Lost in Latent Space

Fontcuberta. Chatonsky. Eldagsen. Astray.

After Boris Eldagsen and I replied to Joan Fontcuberta’s musings on “algorithmic photography” with an article published by PetaPixel, Gregory Chatonsky replied to us with a theory that expands on latent space. Now it’s our turn again. Gloves are off, but it’s all in good fun. You can find Boris' reply on his website, and mine below.

Let me start by saying this:

Chatonsky’s train of thought is a fine piece of engineering—both powerful and aesthetically pleasing—and its derailment a tragedy, a waste of solid material and grand vision.

Inspecting the crash site, it looks like a case of overengineering at first: too much ingenuity in the details making the whole thing prone to a chain reaction of little defects. But the wreck is burying the real reason underneath: it ran on faulty tracks.

I’ll shift gears before I drive this metaphor off a cliff: Chatonsky’s underlying assumptions are faulty if we’re being generous, disastrous if we’re being honest. And all too eager to drive his point home, he rips through photography and its institutions—completely unnecessary casualties.

But this is all very pre-2022 down-to-Earth imagery. Let’s space out:

Gliding through latent space on his galactic assumptions, Chatonsky gets a little lost. Out there, he casually surrenders reality to an intellectual nihilism that is troublingly theoretic and aloof—detached from any axiomatic or at least pragmatic weight that could ground it—fantasizing that two things that look equally realistic are equally realistic (photographs and AI images), that distinctions are fictions (true and false), that uncertainty is absolute rather than relative (a photograph’s limitation to tell the whole truth is on par with an AI lie), and that a flawed system is a failed system (imperfect institutions are automatically obsolete). It’s all very out-there.

Leaning against Fontcuberta who already leaned too heavily on one side of an ontological yes-but-no debate—the truthfulness of photography—their hypotheses topple over in tandem. Reality, meanwhile, hides in the deep theoretic chasm between the two prevailing schools of thought—temples of belief really behind their facades, where photographs are cursed as devious lies or worshipped as holy truths. All too indoctrinated, one overlooks the obvious: the truthfulness of imagery is not absolute but measured in degrees, as I argue in my article Photography Is an Honest Lie, and without guiding a theory along that scale, it gets lost.

As a big fan of moderation to remedy polarization, I stand on solid middle ground with one foot in each camp, true to my chant Nuance or Nothing: I'm the first one to admit that universal and full truths are hard to find (in photography and elsewhere), but I'm also the last one to deny the existence of factual evidence (photographic and otherwise) society can and must agree on to stay intact. Those who don’t grant reality that bit of pragmatic nuance, end up arguing absurdly to keep their theories alive without realizing that they were stillborn.

Metabolism / Metastasis

And so AI “appears”—a deus ex machina saving the day—out of the thin air that cools our CPUs as they power the world (wide web), a logical consequence of digital circumstance. Or so it appears in Chatonsky’s tale, and if you are tripping over the double meaning hidden in the previous clause, I can assure you that both interpretations of “appears” are exactly what I mean.

Of course, those who made the genius genie “appear,” knew it was to grant more than the one wish Chatonsky utters: cure humanity’s media obesity. But let me not get ahead of the plot.

Chatonsky celebrates AI as a metabolic system that helps digest web-induced oversaturation. But metabolism is a zero-sum process: you don’t get more out of it than you put in. LLMs don’t just break down media to spit it back out as statistical revelations in equal measure—after devouring the training data, they can belch out an infinite surplus of slop, most of it even more flavorless than the junk Sam, Mark, and Elon force-fed them through social media funnels: mostly a blend of coffee and cats. They contribute, exponentially, to the very oversaturation they supposedly digests. If we absolutely must go organic and compare circuits to cells, that’s not metabolism—that’s metastasis, and you definitely want to keep an eye on its spread.

The multilateral institutions that have traditionally observed and stewarded humanity’s media flux—imperfectly, yes, but with sufficient degrees of dedication and benevolence for me to put them in their place between these two em dashes instead of exiling them in obsolescence—are needed now more than ever to filter the “mis” and “dis” out of information, to follow up on entrepreneurial promises, and to scrutinize pseudo-intellectual make-belief. Chatonsky roots for their discontinuation not because their function expired suddenly, but because he (alone) thinks they were ill-equipped to fulfill it in the first place. Their job, however, was never to master the flow of all media, but to ensure the circulation of relevant media. Publishers, museums, and press agencies were never in the business of channeling the fun-flow of our vacation pictures and the likes. AI slop as an unchecked inflow of sewage makes the whole stream swell, makes it turbid, makes it harder to treat, which is why we need to fortify these institutions, not take a wrecking ball to their legitimacy. Free spirits dislike gatekeepers, but we need them to keep a stampede of dubious YouTubers, TikTokers, and other self-proclaimed influencers from trampling investigative journalism and those who really have something important—and true—to say.

