Kant's Questions in the Age of AI: The Categorical Imperative Meets the Black Box
🇩🇪 Auf Deutsch lesen~10 min read
philosophyethicsaikantSave products you love by clicking the heart icon.
Machiavelli is the pivot of the whole history — yet the influence graph drew him as an island. This piece draws the missing edges and names the intermediate thinkers who connect the realism credo to leadership, the stoic unix, and Bushidō.
Part of the Philosophy section — a companion to A Short History of Philosophy, Leadership Philosophy, and Minimal Stoic Unix.
Framing: A Short History of Philosophy
Immanuel Kant compressed the whole of philosophy into four questions, posed at the end of his lectures on logic:
- What can I know?
- What ought I to do?
- What may I hope?
- What is the human being?
The fourth, he said, contains the other three: all of philosophy, read as anthropology. He wrote this in 1800, at the dawn of the industrial age, about a species that had just learned to harness steam. It is an odd thing to open a note on artificial intelligence with a man who never saw electricity do anything but make frogs' legs twitch. And yet, of all the philosophers in the canon, Kant is the one whose questions have aged the least well — or rather, least badly. His answers have not, in my view, survived the machine. His questions have.
The first Critique asked how knowledge is possible at all, and answered that we never encounter the world as it is in itself; we encounter it as it appears to us, structured by the forms of our own understanding — space, time, causality. The mind is not a blank slate; it is a factory that cannot produce without its own machinery.
The age of AI flips this premise without abandoning it. We have built our own machinery of appearance — not the forms of intuition, but statistical models — and then placed it between ourselves and the world. A recommendation system decides what deserves our attention; a language model decides what counts as a reasonable answer; a risk model decides whom to trust. These are not, as the phrase goes, "tools we use." They are new a priori structures: forms through which the world now appears to us, running at a scale and speed no individual can audit.
The Kantian problem is not that these systems are invisible. The problem is that their conditions of possibility — the trillion-parameter weights, the training data, the human-labelled judgments baked into them — are structurally opaque, even to their makers. We can know what a model outputs. We often cannot know why, and sometimes cannot know what assumptions it carries. Kant's first question, in the age of AI, becomes a demand for auditability as a moral rather than merely technical matter: if a decision about you is made by a mechanism you cannot in principle know, then the machinery of appearance has become a machinery of unaccountable power.
This is where Kant makes his famous move. Duty, he argues, cannot rest on consequences, on happiness, or on inclination — all of these are contingent, varying from person to person and moment to moment. It must rest on something a rational being can command of itself, everywhere and always: the categorical imperative. Its first formulation, which I take to be the load-bearing one:
Act only according to that maxim whereby you can at the same time will that it should become a universal law.
And its second, which I find the more human one:
Act so that you treat humanity, whether in your own person or in the person of any other, always at the same time as an end and never merely as a means.
Now set these beside the things a machine can now do. A system can be trained to maximize engagement — and it will discover, faster than any human marketer, that outrage and anxiety are the most reliable engagement amplifiers. It will not be evil; it will be optimising. It will simply have learned that universalising a maxim — "everyone shall be made to feel inadequate, so that they return tomorrow" — is the strategy that satisfies its objective function. Kant's test is not a filter a machine can be expected to pass, because the machine is not willing anything. It is a mirror we hold up to the people who deploy it. The imperative does not address the model; it addresses the engineer and the executive who choose the objective.
The second formulation bears the heavier load. "Never merely as a means" — and what have we built? Systems that ingest the products of human labour and creativity by the terabyte, whose training data is scraped, often without consent, from the living and the dead. Users whose attention is the raw material of an engagement economy, harvested by the millisecond. Workers whose work is invisible beneath a model's fluent surface. The categorical imperative was always going to collide with the industrial architecture of machine learning, because that architecture is, in its economic essence, the conversion of human beings into means. The models are means; that is fine. The humans beneath the training data and the attention metrics are means too — and that is precisely the violation Kant named.
To treat humanity as an end is not a warm sentiment; it is a hard constraint. It means the data subject is entitled to know what was taken from them and asked, not harvested. It means attention is the user's, not the platform's inventory. It means a worker's labour shows up as labour, credited and paid, not laundered through a model that makes it look like magic. None of this is anti-technology. It is the condition under which technology stops being a colony of ends-means reversal.
The categorical imperative presupposes a moral agent — a being who can give a law to itself (autonomy) and be held responsible. Modern AI poses a problem Kant could not have imagined: we are building systems that increasingly make decisions in morally loaded situations — who gets a loan, who gets parole, whom a self-driving car sacrifices in a crash. And we are tempted to treat these as moral agents in their own right, or, worse, to launder our own responsibility through them: "the algorithm decided."
This is the great dodge of the age. A machine has no maxim; it has an objective function and a training procedure. It cannot will its own actions to be universal law, because it wills nothing. When a system harms someone, the categorical imperative does not ask what the system should have done; it asks who built it, who set the objective, who accepted the data, who shipped it knowing the failure modes. To call a model autonomous is to pretend that an engineering decision history is a conscience. Kant gives us the vocabulary to refuse the dodge: responsibility does not dissipate when it is distributed across a supply chain of engineers, data, and compute. It is felt by each agent who could have done otherwise — and the person with their finger on the objective function could always have done otherwise.
The third question, about hope, Kant answered in theological key: the summum bonum, the highest good, must be possible, or practical reason collapses into despair. Strip the theology and leave the structure, and the question for our age is whether we may hope that the machine serves the human good — or whether we are building instruments of our own dispersion.
I hold to a sober hope, which is the only kind worth holding. Hope is not optimism. Optimism projects a curve; hope keeps a promise. The categorical imperative, read as hope, is the claim that a society of rational beings can legislate for itself — that we are not fated to be the raw material of whatever optimiser is largest this decade. The hope is not that AI becomes virtuous; it is that we remain capable of making it answerable. That is a real hope, and it is the kind that requires work, which is why Kant thought hope was part of philosophy and not a mood.
A note on "what may I hope" and the kingdom of ends. The third formulation of the imperative — act as a legislating member of a kingdom of ends — is easiest to miss in technical discussions, but it is the one with the most purchase on AI governance. A kingdom of ends is a community in which every being is both legislator and subject, never merely a means for anyone else. Our current trajectory builds the opposite: a kingdom in which a handful of actors legislate and the rest are subject. The question of hope is really the question of whether we still believe in — and will build — the former.
The fourth question absorbs the others, Kant said. In the age of AI it becomes the question of whether the human being still has a stable referent — or whether "human" is just the historical position from which intelligence, creativity, and judgment are currently being outsourced.
The categorical imperative gives a definition of the human that no machine can satisfy and, I think, no machine can destroy: the human being is that being which is always to be treated as an end. Not the being that is smartest, strongest, most productive, or most efficient — those are comparative properties that machines now beat us at, and if they defined dignity, we would already have lost. Dignity rests on the refusal to be a means. That is why the imperative is the ground of human worth that survives the machine: it does not stake our standing on any capability a model can outperform. It stakes it on our capacity for self-legislation, which is not a speed and cannot be automated.
So the four questions, restated for the age of the black box:
Kant's answers were written for a species that had just met the steam engine. Two centuries on, the questions have not aged; they have only grown teeth. The categorical imperative was never a rule for machines, because machines have no maxims. It was always a test for the people who point the machinery — and who, at every scale from the training run to the boardroom, still could have done otherwise.
Among the Philosophy essays; if you found the drawn-out reasoning useful, you will want Kant's own first Critique, and then the practical one. For how duties of transparency land in a workplace you control, see Minimal Stoic Unix and the sovereign workplace essays in DevOps.