An engineer looks at a control panel. Pressure rising, temperature unstable, a combination of signals that no alarm has yet triggered. In three seconds, he decides to evacuate. He is right. Asked the next day, he struggles to explain why — "I felt it." What he drew on in those three seconds was thirty years of incidents compressed into a judgment that no procedure contains.
The philosopher Michael Polanyi captured this in a single phrase, more than half a century ago: we know more than we can tell. He called it tacit knowledge — the part of what we know that governs an expert's every move yet never fits into a manual. You find it everywhere: in the surgeon's hand, in the negotiator's ear, in the way a strategy committee frames its questions rather than in the answers it gives.
That is the part we are losing. Not our data — we talk about that enough. The tacit. And it does not vanish in a breach; it retires.
The knowledge that isn't there
A few years ago, I believed the problem was one of digitization. That it was enough to scan everything, index everything, feed it all to a machine to read. The field corrected me.
In a large industrial organization, most of what is known does not exist in digital form. It lives in people's heads, or in millions of pages never scanned. When it has been digitized, it often takes a form a machine cannot exploit — a diagram, a table, a note in the margin. And by the time you finally go looking for it, some of the experts who could have interpreted it are already gone.
The naïve answer — vacuum up everything, feed it all to the model — therefore fails twice. It drowns in dead documents, and it misses precisely what mattered: the tacit, which was written down nowhere. This answer has a name. It is extraction.
Funes, or why preserving is not accumulating
Borges wrote the story of a man, Funes, struck with total memory. He remembered everything: every leaf of every tree, every instant of every day, in infinite detail. And Borges observes something terrible — Funes was, for that very reason, incapable of thought. To think is to forget differences, to generalize, to abstract. A memory that keeps everything sorts nothing. It does not think; it archives.
This is the answer to those who confuse preserving with accumulating. To preserve is not to keep everything — it is to choose. To keep the living reasoning, not the dead document. A principle recent research has formalized in a simple phrase, less is more — LIMO: beyond a certain point, adding material degrades thought instead of enriching it. Selection is not a loss; it is the very condition of intelligence.
Extraction accumulates and dilutes. Preservation curates and concentrates. They are opposite gestures — and it is the second, not the first, that keeps what makes an organization singular.
Thinking with the same tool
Here is why this matters more than it seems — and I want to be precise, because the point invites misunderstanding. A large language model computes no "average": it is deeply contextual. You steer it, and it follows, through a space of meaning of dizzying richness, the direction you give it. The same model, depending on how you question it, produces the banal or the remarkable.
That is precisely what should give us pause. For, left to a generic context and to conventional questions, a model follows the most probable slope: the already-written, the already-thought. It will be singular only if the context you entrust to it is singular. Yet at scale the opposite happens — and the cause lies not in the machine but in the use. When entire organizations converge on the same tools and question them in the same way, the diversity of outputs fades because the diversity of questions fades. The uniformity of the answers reflects the uniformity of the demands.
The philosopher Byung-Chul Han put it for our age: information is not knowledge; its sheer mass produces noise, and noise dissolves singularity. The enthusiastic director who places the same assistant in the hands of all his managers believes he is gaining productivity. He does not see that he is asking a tool fed on common knowledge to write the next singular chapter of his own story.
The real risk, then, is not the data breach. It is the contamination of thought — the slow erasure, in the name of innovation, of what should have stayed singular. And this observation points, in negative, to the remedy: since a contextual model is worth only the context you give it, everything turns on the nature of that context. The common knowledge of everyone, or your own.
The poison and the remedy
None of this is new, and that is what should alarm us. In the Phaedrus, Plato tells how a god offered writing to King Thamus, praising it as a remedy for memory. The king refused: this craft, he said, would give "the appearance of wisdom, not wisdom," and would "implant forgetfulness in the souls" of those who relied on it. Twenty-four centuries later, we are having the same debate, with an infinitely more powerful tool.
