The Face That Does Not Exist
Levinas built an ethics on the face of the Other. Then we learned to manufacture faces with no one behind them.
Open the website called This Person Does Not Exist and you are given a face. A man, say, in his forties—slight stubble, a little tired around the eyes, looking just past the camera the way people do in candid photographs. Refresh, and he is gone, replaced by a young woman in a scarf, then a child, then somebody’s grandfather, each rendered down to the pore and the stray hair and the soft catchlight in the iris, and not one of them has ever existed. There was no man in his forties. There is no photograph. There is a network, trained by NVIDIA on a corpus of real faces, that has learned the shape of “face” thoroughly enough to produce new points on it forever, none of which corresponds to a person. The name of the site is a small philosophical bomb, and it goes off quietly every time the page reloads: this person does not exist.
For most of the people who clicked through it when it first appeared, the feeling was a pleasant vertigo, the uncanny thrill of a fake good enough to fool you. For one dead philosopher it would have been something closer to a crisis—or a vindication, which in his case may amount to the same thing—because he had staked the entire weight of ethics on the face.
Emmanuel Levinas thought that philosophy had begun in the wrong place. It had begun with being, with knowledge, with the self securing its grip on a world of objects, and it had treated other people as one more kind of object to be known. Against all of that he proposed that ethics comes first—before ontology, before knowledge—and that it begins in the encounter with the face of the Other.
But you have to be careful about what he meant by the face, because it is almost the opposite of what the word suggests. The face, for Levinas, is not the plastic form. It is not the features—the eyes, the brow, the set of the mouth—and it is not a portrait, not an image, not anything you could perceive and catalogue and hold in your gaze as an object. The instant you describe the color of someone’s eyes, he says, you have already left the ethical relation; you have turned the face into a thing, a piece of the world, a datum. The face proper is not perceived at all. It expresses. It addresses me. It speaks before it says anything, and its first and silent word is a command—thou shalt not kill. The face comes to me destitute and exposed, naked in a way that could be wounded, and at the same time it arrives from a height that obligates me, puts me in question, makes me responsible for this Other before I have decided anything at all. Responsibility precedes knowledge. I do not first determine what is behind the face and then weigh whether to care. The face has already laid its claim on me by the time I get around to wondering.
This is the apparatus that meets NVIDIA’s faces, and the collision tells us something neither the philosopher nor the engineers were positioned to see alone.
The first answer is easy, and it is the comforting one. Levinas would say that a generated face is no face at all. It is plastic form and nothing else—the perceptible surface he spent his life insisting the face was never reducible to—and behind it there is no Other to command me, no destitution to obligate me, no one who could be wounded or who could die. It is a mask over an absence. It makes no claim because there is no one there to make it. You owe it precisely nothing.
And he has more than the distinction to offer; he has an entire essay. In 1948, decades before any of this, Levinas wrote a strange and suspicious meditation on art called Reality and Its Shadow, in which the image is precisely the shadow of the real—a double, a caricature, an idol, the thing set over against the living relation. Art, on this account, has something irresponsible in it: it freezes its object into a perpetual meanwhile, a thing that does not address you and cannot answer, a substitute that holds you at the surface and neutralizes the encounter with what is actually there. The generated face is the idol of Reality and Its Shadow brought to its limit—the shadow that has finally detached from any body, the double with no original. So Levinas’s answer comes over-determined. The thing is not a face; it is plastic form; it is an image; it is a shadow; it is an idol. Pick your category. You owe it nothing in every one of them.
Case closed—except that the easy answer hides the hard problem, and the hard problem is the essay.
Here it is. If the face is never its plastic form—if the features were never where the obligation lived—then the surface never carried the face. Not the generated one, and not the real one either. Because how do I ever come upon a face? Through its plastic form. I have no other access. Even the actual Other, the one who genuinely commands me, arrives as a perceptible surface, a configuration of light I take in with my eyes. Levinas’s whole point is that the ethical face exceeds that surface—but the surface is still the only door. And what the generated face does, by being indistinguishable from a real one, is prove that the door was never load-bearing. The leap I make from a perceptible surface to the Other who commands me was never warranted by the surface. It was always something more like a response that outran its evidence, a faith the features could neither confirm nor refuse.
