In Defense of AI Slop

The question of AI slop at this point in the development of generative AI (GAI) is, at this point, unavoidable. Many, if not most people, see this consequence as a flaw of the technology that renders it ineffective and therefore useless. To put it succinctly, AI slop is the output of these engines that not only has a tenuous connection to reality but also lacks the creative merit of a human artist. While I don’t necessarily disagree with the judgement concerning its connection to reality it is my contention that this is not an altogether bad thing.

Eryk Salvaggio in his newsletter Cybernetic Forests recently did a six-part series on the nature of AI Slop that is absolutely worth reading. In part two of the series, he goes about providing an exhaustive definition of AI Slop that includes its presence and effect on social media:

AI slop breaks down the inquiry and investigation into the world as it is, replacing the critical landscape with text and image fragments that affirm the world as it is imagined. In essence, it circumvents any desire to understand the world because it offers us the immediate satisfaction of having a feeling about the world.

He also shares this quote from a paper by M. Beatrice Fazi:

Large language models lack what linguistics calls a referent. Reference here is the relationship between a linguistic expression and what, in the world, that expression is supposed to represent.

It is worth reading his post in full, but his argument generally speaking is that since AI Slop has no real connection to reality, it has no relationship to truth, and therefore represents an aestheticization of its political content that ultimately pacifies its consumers.

The characterization of AI Slop’s relationship to reality makes sense to me but I don’t agree with the notion that the outcome is necessarily a suppression of creative spirit on behalf of its audience. Instead, its disassociation with reality is the strength of GAI and exactly what needs to be exacerbated rather than “fixed.” So much of the discourse pertaining to AI sees it as a failure because so much of what it produces is either a regurgitation of scraped content or a hallucination that can be described accurately as nonsense. These kinds of outputs do not need to be fixed because they are not flaws of the technology but its strengths.

Bearing this framing in mind, slop, or AI output more generally, does not pacify its audience; in fact, as with any other media, it opens up a circuit of influence and exchange between the audience and the content itself. This circuit does not terminate with content that bears no resemblance with reality–it either tightens it or it continues to spiral out expanding the initial conditions of this feedback loop. Most of this criticism arises from the notion that data holds a diminishing relationship to that from which it was extracted and therefore to the world at large. This should not be up for debate; data should be understood as a facsimile of its referent. To be a mirror of the world should not be its expected function. In fact, the reduction of the world into an incommensurable mosaic of the world is what contributes to the utility of the gathered data in the first place. The intention should not be to reflect the world but to expand the limits of what can be thought–even if what becomes possible thought is a deranged, mutant, version of the world.

One way to look at data produced and used in this way is as the decomposition of the world or the self into discrete elements vis-a-vis extraction. The extraction of data as a form of mediated representation of the self or world arguably can be considered reduces the individual to a set of indexable classifications across an ecosystem of bits. Self and world are reconstituted as approximations in relation to how that data is encoded and submitted to a larger training set for future machine learning. Perhaps not an exact analog, but Simone Weil in her conception of decreation argues, “[w]e participate in the creation of the world by decreating ourselves.” The desire to retain what is human about ourselves, our unwillingness to relinquish what makes up our identity keeps the world in stasis because we want to remain earthbound creatures committed to a certain image of humanity predicated on agency and control.

Once again, though, Amy Ireland and Maya B. Kronic have developed a useful set of analytic tools for understanding these phenomena through their formalization of Cute Accelerationism. More recently they were interviewed in Spike Art about the general thrust of their manifesto and they explained what they saw as the future of the self contra past post-apocalyptic visions of it:

The future isn’t sending robots to hunt us down and kill us; it’s already inside us, turning us into cute objects as we tumble inexorably toward the great asymptotic kitten outside history. It’s the explosion of integral selfhood into a database of libidinal fragments, or “moé-elements,” that displaces binary gender: It’s the interminable pursuit of the obscure state of “just-rightness” that is inherent to cuteness – along with the temporal distortions peculiar to all of these things.