Chatonsky psalms could have us believe that AI is a “process [that] makes [reality] navigable,” but what places will the almighty algorithm get us to, if it shifts the very landscape it “helps” navigate before our eyes, and colonizes that fragile ecosystem with otherworldly fictions that distract our attention (or what is left of it)?

From Code to Vectors / From Process to Result

Signing a premature surrender, Chatonsky writes off traceability—and with it the institutions that follow traces—as a technical impossibility because the belly of the machine doesn’t adhere to deterministic causalities. But he is overcomplicating a simple matter: nobody is trying to trace what’s happening inside the CPU, only what comes out of it. We don’t need to understand the process to flag its tangible result. Traceability is not an abstract pursuit but an administrative obligation shared between the machine’s manufacturers and the content factories that employ it. MADE IN CPU. Consumers have a right to know where a products comes from—after all, we’re paying for it with our attention and internet identities—and for those giants of informatics who built the systems, it’s a tiny act to install markers of provenance. Of course, there will always be ways to circumvent the circumventable, but that cops-and-robbers dynamic applies to, well, everything, and inventive robbers make cops not less but more relevant.

Marveling at the mysterious little box with the big secret inside, Chatonsky describes the latent space it harbors as “a continuous field of forces where one can glide from one concept to another” with “no stopping point, no moment of capture,” always in transit along “a continuum of possibilities.” And getting so carried away by the transitional properties of AI image creation, he wants to surf that wave function past the crest, past the very real point where possibility collapses and certainty emerges: just like reality—at the quantum level—is a probability function enfolded inside a logic that eludes us, AI’s statistical wave breaks upon observation. The moment your LLM is done chewing a prompt, it flicks that bubble gum of possibility against your screen from the inside, and where it lands, elasticity takes a fixed shape for you to look at. Of course, you can let the machine chew again, but probability becomes entity at every interval of observation, and that’s when you can label it.

While Fontcuberta littered his theoretical picture with uneven parallels he drew between AI and photography, Chatonsky forces fat and distinctive black and white brushstrokes all over the hypothetical masterpiece this could have been. Maybe I can contribute a little color in the spirit of continuous modification: photographs—as well as paintings, songs, and even something as solid as sculptures—can be modified continuously as well, if not by the same means. They are not thaaaaat deterministic. Crop it a little more, add one more brushstroke, take out a note, sand the edges. It is only when we share a creative work with the world, that it becomes somewhat permanent. Boris’ Electrician, as an AI image, is no different in that. It hangs on a wall, labeled with an edition number and signature.

When we compare an AI image with a photograph, we’re comparing results. No matter how transitive or deterministic their procedural route, a result is a result, a tangible manifestation. One that can be iterated, of course, but one that is something, whatever thing, once it’s out there.

Traceability is a non-negotiable responsibility of the machine’s overlords, even if they are at a speechless loss to explain the inner workings of their own creation—a mind-blowing reason to regulate it, in fact. They have created something we understand maybe as little as it understands us, and Chatonsky's low level of discomfort, as he stares at his screen without flinching, can only be explained with the lazily reclined position of somebody who has given up before the fight even started.

The Inversion of the Graft / The Inversion of Reality

Only an intellectual could be so foolish to flip reality upside out and inside down for the mere sake of argument and provocation. A kid can tell that the nature of photography didn’t change just because AI latched onto it. Its nature is nature, after all—the natural world—and reality doesn’t disappear underneath the imaginary worlds we strap to its back.

Composting Foncuberta’s graft by reversing its logical chronology, Chatonky’s theory morphs more and more into fiction: a dystopian plot where photography becomes meaningless LLM food without an essence of its own. Where Fontcuberta already went too far, his disciple wants to go further to radicalize the theory, and he confesses as much in his manifesto.

Of course, here in the real world, exploiting a photograph as input for something else—may that be an AI image, a graphic, or a painting—doesn’t change the photograph and much less the genre of photography.

AI does not exceed photography the same way a fictional plot doesn’t exceed a nonfictional narrative. They are different genres. AI offers certain possibilities photography cannot afford. But LLMs are not plugged into the real world like a camera is, and that connection is meaningful. I will go out on a limb here and say that the latter will remain the preferred device to take on a vacation or have around on your wedding day. Sure, you could have AI hallucinate the ceremony, but that would miss the point by an insane degree.

An excess of visual possibilities alone is neither desirable nor valuable or meaningful. Scarcity is. And photographs become scarcer the more pictures AI produces. There is meaning in reality, in knowing that a photograph is made from something that really happened, a moment that occurred, a tangible piece of history. For all its poetic potential to manifest fragrances of the subconscious reminiscent of human inclinations, AI’s connection with us is indirect, via an outside world it cannot touch, that only we can touch to tell the machine about it.