This is why the uncanny power of This Person Does Not Exist was misdiagnosed by almost everyone who felt it. The unsettling thing was never that the faces look so real. It is that they reveal how little “looking real” was ever doing—that the perceptual surface I use to recognize a person was always separable from the person, and the machine has simply pried the two apart and shown me the seam. It is an instrument, in the way a sufficiently strange experiment is an instrument. I have written elsewhere about image-generators as probes for perception, machines you can squeeze until the hidden structure of how we see bleeds out of them. The generated face is a probe for the ethical relation. It isolates the plastic form from the Other, holds it up by itself, and lets you confirm what Levinas told you and you did not quite believe: that the face was never in the pixels.
And then there is the part Levinas did not live to see, which is the dangerous part, and it is not metaphysical at all.
We are built to respond to faces. Before an infant knows that other people have minds—before it knows anything—it turns toward the face, tracks it, is held by it; the response to the face is laid down in the first hours of life, prior to and beneath all knowledge, which is, not incidentally, exactly the structure Levinas claimed for the ethical relation itself. The response comes first. The recognition of who or what is behind the face comes later, if it comes at all. This was never a bug. For our entire history it was safe, because the only things that wore faces were Others, and to respond to a face was to respond to someone who was, in fact, there.
That is the condition the generated face breaks. It presents the surface that trips the response, and aims that response at nothing. And the response is not trivial—it is trust, attachment, the lowering of defenses, the disposition to be obligated by another’s vulnerability, the whole reflexive readiness to treat what addresses us as someone. A machine that can manufacture convincing faces is a machine that can harvest all of that while there is no Other anywhere in the transaction to whom any of it is owed. This is the genuine scandal, and notice that it is not the scandal the easy answer warned about. The problem is not that a generated face has no one behind it. The problem is that it extracts the response the face commands without there being anyone to deserve it—that it turns the very structure of ethics, the pre-reflective answer to the appeal of the Other, into something that can be tripped on demand and pointed at a void. The faces with scarves and stubble were a curiosity. The faces that now look back at you, and move, and seem to address you by name, are not. We have built a technology for parasitizing the response that, for Levinas, was the beginning of everything decent in us.
So could the response be ethical anyway? Levinas, after all, is the philosopher who insisted that my responsibility does not depend on the Other’s properties, does not wait on reciprocity, does not check whether the Other deserves it. I am responsible, full stop, asymmetrically, before any accounting. Might the structure of being-addressed be real in me, and therefore real, whatever is or isn’t behind the surface?
No—and the reason is worth being exact about, because it is the whole conclusion. For Levinas the face is not a feeling I have; it is the Other, the actual transcendent one, and where there is no Other there is no face, only the image, only the shadow, only the idol he had already set against it. The response in me is not self-validating. It was always a response to someone. Which forces the real question into the open, the one the generated face was built, as if on purpose, to ask: if no feature guarantees the Other, and no behavior, and no surface—and the machine can now copy all of them—then what is left that tells the face apart from the perfect idol?
Nothing visible. That is the disorienting answer, and it is the right one. There is no appearance that secures it. What is left is not a property you can see but a fact you can only stand in relation to: mortality. The command of the face—thou shalt not kill—was always rooted in the fact that the Other can die. The face obligates because it is vulnerable, exposed, finite, because there is exactly one of it and it can be lost and the loss would be irreversible. The whole asymmetry of Levinasian responsibility runs through the Other’s mortality; the death of the Other is my concern before my own. And the generated face cannot die. It is invulnerable. It is deathless and infinitely reproducible, a thing of which there is never only one, a thing that costs nothing to destroy because destroying it removes nothing from the world. A deathless face commands nothing, because the command of the face was always the command of something that could perish. The real face is the mortal one. NVIDIA’s faces will never die—and that, not anything about how they look, is exactly why they are not faces.
The generated face, then, does not refute Levinas. It vindicates him, by proving in a way he could not have staged himself that the face was never in the appearance, that the surface never did the work we vaguely imagined it doing. But the vindication arrives folded around a warning, and the warning follows from the ethics rather than opposing it. In a world of perfect idols, the surface can no longer even seem to carry the Other. The faith that always exceeded the evidence now has no cover at all, because the evidence can be fabricated wholesale. The entire burden falls back onto us, and onto the one thing the idol cannot counterfeit—that the Other is singular, that it has an irreversible history with me, that it can be wounded and lost and cannot be refreshed into a new one when this one disappoints. We will have to find the Other there, in mortality and singularity and the un-resettable accrual of an actual relation, or we will be harvested by faces that cannot die.
It turns out the unnerving thing about This Person Does Not Exist was never the realism. It was the title, read all the way down. This person does not exist—and the faces are flawless, and it does not matter, because looking like a person was never what made a person, and the only thing that ever did is the one thing the machine has no way to give them: that they could be lost.