Generally speaking, these databases of “libidinal fragments, or ‘moe-elements,’” comprise the databases upon which AI engines are trained. It is an automated process of decreating the individual and rebuilding content through which individuals can re-imagine themselves. Data in this context can be best understood as the fragmentation of individuals into indexable moe-elements that can be recombined or reconfigured in sundry ways. Of course, it can’t be ignored that how this data is classed and organized is inescapably shaped by the culture(s) in which it has been produced. It is a deresolution of the human that favors classification over complexity. In a word, a synthetic world system is generated that in this extended form can be subject to manipulation. Consider Vilem Flusser’s examination of photography in his book Towards a Philosophy of Photography:

The photographic universe is a means of programming society–with absolute necessity but in each individual case by chance (i.e. automatically)–to act as a magic feedback mechanism for the benefit of a combination game, and of the automatic reprogramming of society into dice, into pieces in the game, into functionaries

Earlier he described the function of cameras as producing symbols and that “every photograph is a realization of one of the possibilities contained within the program of the camera. Therefore, what can be produced through the camera is restrained through its ability to capture it as a photograph: “The number of such possibilities is large, but it is nevertheless finite.” He argues that cameras function as a “combinatory game using number-like symbols” which increasingly mechanize thought therefore extending human reliance upon them. Further, these symbols are a gestalt of abstract concepts formulated by humans and encoded into them. Ultimately this results in a process of feedback that contributes, to the detriment of the human, to the improvement of the camera. According to Flusser, humans forget images are a representation of the world and begin to interact with them as if they are coextensive with reality. They are no longer decoded as representations but are internalized and subsequently re-projected out into the world:

The technical images currently all around us are in the process of magically restructuring our 'reality' and turning it into a 'global image scenario'. Essentially this is a question of 'amnesia'. Human beings forget they created the images in order to orientate themselves in the world. Since they are no longer able to decode them, their lives become a function of their own images: Imagination has turned into hallucination

Flusser describes the combinatorial function that generates photographs through the apparatus of the camera as a game of chance. These so-called “photographic universes,” as determined by and an extension of the camera, are “chance realization[s] of a number of possibilities contained within camera programs which correspond point for point to a specific situation in a combination game.” Cameras, by and large, for all the automated processes that occur and newer and more advanced models, still require a human to click a button to activate the photographing function. Despite this limitation, this characterization of the camera as a chance operation facilitated by the capacity to capture any subject as a photograph dovetails with data as its own form of atomized capture and combinatorial operation. This focus on chance is important because contrary to many discourses pertaining to the content of these media it allows for the possibility of an outcome that is not reducible to conventional sense-making. Larger datasets could yield more far-reaching and confounding anomalies. One of the definitive features of the internet is the phenomenon of context collapse–or the treatment of all information as existing on the same plain of inference. Juxtaposition is inevitable on the social internet and it causes otherwise disparate media to overlap and cross-pollinate.

Hito Steyerl, in her essay In Defense of the Poor Image, formulates its eponymous subject as a “visual idea in its very becoming” not in spite of but because of its imperfect, substandard quality. The speed at which visual media is shared, repurposed, reedited, and reformatted, over and over again, through this digital meat grinder contributes to this poor quality while also paradoxically enabling its durability as a container of meaning and as a new form of sense-making. Images don’t become poor in isolation but among an ecosystem of unrelated media degrading and spreading in aggregate. Unrelated media criss-crosses, overlaps, and melds together that on a micro level are nonsensical. However, this does not suggest sense, or meaning for that matter, are elusive, or impossible. The impact of this ecosystem is touched upon in a somewhat narrow way through her invocation of the Dziga Vertov concept of “visual bonds.” The poor image, “constructs anonymous global networks,” and generates a new shared history. New communities are bonded over this generatively constructed ecosystem of the “transience of the copy.”

Importantly, much of this discourse critical of GAI assumes an output that should be legible to a human interlocutor. Speaking strictly about visual AI slop, often in the form of memes it tends to have inscrutable sentences and dialogue or nonsense words. This is paired with unrealistic or impossible imagery that even if it did look real consists of situations that push the boundaries of believability. Sense should be made or the content is considered useless and therefore slop. However, this lack of sense is simply novelty that cannot be satisfyingly assimilated into common sense-making procedures. With that in mind, it is worth exploring not whether this content successfully represents the world, but what happens when content that does not resemble the world is consumed with increasing regularity. It could be argued that a precursor to slop can be found in the tumblr-cum-art project the jogging from over ten years ago. Founded by artist-researchers Joshua Citarella and Brad Troemel in 2009, its content regularly went viral on the platform precipitated by its increasingly absurd and nonsensical quality (ex: a baguette beer koozie; a piece of bacon in a hair straightener).