That is the photographic essence escaping Chatonsky and hopefully no one else: photons. They are one of our most significant correlational bonds with the original world, long before they start and long after they finish nurturing AI’s statistical imagination. Channeling their light, the camera becomes a magical tool to slice time.

And AI’s essence? What does the machine run on to excavate those “visual truths nestled in the interstices of our collective memory?” The images we accumulated in our databases throughout thirty years of internet history were not equally truthful depictions of reality across the board. With the more recent age of social media contributing disproportionately and skewing the picture heavily, the training data is foremost a show of how we wanted to pass the world off to others: pure sunshine, glossy perfection, and personas rather than people, let alone the personalities underneath our flesh costumes. Who we are online isn’t who we are. These aren’t “memories” so much as they are reminders of how we want to be perceived when everybody is looking. AI’s essence, then, is makeup.

Chatonsky dreams that there is “no reason to believe that photorealism is inherently more “realistic” than the “realism of the possible” offered by AI,” because it taps into the subconscious, the imaginary, which is, indeed, deeply human. But what happens inside of us and outside of us are two things that we can absolutely and must definitely give different names. Our perception might become our personal reality, but there are universal facts societies agree on to keep their collectives of individuals from scattering.

Chatonsky pretends that there is no realistic difference between, let’s say, a photograph of the commander in chief stumbling up the stairs of Airforce One, and an image of him piloting a fighter jet. And while he couldn’t fool a child with that, I’m afraid some intellectuals might fall for it.

A realistic look does not make a fake real. A piece of glass with a brilliant cut does not become a diamond. But in Chatonsky’s media economy, the machine’s statistical averages are of equal if not higher value than the original world they draw from, and so reality is discounted below its actual worth. Of course, in reality, reality is precious, and more precious now. The more AI inflates that media economy, the more we will treasure human-made content.

For AI pretty much everything is possible—like in a dream—and that’s what distinguishes it from reality. The comparative scarcity of possibilities makes reality real, and realer than AI hallucinations.

The Era of Generalized Suspicion / The Era of Generalized Capitulation

To start a paragraph titled “The Era of Generalized Suspicion” with a misattributed quote could be a prank I’d tip my hat to, or maybe a self-fulfilling prophecy, or maybe just poor research. But I have another suspicion: a few months ago, a social media post slapped me in the face with my own face—or rather an AId interpretation of it based on a real photo—and the text underneath mixed up Boris and me. This here, reminds me of that then. Moreover, the language of Chatonsky’s superb text has AI written all over it. A little too smooth around the edges. More-more-over-over, why wouldn’t it be AI-assisted, given his praise?

If you want to get to the bottom of it, there are plenty of AI detectors in your clickable vicinity. How much you can trust their judgement is a good question, but the better question is this: if his text is AI-assisted, how many of its preposterous claims are attributed to a human’s obsession with intellectual attention, and how much of them accounted for by a hallucinating machine getting high on its own supply?

While that truth might remain mostly unknowable, this one here is fully certain: the quote Chatonsky smuggled into my Fontcuberta reply (“If all doubts paralyze us, those in power win”) isn’t mine. It isn’t Boris’s either, by the way. It is an overly generous paraphrase decorated with counterfeit quotation marks. The truth can be found anytime, not very hidden, in the PetaPixel article.

Chatonsky celebrates Fontcuberta’s generalized doubt as a win: finally we all realize that photographs have always been open to interpretations. But that was never a “truth we were hiding.” Those institutions never claimed to have a monopoly on the truth, the whole truth, and nothing but the truth. Nor did it take AI to find out—experts always knew and laymen still don’t know.

And I must insist on degrees once more: an image can lie a little or a lot. There’s a difference. Give nuance, give truth. Only because not every picture is proof of everything, we can't conclude that no picture is proof of anything. In the same vein, an inability to guarantee the authenticity of all pictures, does not imply that none can be authenticated.

The image might have never been full proof but it was usually proof enough for all practical purposes. The “interpretive battlefield” was much smaller, much less contested than Chatonsky makes it out to be. But AI has the power to expand it infinitely, and turn a battle into a war.

Fontcuberta’s and Chatonsky’s generalized suspicion is a generalized capitulation: “accepting undecidability.” But the veracity of an image is still within reach and we have to keep it there if we want to stay in touch with reality. We didn’t lose truth categorically in 2022; it is the process of finding it that became exponentially more difficult, thereby increasing institutional responsibility to hold those in power accountable. We don’t need to dig into all imagery out there to unearth the true nature of every single cat picture, but we must keep up with the propaganda of those in positions of influence and power.

