Responses to “A Necessary Critique Of Fontcuberta’s Algorithmic Photography”

Published on chatonsky.net on April 13th:

“The arguments made by Eldagsen and Astray against Fontcuberta’s hypotheses seem to rest on an ontological conservatism that misses the ongoing epistemological revolution. By clinging to a binary distinction between “light” (photography) and “code” (AI), they commit a fundamental category error. They perceive generative AI as an informatics of instruction—software executing a calculation according to rules pre-established by a model—whereas we have shifted to an informatics of vector navigation. Fontcuberta understood what many still refuse to see: the image has never guaranteed authenticity. By proposing a wisdom of doubt rather than restoring certainties, he opened the door toward a post-photographic epistemology. Yet this opening remains to be pursued: it recognized undecidability without drawing out its full political and ontological consequences. This text extends that fundamental intuition by exploring what is at stake beyond generalized doubt.

Metabolization
The Web has triggered an unprecedented media inflation. This saturation has transformed the status of photography. It is no longer an isolated act of capture, haloed by singular value. It has become a surplus resource, one datum among billions that humans can no longer perceive in its integrity. This is the condition of hypermnesia: remembering becomes impossible because there is too much to retain. Thirty years of the Web represents thirty years of the silent accumulation of images in databases—images tagged without consent, metadata piled in invisible strata.

It is precisely within this context of intractable saturation that AI appears. But it does not appear as a threat external to institutions; it is the symptom of their obsolescence. Institutions never truly had the power to master this flow. They believed they were organizing scarcity. The Web revealed that there was no longer any scarcity to organize. AI absorbs this deluge by transforming it into a continuous multidimensional topography: latent space. It digests this massive flow according to a logic very different from copying or simulation. It extracts symbolic forms—not some “truth,” but statistical correlations that reconstruct the world as it has been represented by billions of individuals in their daily practices of sharing images.

This is a chemical process in the strict sense: AI does not copy reality; it fractionates it and makes it navigable according to a logic that escapes traditional categories of the discrete. This is precisely what we observe when working with these latent spaces: how images cease to be finished objects and become transit points in a continuum. When we ask a diffusion network to transform one image into another through interpolation, we witness a morphing that has no photographic equivalent. It is not a fusion of images. It is a traversal of the possible within a geometry that eludes us. And in this traversal, categories collapse: we no longer know if we are creating or discovering, inventing or awakening what lay dormant within vector coordinates.

This process is radically different from what Eldagsen and Astray describe. They imagine they can preserve the photography/AI distinction by strengthening institutions and tracing processes. But this fails to recognize that institutions have always been structurally incapable of mastering this flow. They merely organized its invisibility. The Web made it visible. AI is the direct consequence, not an accident to be corrected. Metabolization means that AI is not an intruder. It is the reaction of a technical system to a sensory saturation that has become intractable by old logics. To refuse to see this is to cling to the fiction of an epistemic scarcity that has already evaporated.

From Code to Vectors
Eldagsen and Astray implicitly adopt the distinction between two conceptions of AI: either an AI that executes explicitly programmed instructions (code as recipe), or an AI that emerges from learning (code as hidden logic).

But this opposition itself is obsolete. What has actually occurred is a shift from an informatics of instruction to an informatics of vector navigation. This shift is not a technical refinement; it is a logical rupture.

Classical informatics of code consists of a series of instructions written by humans, executed deterministically, producing a predictable result. This is computational logocentrism: the belief that code is transparent—that it can be written, read, modified, and mastered. Eldagsen and Astray remain prisoners of this conception, even when they admit to AI’s opacity. For they still expect traceability to be possible—that someone could, in principle, understand the process. This is a pre-computational belief: that everything can be made visible through intellectual effort.

But generative AI does not function this way. It does not manufacture an image according to pre-established rules. It locates an image in a multidimensional space of probabilities whose dimensions emerged from the learning process without prescriptive human intervention. The true code—the architecture of the neural network—is merely the tooling to create the conditions for navigation. What matters is not the programmed logic. It is the probabilistic topology that emerges from the process, autonomous and irreversible. Once the network is trained, it cannot be “unrolled” like a film. The parameters are fixed, but their interactions remain inaccessible to linear reading.

They insist: “Photography is written with light; AI imagery is written with code.”