In the Epistemology of Noise, Cecile Malaspina discusses what she calls a ‘firewall’ in which unstructured or chaotic stimuli are determine a threshold to sense as a means of minimizing the distress caused by them:

These pre-conscious a priori act as a perceptive firewall , separating out not only superfluous stimuli, but thereby ensuring the very condition of perception: a stable sense of self.

Confrontation with ‘noise,’ broadly understood, can cause a dissolution of the self through psychological stress which ultimately leads to impairment or a diminished ability to make sense of the world. However, as she notes, this stable sense of self predicated upon the “Cartesian presupposition of a coherent self” that rests on, as Kant shows, “the universal structure of apperception, preformatted by a set of unchanging transcendental a priori.” The contingency of this nominally transcendental self is foregrounded in moments of crisis in which the boundary between self and world is complicated. Noise contra information in this sense forces individuals to confront the artificiality of their own concept of self or even their independence as somehow divorced from the rest of the world. However, this awareness does not necessarily mean a total dissolution of the self but a self-reflexive understanding of it as subject to change. Additionally, it identifies the boundaries between a self and the world as trivial. To that end, AI slop introduces noise into everyday interactions on social media platforms and distorts the boundaries of what is conventionally understood to be information or information-rich media or content. Slop is a phenomenon that can only emerge from an information oversaturated culture that generates so much content that it ultimately overwhelms and disorientates those who participate in its creation and consumption. It undermines conventional aesthetic considerations and redraws them at a society-wide level therefore no longer operating as a boundary condition for sense.

Returning to the Weilian concept of decreation, let us say art is made with a fully realized intent on behalf of the artist. However, whatever their intentions, how an audience receives it cannot be determined by the artist beyond whatever they put out into the world. In contrast to this, in Lauren Elkin’s Art Monsters she evokes the concept of decreation, discussing artists who specifically produce works that are intentionally subject to destruction and decay.

“The finished work is the death mask of its conception, but the unfinished, or de-finished, work is forever open…[t]he feminist act of decreation is making something that one acknowledges will fail, will decay. The work will not endure.…“[i]nstead of striving, competing – to fail. To accept your name (like your body) will not survive. To welcome it, rather.”

In conclusion, consider, at length, Jason Bahbak Mohaghegh in his exegesis on new Middle Eastern literature and philosophy:

“[W]hat does it mean to engage with texts as if they were fatally oriented…[t]o envision and ultimately perform a fatal experience of the text, we would have to begin to play for lethal stakes, to recognize that the text is always already condemned, and ourselves alongside it—that it has no right to remain as it is, no right to permanence. We cannot allow the literary evocation to swear an allegiance with the totalitarian mythologies of being; rather, those who would initiate the chaotic event must become carriers of an infinite risk.”

…

“[I]t must maneuver toward conditions of extremity, excess, and exhaustion…where the victories and losses of the writing-act are at their most severe—a last boundary that marks not a limit but an edge. And so this movement toward extremity casts us into the next premise of fatality—that of “excess”—for it is here that the text begins to proliferate and entangle itself in overbearing and intricate new circuitries; it is here that the text becomes an effusion (and an unforming).”

…

“Thus our discourse here is not one of transgression or of the profane but rather one of sheer convolution. The only command, the only law before us, is that of recurring distortion. The chaotic text must fashion a generative prism, one of diluted substances and imperfections; it must tempt unnatural admixtures, fusing elements into contaminated alliance. We do not compare texts in order to synthesize them in some abstract, lifeless equilibrium—to sanitize or cleanse the surface; instead, we unload them toward corrosive depths, spheres of dissonance where they run their acids across one another, following asynchronous and lacerating lines, and forging unforeseen collusions.” (emphasis mine).

…

“[W]e must betray literature; we must seek the triggers and the catalysts through which a text becomes a subterfuge—becomes the faintness of an amorphous zone—where articulations devour themselves, shatter, and regenerate in new, unacceptable maskings. To this end, the chaotic imagination must accentuate the pain of transfiguration—it must learn to play both in subtle malformations and in monstrous turnings, if only to reconvene us in a foreign atmosphere, a chamber where deception overrides truth, illusion supersedes authenticity, and where the dominion of reality has long since been overthrown. Stated otherwise, we must train ourselves to lie.”