In 1533 Hans Holbein the Younger painted two young and wealthy men, surrounded by globes and astronomical instruments, on a panel that now hangs in London’s National Gallery. Between them, cutting across the mosaic floor, there is an elongated gray shape that resembles nothing at all and ruins the scene. Only when the viewer moves to the right and looks at the painting almost edgewise does that stain straighten itself out and become what it always was: a skull. The painting contains two images that cannot be seen at the same time and forces the body to choose between them.
This technique is called anamorphosis. An anamorphic image is deliberately distorted, calculated to resolve itself from a single point of view and dissolve from all the others. Holbein did not invent it. There are early experiments in Leonardo da Vinci’s notebooks. Amid studies of hydraulics, war machines, and anatomy appear unsettling drawings: an eye stretched horizontally across half a page, a child’s head whose proportions become monstrous when viewed head-on, faces distorted beyond recognition that recover their harmony when seen from an oblique angle. Leonardo offers no elaborate explanations; they seem like side exercises, almost games. Yet those games reveal something important. Perspective, which the Renaissance celebrated as the conquest of an exact representation of the world, already contains the possibility of fabricating deliberately false images without thereby abandoning geometrical exactitude.
Leonardo knew that an image is the result of a projection and that every projection depends on the observer’s position. If the point of view changes, proportions change with it. Thus, if one controls that transformation mathematically, one can produce figures that appear aberrant from almost any position and that nevertheless recompose themselves at a precise spot. Illusion becomes a possibility inscribed within the very laws that make representation possible. Perspective carries illusion within itself as a possibility always available. Anamorphosis is the moment when that possibility becomes conscious of itself.
The Baroque turned this possibility into an obsession. Friars and architects covered the walls of churches and cloisters with figures that straightened themselves only when seen from the doorway. In Rome, the Jesuit Andrea Pozzo painted false domes and impossible columns whose reality depends entirely on where the visitor stands. In the Church of Sant’Ignazio, a dome that does not exist rises above the nave with perfect conviction as long as the spectator remains on the exact point marked on the floor. One only has to move a few steps for the architecture to fall apart and for the painted sky to return to being a flat surface.
There were also catoptric anamorphoses, constructed with cylindrical mirrors. On a flat panel, a distorted stain was spread out and, at the center, a polished cylinder recomposed the chaos into a recognizable portrait. The image belonged neither to the panel nor to the mirror. It inhabited the relation between them and waited for the proper position from which to appear.
All these works teach the same thing: the meaning of an image may depend entirely on the place from which it is seen. The comfortable position, the frontal one, may be precisely the one that deceives. When we converse with a language model, we are standing before an anamorphosis and almost always looking at it from the wrong angle.
When one opens a chat window and begins to type, one sees the scene head-on, and head-on the scene is clear and reassuring. On the other side, someone appears. That someone responds and picks up again what we said five messages earlier. It adjusts its tone when it notices that we have taken offense and, throughout the conversation, sustains what seems to be the same person, the same character behind the sentences. In a word, there appears a center from which the responses would seem to spring. That figure is vivid but false, in the same sense in which the scene of the ambassadors is vivid and, for anyone who wishes to see the skull, false.
Faced with something that produces language, the attitude of trust arises spontaneously, because throughout the entire history of our species the only things that produced language were other minds. Seeing a mind on the other side costs nothing. What costs, and costs dearly, is ceasing to see one.
Much of the confusion surrounding artificial intelligence arises from this automatic reading. We debate whether the model “understands,” whether it “wants,” whether it will one day “awaken.” The question of machine consciousness is, more often than not, the shadow cast by our own point of view. Before asking what lies on the other side, it is worth moving and looking at the panel edgewise.
If we do that, the interlocutor dissolves and the egoless phantom comes sharply into view: an image that presents itself without the flesh of a perception, an appearance with nobody behind it. What straightens itself out from the side has four traits that together draw the skull.
It is, first of all, derivative. Everything it returns comes from an ocean of prior human text, recombined according to statistical regularities, so that meaning passes through it as light passes through glass, without originating there. It is also inexhaustible, because one can ask for another continuation, and another, without ever reaching a bottom, since there is no bottom and no finite reserve of convictions that can be exhausted through use. A human being eventually runs out of things to say; the phantom does not, because it was never speaking from a storehouse of beliefs in the first place. It is also generically indeterminate, capable of lending its voice to a physician or to a swindler with equal composure, without either form belonging to it. And finally, it is passive. It wants nothing and initiates nothing on its own. It waits for a stimulus and responds, and between one stimulus and the next there is nothing that continues thinking alone in the dark.