Bernard Stiegler gave this paradox its name: the pharmakon. The same substance is remedy or poison depending on how it is used. The technology that augments can also un-teach; the one that liberates can also dispossess. Stiegler spoke of proletarianization — the gradual loss of a know-how captured by the technical system, until the one who knew no longer knows. Extraction proletarianizes thought. The question, then, is not whether artificial intelligence is good or bad. It is under what condition it becomes remedy rather than poison.
Proof must become invisible
The condition is trust. Not the trust one asks for — the trust one can verify.
In 1995, a small icon changed the Internet: the padlock. It did not make the web smarter; it made it trustworthy. You could finally entrust it with a transaction, a secret, an identity. Then something remarkable happened: the padlock faded from our attention. No one thinks anymore about the protocol securing their payments. Proof became ambient, invisible — and it was precisely at that moment that it made everything possible.
Artificial intelligence is waiting for its padlock. It produces brilliant answers of which no one can say, on reading them, whether they are true. An architecture of veracity is its missing keystone: every claim set against its sources, traced, dated, replayable identically. Not one more model that speaks, but a layer that attests. Evidentia: what is proven becomes, in the end, evident. The day proof becomes invisible as the padlock became invisible, trust ceases to be a question — and the pharmakon tips to the side of the remedy.
One will object that tacit reasoning, by its nature, cannot be proven. That is true of an intuition left to itself. It is no longer true once it is preserved as verified, contextualized units, connected to what grounds them. What is then returned is not an opaque answer, but a path you can follow, contest, replay. Proof does not replace judgment. It gives judgment back what it needs to decide.
The remedy: to augment
Once veracity is made trustworthy — and sovereign, that is, running on our own soil and under our own rules, for sovereignty is not a flag planted on a data center but an architecture — the remedy can operate. This remedy has a simple name: to augment. Never to replace. To augment.
To augment the expert, whose reasoning outlives his departure — not a double that would speak in his place, not an oracle that would spare his successor the effort, but material that a living person reclaims by discussing it and pushing against it. To augment the organization, by making visible what it knows without knowing it: that immense share of its collective intelligence, scattered, that no one connects at the right moment. And when that share becomes visible — when the cognitive phenotype of an organization, its own way of reasoning, finally appears — something happens that search never gives. Search answers the question you ask. Revelation connects what you did not think to connect: a line of reasoning abandoned yesterday and a method published today, an expertise that lay dormant two floors below. These encounters owe nothing to chance. They become possible the moment an infrastructure preserves the full context of each piece of knowledge and lets value attract value. Nothing resists value: two sound pieces of knowledge, brought together in the right context, produce more than their sum.
That is exactly where preservation stops being a defense and becomes a force. We do not keep the past in order to keep it. We keep it so that it goes on thinking.
What remains in our hands
Europe missed the platform race, and turned it into a ritual of self-flagellation. But you do not win yesterday's battle; you choose today's. The infrastructure of knowledge is one such battle — still open, its standards unset. And our supposed weaknesses — our demands for transparency, our culture of respect for singularities — become advantages here, because an architecture built for extraction structurally cannot satisfy them.
What remains, then, is the only choice that matters. Submission: adopting the tools that think in our place, becoming excellent at extracting value from them, at the price of a slow proletarianization of our own judgment. Or preservation: keeping what we know, and above all the way we know it — that tacit knowledge Polanyi named, which the document does not hold, and which a man's departure takes with him.
Knowledge is alive only on condition that it circulates. Our task is not to embalm it, but to see that what deserves to be passed on can be — verified, sovereign, alive. And that the decision of what deserves it remains, always, in human hands.
It may be the last thing a machine should never choose in our place.
This is the thesis (Urs) is built on: to preserve living reasoning rather than the dead document, to make every proof verifiable, sovereign and replayable, to augment the expert and the organization without ever replacing them. Explore the infrastructure →