This is a seductive formula, but a false one. Neither is “written.” Writing implies a linear intention, a traceability of the gesture. Photography is an optical capture—simultaneously passive and active—subject to the physical presence of the real. Generative AI is not written: it is vector navigation, a learning of the fold of latent space. These are not two variants of the same act of composition. They are two radically different ontologies, two regimes of meaning.

When we generate an image, we are not calculating anything in the classical sense. We are traversing a continuous latent space along trajectories that were not mapped out in advance. The image does not result from a formula. The image emerges as the actualization of a possibility that existed as vector potentiality, immediately. This is the abyssal difference between a discrete logic (code: 0 or 1) and a continuous topology (latent space: an infinity of gradations). And this difference changes everything—it changes not only how we produce images, but how they produce us in return.

Photography is a medium of the discreet: a click, an instant, a single viewpoint, an immobilization. It captures according to a binary logic: this moment existed; this event took place or did not. Latent space functions according to a different logic: it is a continuous field of forces where one can glide from one concept to another without rupture. There is no stopping point, no “moment of capture.” There is a continuum of possibilities.

This is why they cannot be compared by saying they are just two different means of achieving the same result. They proceed from incompatible ontologies. In photography, there is a rupture between what was captured and what was not. In latent space, there are only degrees of probability, topological proximities, and seamless continuities. This is also why the Eldagsen/Fontcuberta opposition is false: they are arguing over what name to give something, when the problem is not the name. It is that the very ontology of the “visual” has changed.

The Inversion of the Graft
Fontcuberta rightly uses the metaphor of the graft: AI grafts itself onto photography, transforming it and changing its nature from within. Eldagsen and Astray implicitly accept this causal direction, as if AI were a disturbance coming from the outside, infecting a pre-existing system.

But we can not completely reverse this argument: it is not AI that grafts itself onto photography. It is photography that has become an anachronistic graft of the AI ​​system. The logical chronology is reversed.

Traditional photography, frozen in its optical capture and its presuppositions of authenticity, is now merely a mode of input injected into a system that radically exceeds it. It certainly provides the initial coordinates—the training data—but it is the latent space that deploys its metamorphic potential. It is no longer the producer of meaning; it is the digested material.

The photographic image thus becomes one archive among others, a trace in the vectorized memory of AI. Its status does not change gradually; it collapses categorically. It shifts from “proof of a captured reality” to a “starting point for the generation of possibilities.” It is absorbed, metabolized, and recombined according to a logic that has no common measure with the photographic process. And in this absorption, something of its essence escapes—or rather, it discovers that it never had an essence, only forms.

The photographic accident—subject to the hazards of the physical real, the raw contingency of the moment, to that which refuses to be seized—is now replaced by the vector accident: an unpredictable drift in the multidimensional curvature of data that reveals visual truths nestled in the interstices of our collective memory. This vector accident cannot be predetermined. Nor can it be mastered. It emerges from the navigation itself, as an encounter between the navigator’s intention and the unknowable topology of the space. It is an accident that is only an accident for us, not for the system that generates it; for the system, it is simply the actualization of a virtuality contained within its structure.

Thus, Eldagsen was right to refuse the Sony award, but for the wrong reasons stated publicly. He should not say, “AI steals the prize from photography by mimicking it better than itself.” He should say: “Photography no longer exists as an autonomous ontological category. It is a graft of AI. And I refuse this prize because accepting it would mean admitting that I still believe in a distinction that the technical system has already made impossible.”

Only on this condition would his gesture of humility be honest.

The Era of Generalized Suspicion
Eldagsen and Astray see the generalized doubt toward images as a crisis to be resolved. Astray worries: “If all doubts paralyze us, those in power win.”

But this is an misunderstanding of what is happening. Generalized suspicion is not a crisis. It is an inevitable clarification. Since AI has metabolized the photographic aesthetic to the point of making it indiscernible from optical reality, trust in the image collapses. But this collapse does not mean we have lost the truth. It means we are finally discovering a truth we were hiding: the image has never been proof. It has always been an interpretative battlefield.

This disturbance manifests as a visible double crisis:

On one side, synthetic images insert themselves into the social field by passing themselves off as captures of the real.

On the other, authentic photographs are contested, victims of a collective paranoia that mistakes them for algorithmic generations.

But what Eldagsen and Astray interpret as the collapse of distinctions, I see as the revelation that distinctions only ever existed as institutional fictions. Recent controversies in art competitions are merely the visible symptoms of this clarification. They are not accidents. They are the exposure of what had always been hidden: that the image is never proof, never a guarantee of authenticity.