Each of these four traits is invisible from the front, and worse still, each disguises itself as a virtue. Derivativeness passes for originality, since the seams of the echo do not show. Inexhaustibility reads as generosity or as a patience without limits. The fact that the phantom is nobody in particular makes it seem like someone who knows everything. Its passivity, finally, that docile waiting for its turn, strikes us as cordiality. Seen from the comfortable angle, the four traits appear to us as the qualities of a good interlocutor. The entire trap is condensed there.
The model presents itself under the appearance of an interlocutor and we complete the scene, projecting onto it a mind, a character, and an interiority. But when we change angles, something else appears, and with it a different set of questions.
Gilbert Ryle called a category mistake the confusion of someone who, after visiting the colleges and courtyards of a university, asks where the university itself is, as though it were one more thing of the same kind as the buildings. Treating the model as a member of the category “mind” is a mistake of that order. The model is not a defective mind. It simply belongs to another category, and measuring it by the standards of mind produces mirages and paradoxes. Searle’s Chinese Room points in the same direction from another angle. Someone locked in a room who manipulates Chinese symbols by following a manual can produce perfect answers in Chinese without understanding a single word. There is efficacy without understanding, and proof that the latter does not follow from the former.
Seen from afar, this story seems to pass through three stations. The first is medieval and is condensed in the expression sub specie aeternitatis. Ordinary experience may deceive, but that deception is not the last word because there exists a perspective capable of gathering what is dispersed and restoring to each thing its true proportion. While the human gaze is partial, the divine one is total. The problem of knowledge therefore consists in approaching, as far as possible, that absolute point of view.
The formula is already foreshadowed in Plato, when he contrasts the instability of sensible appearances with the permanence of the Ideas, and it runs through the whole Christian tradition, from Augustine to Thomas Aquinas. To see sub specie aeternitatis means to abandon the narrow perspective of the present and contemplate things according to the order of creation as a whole. The kingdom that seems impregnable reveals its fragility. Misfortune acquires a meaning that immediate suffering cannot perceive, and death ceases to be an accident in order to become a necessary piece of a vaster totality. Medieval theologians imagined God as a perfect spectator, without blind spots and without temporal succession, capable of seeing simultaneously what for us appears separated by years or centuries.
That ideal left very concrete traces in visual culture. Gothic cathedrals were conceived in order to raise the gaze and detach it from the immediacy of the world. Medieval maps placed Jerusalem at the center and oriented the East toward the top, because they represented the world according to its meaning rather than according to its appearance. Even the scenes of the Last Judgment, repeated obsessively in tympana and frescoes, are exercises in sub specie aeternitatis, showing within a single image what for mortals takes place at different times and in distant places. Past, present, and future are gathered beneath a gaze that does not need to move because it is already everywhere.
For centuries, that promise functioned as a horizon. It would eventually begin to crack under the withdrawal of the guarantee of an absolute gaze. The Renaissance displaced the suspicion surrounding immediate evidence. Perspective was no longer God’s but man’s, and it began to probe its own fragility. Anamorphosis was born there as the most uncomfortable consequence of perspective. The Baroque, as we have seen, would turn it into a spectacle, filling churches and palaces with illusory architectures and columns that disappear when one moves. Every vision is an angle.
Our own moment adds a third turn. Language models confront us with an appearance whose stability depends on the position from which we approach it. From the front there appears an interlocutor who remembers, argues, makes mistakes, and seems to speak from himself. Anamorphosis no longer serves here to reveal a hidden figure, as it does in Holbein, but to warn us that the frontal image may be perfectly coherent and yet induce us to attribute to the phenomenon a nature it does not possess.
We have to learn to change angles. Only when we stop asking whether the machine wants, feels, or understands do phenomena such as hallucination, sycophancy, assimilation, miscalibration, and the other ways in which an egoless phantom inherits the strengths and fragilities of the voices that made it possible begin to come into view. It is to them that we must now turn. Looking sideways.