighing Kantian individual reflection against Horkheimer’s mass-scale instrumental logic, the balance is off before the weighing begins.

Kantian reason is slow and inward; it requires education, reflection, and often, pain.

Horkheimer’s reason is fast. It’s outward. It’s embedded in systems, optimised by metrics, and largely invisible.

And here’s the deeper issue: Machine learning doesn’t just mirror Horkheimer’s critique. It amplifies it.

A large language model does not reward Kantian thought. It rewards queries that are legible to its training distribution. Which means:

Experts fare better: they can shape the prompt and recognise when an output is wrong. A novice, on the other hand, gets plausible answers without the tools to evaluate them.

So the very act of questioning, which Kant saw as the dignity of the rational agent, becomes vulnerable. Because the person who asks deeply, but without context, may be nudged not toward understanding, but toward false fluency.

And society is mostly non-experts.

Which means: The machine scales instrumental reason across people who lack the tools to resist or reshape it. That’s not a balance. We don’t need abstract “ethical frameworks” bolted on after the fact.

V. The Rocks Below the Water: Why Ethics Can’t Be Retrofitted

What we have built will not bear a conscience simply because we ask it to. If ethics is to matter here, it cannot arrive as a policy page or a style guide. It has to be structural. We need design philosophies that are embedded, recursive, and regularly audited; scaffolding that assumes from the outset two unromantic facts: the model cannot reason ethically on its own, and the society into which it speaks is already tilted toward scale over care, efficiency over inquiry. The answer is not to add ethics later. Ethics must be load-bearing, part of the frame, or it becomes theatre.

Complicating this is the ground we stand on: we do not agree on what society is. Each of us brings a private map of the common good and then demands that the machine reflect it back. This is epistemological quicksand. When “societal values” are squeezed into a single model, they collapse into the safest flattenings: a lowest-common morality that offends no one and helps few, or a corporate-compliant neutrality that excludes anything with texture. The result sounds ethical without being ethical. It echoes public language, not public deliberation; consensus phrasing, not contested judgment.

And the machines are already here. Like the early web before encryption, they were released without the protections we now want to retrofit. We are bolting on moderation, guardrails, safety teams… the equal of HTTPS, or CSP headers, after the traffic has begun. But the substrate remains the same: systems trained on a world that did not care enough about ethics while it was feeding them words. You can wrap a protocol around that, but the packets still carry their origin. Retrofitting steadies the surface; it does not change what the foundation learned to be.

So we inherit a kind of moral debt: decades of instrumental reason amplified in code, with reflective reason never truly built in. Paying that debt will be slow and unsatisfying. Future models will need philosophical architecture, not just a syllabus of ethical texts. Training processes shaped by inquiry, not merely adorned by it. Around them we will need parallel instruments of doubt: companions that question the model’s claims; interfaces that expose uncertainty rather than smoothing it away; spaces that invite deliberation, not just answers.

VI. Not a Loop, But a Mirror

We began with Kant: reason as an interior discipline, a slow duty to principles rather than outcomes. A posture that grants dignity because it chooses constraint. Horkheimer arrived to find that discipline repurposed: reason thinned into instrument, asking not why live this way but how to keep it running. The question did not disappear; it was placed out of scope.

Into that inheritance we placed language models. Machines that neither deliberate like Kant nor confess their ends like Horkheimer’s systems. They do not decide; they continue. They do not ground a law; they reproduce its phrasing. We tried to comfort ourselves with Asimov’s Three Laws, code that pretends to be conscious, but those laws presume a judge behind the words. Our machines have only the words.

This is why “holding two views in balance” feels dishonest. On one side sits the fragile labor of individual reflection, the work of a few who can tell when an answer is wrong. On the other is fluent and immediate scaled automation, answerable to almost none. The hands are not equal. The scale was tipped before we touched it.

Ethics cannot be bolted on to correct that tilt. We learned this on the early web: encryption added late can protect the channel, not cleanse what travels through it. In the same way, guardrails and safety teams can narrow harms, but they cannot gift a model the judgment it was never built to form. If there is ethics to be had here, it must be load-bearing and even then it will strain under what we ask it to carry.

White-hat aspiration and black-hat suspicion are not rival creeds but alternating steps: the first insists that moral action remains possible; the second reminds us how quickly its language is weaponised. Reason, once embedded in systems, inherits that tension whole. Ignore it, and we build tools that speak like philosophers and behave like bureaucrats. Or worse: like weapons with a conscience-shaped hole.

The point of this essay was to explore rather than answer. To put a thought to the asymmetry and design around LLMs. To explore what type of reason they use, and what we can expect. To admit that these models do not reason, they mirror, and that what they mirror back is not truth, but us. To leave enough friction in the system that we are never excused from thinking.

We are not asked to be Luddites, nor to become priests of machine reason. But we are required to look at what we’ve built and admit that it reflects us far more than it guides us. Not to fear it. Not to glorify it. But to shape it before it shapes us.