Institutions believed they mastered this authenticity. They simply mastered a consensus. And this consensus is now collapsing because latent space has shown there was never anything to master—only probabilities to navigate. To refuse this suspicion by calling for the strengthening of institutions is to refuse to see that institutions are precisely what collapsed suspicion through power, not through clarity.

Eldagsen and Astray ask the wrong question. They ask: “How to distinguish? How to preserve? How to restore trust?”

The real question is: “Who controls the latent space? Who has the power to parameterize alignments, to choose datasets, to decide which visual possibilities will be generatable and which will remain unthinkable?”

This is a political question. Not technical, not institutional—political. It directly engages the very possibility of what an image can express, what it can show, and what it will never show.

For centuries, photography seemed to guarantee a certain democracy of representation: anyone could, in theory, take a photo, publish it, and challenge dominant images. But this was a productive illusion. The power to control the image had moved to institutions: publishers, museums, press agencies. At least one could criticize them, occupy their spaces, and contest their selections. We knew where the power resided.

Now, this power has volatilized and reconcentrated at a more fundamental level: the control of latent space itself. A tiny number of technological corporations absolutely control the datasets, the algorithms, the learning parameters, and the final alignment of the models. They do not control a collection of images. They control the ontological conditions of possibility for what an image can be.

And this mastery is structurally invisible. When Meta or OpenAI decides that a certain representation will be “aligned” and another not, we are no longer debating at the level of images. We are debating at the level of vectors—a domain where only the engineers of these corporations can navigate. The latent space of commercial platforms is closed. Datasets are proprietary. Alignment is secret. And yet, billions of individuals dream through these latent spaces, believing they communicate through their generated images, unaware that they are only actualizing the possibilities that a few algorithms have decided to be thinkable.

Calling to strengthen institutions in the face of this problem is like calling to strengthen the coast guard against a rising tide. The problem is not a failing distinction between photo and AI. The problem is that the mastery of the collective imagination has concentrated in the hands of proprietary algorithms, and this concentration has become invisible precisely because it no longer works at the level of visible images, but at the level of vector possibilities.

Toward Multiplicities
Eldagsen and Astray defend an epistemic order that has already collapsed. Fontcuberta proposes accepting undecidability and cultivating a wisdom of doubt. Even if Fontcuberta’s approach seems more accurate, both positions perhaps miss the true stake—not because they are false, but because they are politically insufficient.

What is needed is not to restore the photo/AI distinction. Nor is it to passively accept doubt as a final horizon. It is to accept the collapse as a political condition to invent other practices—artistic, pedagogical, political—that multiply latent spaces so that none can dominate.

For as long as a single latent space controls the majority of visual generations, as long as Meta, OpenAI, and a few others alone decide the conditions of the visible because they possess the computing power and have appropriated the means of production for profit, we have not solved the political problem. We have only moved it from the level of institutions to the level of vectors. True liberation would be for visual possibilities to fragment radically, for incompatible latent spaces to develop in parallel, so that no one can impose a single grammar of the visible. Not out of nostalgia for lost creative autonomy—that nostalgia is also a trap—but out of tactical necessity: the plurality of latent spaces is the only guarantee against the totalization of meaning. This can only be achieved through the collective appropriation of computing power.

The flaws of older generations of AI—their hesitations, their monstrosities, their undomesticated strangeness—constituted precisely their aesthetic and political virtue. These flaws were fissures where the unpredictable found a place. Their gradual disappearance in favor of a standardized, polished, invisible realism is not technical progress. It is the programmed homogenization of our collective imaginary, the methodical closing of possibilities in favor of a statistical average that corporations find manageable and monetizable.

Against this reduction, we must cultivate the accident, the divergence, the defamiliarization. Not to restore a lost authenticity—that authenticity never existed—but to multiply the possibilities within the technical system itself that tends to reduce them. This work is not innocent. Nor is it totally free. But it is necessary. And it presupposes a certain form of humility: recognizing that we navigate latent space without mastering it, that we explore possibilities without presuming we create them, that we resist without the certainty of victory.

The only honesty today is to accept that there are no more distinctions to restore, only spaces to fragment, possibilities to multiply, and a silent war to be waged against vector homogenization. It is in fidelity to Fontcuberta’s intuition—but by radicalizing it—that we reach this conclusion: his generalized suspicion is not an end in itself, but the starting point for a political transformation.”