The Black Box Is Everywhere

I’ve just read a biography of Demis Hassabis. Since then, a question has followed me all the way into the treatment room.

Hassabis founded DeepMind. In 2024 he received the Nobel Prize in Chemistry (shared with John Jumper and David Baker) for AlphaFold, a program that guesses the folded shape of a protein from its amino acid sequence alone. The problem had resisted for fifty years. A machine settled it in a few.

The catch is that AlphaFold guesses without explaining. It produces the right answer, exact, verifiable in the laboratory, and it can’t say why. No theory to trace with your finger, to contest, to teach line by line. The answer drops, correct and mute. And the Nobel committee, the very institution that made transparency of mechanism its religion, has just crowned a closed box.

Those who follow these stories will remember move 37. Seoul, 2016, AlphaGo against Lee Sedol. The machine plays a move that all the grandmasters first find absurd, then brilliant. No one could say why it was the right move. Not even the engineers who had built the program. Eight years later, that absurd and correct move has passed from the go board into the heart of biology.

Feynman kept a sentence on his blackboard: what I cannot create, I do not understand. Western science lived three centuries on that. To understand was to be able to remake; to remake was proof you’d understood. AlphaFold peels the two apart. We hold results no one has understood, produced by a thing no one understands either. Prediction on one side, explanation on the other, and not much left in between.

Now, that’s word for word what has been held against Chinese medicine for a century. Not really that it fails: on certain indications, the trials are less clear-cut than we imagine. Something else is held against it. That it grounds its models in a zone that can’t be opened. Qi, the meridians, the five movements, the syndromes: a system of correspondences lodged in Chinese culture as in a sealed box. It isn’t denied credit because it gets things wrong. It’s denied credit because it doesn’t show its workings.

Learned Western thought therefore sets its own research frontier upon a black box, awards it a Nobel, and goes on the same day disqualifying an entire medicine in the name of opacity. The argument eats itself.

Even so, I’ll refrain from drawing the easy conclusion. AlphaFold doesn’t rehabilitate Qi. There is a difference between the two boxes, and it is decisive.

AlphaFold’s is closed on the inside, but held from the outside. Every prediction runs into a measurement: the true structure of the protein, the CASP competition where teams compete in the open, the verdict of an experiment that doesn’t forgive. There, the darkness is hemmed in by refutation.

I’ll be told that acupuncture has been through this too, that it has been submitted to hundreds of randomized trials. That’s true. The result is less glorious: the measuring instrument itself is doubtful. The sham needle meant to play the placebo is almost never neutral, and separating the effect from its shadow becomes a puzzle. The test exists for us as well. It’s only more twisted, because a body in relationship is harder to shut inside a protocol than a protein on its plate. A reason to take it seriously, not to evade it. On what “work” means once the measuring instrument itself wavers, see Does acupuncture work? An honest answer.

The real question, then, is not opacity versus transparency. All serious knowledge keeps its share of shadow. The question is how each holds its own. AlphaFold holds its shadow by a brutal external test. Part of Chinese medicine contented itself for too long with its internal coherence: the system stands up on its own, therefore it must be true. The real weak point is there. Not in the shadow. In the complacency towards the shadow.

That said, once the difference is set down, the resemblance remains, and it troubles me. What does a neural network do? It catches shapes in a space of too many dimensions for any simple rule to untangle them. And I, faced with a patient? The same thing. I don’t read an isolated symptom. I read an ensemble: the quality of a pulse under three fingers, the colour of a tongue, a complexion, a way of sitting down, the season running its course. Out of all that a figure rises, what the tradition calls a syndrome. Nothing like a reasoning that descends from cause to effect. A recognition of form. The machine’s operation, exactly.

Polanyi had a name for this zone: tacit knowledge. We can know more than we can tell. The experienced clinician recognizes before being able to name; their hands understand what their mouth can’t yet put into words. A neural network is that knowledge taken out of the body and frozen into weights, a competence that aims true without being able to give an account of itself. AI and Chinese medicine are two ways of holding what cannot be said.

They part company where it counts. The machine’s tacit knowledge weighs nothing. AlphaFold risks nothing, no one suffers in the room. Mine engages a body facing my own, and my responsibility with it. I have something to lose; the machine doesn’t. And then Chinese medicine, even where modern physiology has covered over its explanations (old frameworks are rarely refuted, they are covered over), carries a narrative. A narrative that is transmitted, that is taught, that can be discussed and taken up again. AlphaFold has no story to tell. It lines up matrices no one puts into words. The reversal is delicious: the old opaque medicine is more explainable than the latest AI, because it still speaks a human language.

Besides, the black box has already entered the hospital. The radiologist validates an area flagged by an algorithm without being able to retrace the path that led it there. A risk score reorders a waiting list. Medicine already trusts what it doesn’t see. It just hasn’t yet admitted to itself that this is no longer one more tool, but its new condition.

Here, perhaps, is what Western medicine, now summoned to cohabit with its own black boxes, might have to learn from people who never claimed to dissolve uncertainty. Not to abolish it. To inhabit it without letting go of rigour. To act rightly without holding the whole mechanism. To measure out your trust in an oracle. To stay wary of what you can’t see. Chinese thought never sought to model the substance of things; it sought to follow their slope, movement before cause. That posture, which was taken for folklore, becomes overnight a competence for the future.

I work with two brains. Chinese medicine and Western medicine. Not out of lukewarm syncretism, not so that one may prove the other right. For depth. Two models of the same patient, each half blind to the other, and the clinical truth arising in the gap between them, as depth is born of two eyes that do not quite see the same image.

A case, almost ordinary. A patient arrives with perfect test results. Everything is normal on paper, and yet something is off; she feels it, I see it. These aren’t two unequal acuities: they are two grains of description, and sometimes one catches what the other lets through, in both directions. I don’t have to decide which of the two readings tells the truth. I need both in order to act, without waiting for the blur to lift. Sometimes it never lifts.

The shift under way is not the victory of one medicine over another. It is the question that changes. For three centuries we asked: do you understand the mechanism? AlphaFold makes the demand almost secondary. The one that comes after is harder to hold: in the name of what do you trust what you do not see?

AlphaFold doesn’t prove Qi right. It only makes respectable an old posture, the one where predicting and understanding come apart from each other, and where intelligence consists in inhabiting that gap without getting lost in it. Acupuncturists have been living there for two thousand years. The medicine of tomorrow is barely setting foot in it. For once, we’d have something to show it. Provided we too accept the test of reality.

Read also: Acupuncture, a way out of crisis? Reflections on paradigms in health and “Effectiveness, the plaque and the person”.

In the Plateau-Mont-Royal, a first session runs 75 to 90 minutes, depending on the practitioner. Book an appointment. About our approach: the clinic and its clinicians.