There have always been degrees of authenticity, and they matter. A press photograph published alongside an explanatory caption, expanded on by virtue of a thorough journalistic text, published by an independent news outlet, inspires significantly more trust than some AI post churned out by a viral social media account with questionable motives. The messages these two pictures transport are neither equally trustworthy, nor equally true, and we can very well discern them, and have to do so without a doubt. Trust in the image has not collapsed altogether, and it doesn’t have to, and hopefully never will.

A picture that tells half the story is different from an outright visual lie, similarly to the difference between a staged fine art photograph that might tell a truth about the inner workings of a mind and heart and soul, and a documentary photograph that reports on the truth of an event that occurred outside of the photographer. There is a difference between optical truth and truth of meaning, and between pictures whose degree of truth is 1%, 50%, or 99%. If you want to take your chances, flip a coin, but don’t play the lottery.

Chatonsky is ironing these degrees of authenticity, until there is only the flat fabric of unknowability. But doubt is a means, not an end; the beginning of an investigation, not its conclusion. If neither veracity nor falsity follow it—either because we accept undecidability or because we overwhelm our institutions with AI slop—doubt becomes noise and we get stuck in limbo. A permanent alert of generalized suspicion doesn’t alert anyone anymore; it only causes fatigue.

My trust in today’s outlets of legacy media remains fairly firm, or at least my trust in the outlets I entrust with my attention. Meanwhile, social media has become such a playground for greedy stakeholder algorithms that it lost my trust and with it my attention.

To call the distinctions between a real and a fake image “institutional fictions” misses the fact that facts are facts—have been, are, and will be—and that AI images can be debunked as such if we make an effort. Chatonsky accepts, in advance, a world plagued by AI slop, and riddled with perpetual doubt.

Towards Multiplicities / Towards More Than Multiplicities

Democratization of LLMs, appropriation of computing power, roughening smooth AI-average-aesthetics—yes!, yes!, yes! (check out my second AI stunt for an opinion on algorithmic taste). But let’s not insist that a solution is the solution, or that these approaches are flawless and hazardless. Chatonsky bends with today’s intellectual tendency to synthesize reality’s complexity into elegant formulas, no matter how reductive.

We can aim for his multiplicities without giving up on institutional scrutiny. Multiplied latent spaces are not more realistic than unified traceability. So, while I can very much get on board at those last stops along Chatonsky’s train of thought, I'll get off before it rips through photography and institutions like cannon fodder because that is an unnecessary sacrifice to drive the point home.

To acknowledge the black holes in latent space—dense power concentrations and near infinite bias—and then call for more such spaces, is a call as daring as it is risky, especially if we are mistaken about how they operate and under whose control.

AI’s early mutations, like the six-finger hands on melting people, were glitches. Programming rectified these technical difficulties, yes, but the homogenization mirrored in AI’s contemporary aesthetical averages cannot be blamed solely on the manipulative engineering devised by obnoxious tech bros—we’re all complicit in it. We provided the biased training data when we uploaded but the most likable versions of the world and ourselves, polished superficiality and cosmetic beauty, a mainstream of generic and vain misrepresentations that don’t reflect reality’s perfect imperfections.

As for the power structures holding these machines hostage: we, the people inhabiting democracies, bring a lot of bargaining power to the negotiations—our merged monies and political influence. Without us as clients, corporations are penniless; and without us as voters, governments are powerless. The reason we give in to their popular and populist temptations so easily is simple and brings me back to the AI genie:

Contrary to the tale that Chatonsky tells, it grants more than one wish, and they all have a catch. Chatonsky wishes for AI to unfold reality’s fabric with its revelations. But it can also create for you and think for you. Those are three fat wishes right there, but all echoes of the same desire that has conjured every technology:

CONVENIENCE

Conveniences are obvious, but their consequences subtle—short-term comforts with long-term impacts: create a cute little flyer today, spiral inside a perpetual tornamitsunado of AI slop tomorrow.

Convenience is a loan. You pay it with interest, little by little, conveniently oblivious to the sum total. And AI’s interest rate is staggering, a bit of convenience at an exorbitant price: think less today, lose the capacity to reason tomorrow.

Seductive and addictive like any other drug, convenience becomes a costly habit over time, and as dependency tightens its grip, you unlearn how to live without it. That’s when corporations cash in, true to their motto: give convenience, get profit. The tech-trick is to offer those services gratis on paper, while charging us our marketable identity in the fine print—because we are too numb to care about footnotes and what's between the lines. While we finally begin to grasp how expansive and expensive our social media presence is—as those billionaires sell bits and pieces of us to the highest bidder—the notion that ChatGPT knows us more intimately than our mom or spouse is still buffering.

Genies only act as good or bad as the masters who prompt them. Now this one is out of the bottle, and I won’t contradict Chatonsky here. Shouldn’t we hold on to the bottle though, make sure its neck doesn’t break off with the genie out? And, on a final sidenote: if the genie fulfills all wishes so masterfully, is it still a servant?