On large language models, archives, generation, slop and art

Holly Herndon, an artist who works with language models and generation, said in a recent interview she gave to Der Spiegel, that “people think AI is some kind of an alien, when it is actually something from us, our culture.”

There is a lot to unpack here, the depiction of AI as something from “our culture” leads to questions: what exactly is “our culture” supposed to mean? Who is meant by “we”? One wonders if “we” did not learn anything from the process of examining the archiving and knowledge organisation and categorising practices of the Global North that are built on colonisation, slavery, and the exploitation of people and land.

Mystification of large language models (LLMs) is a phenomenon that has emerged over the past decade, before language models were forcibly made part of the daily life of many. The LLM was mystified as something that promised a grand future, an exciting phenomenon that only a few great minds thoroughly understood. Our AI future was something that should be celebrated without qualms. In the current decade, LLMs have already been integrated into internet search engines, and embedded into almost every app, and with that ubiquity, they have become completely demystified. Mystification can only happen as long as the concept stays abstract and insulated from daily reality where there is no room for idealisation and mystification.

When the data centres that power LLMs are not leaving enough electricity for neighbouring residents, and they are soaking and stealing water from the surrounding areas, the magic has to become further demystified: LLMs become so fleshy that they become the opposite of mystical.

Almost too fleshy in material, as in the film Ferat Vampire, where cars are sucking people’s literal blood through their gas pedals.

Upír z Feratu, 1982

The AI bubble has been imposed on people, while in the background companies invest immensely in large language models—together with the infrastructure they require. Since the capitalist class has invested so astronomically over recent decades in AI, they now have to push the technology equally hard to establish it, to make it profitable, and make billions of people use this infrastructure. From a sense of capitalist class solidarity, they impose LLMs to protect their mutual investment, with great determination.

The aim is to make the public, major companies, and, especially, governments use LLMs until these models become something they can’t give up, similar to other supposedly convenient big tech tools people continue using out of habit while ignoring their problematic aspects, even when using them becomes a burden.

As the capitalist class wants to stick AI into anything and everything possible in our working lives, the message is clear: the contemporary art market had better embrace AI, too, or face being left behind, and so it does. Soft power is crucial, the AI industrial complex needs it too. Art offers a way to launder, even to valourise, the possibilities of LLMs for audiences.

(Break)

When we consider the fact that LLMs are, in fact, massive archives, then we can build on the knowledge we possess regarding archival colonialities when thinking about these models.

In their heyday, the archiving practices of Western colonial powers claimed to serve a greater cause—in their language “civilisation” and “humanity”—all while they categorised extracted artieacts with the aim of expansion and entrenchment of their empires. Categorising the Other in a specific way nurtured their white supremacy.

No person in their right mind should accept the idea of making something liberatory out of the very same colonial archives. For someone who wants to engage in decolonial archival practice, the first thing you should do is to refuse to work with those archives and instead start archiving otherwise; practice counter-archiving, and build your own archive. When we apply this logic to LLMs, we can see that the current liberal discourses on copyright, “discriminatory AI”, or even “AI ethics” etc. barely scratches the surface of the core problems of the technologies. Centralised archives and museums represent “our” culture, as much as big datasets represent “our culture”.

“AI is not a tool, but a socioeconomic instrument of power.”

The spectacle becoming de rigueur in contemporary art in the 2020s is a consequence of the embrace of virtual and augmented reality technologies and, more recently, AI. The augmented reality bubble did not last long, despite being hyped to a prominent place in the art market. It started losing air within a few years after failing to establish itself as a viable genre with monetary value within the semiotic economies of contemporary art. VR and AR, which have their origins in the defence sector, offered allegedly enhanced sensory experiences, not unlike purported augmented empathy, augmented feelings that (apparently) become available to humans uniquely through these technologies. After these technological and discursive interventions fizzled, AI has been pushed aggressively within contemporary art in various ways. While performance art has always been a spectacle, for example, the current state of AI transforms it increasingly into auto-generated slop that aims to evoke feelings of astonishment in a similar fashion to the way non-art generated images do. While the generation of texts and images is being forced on us in every aspect of our quotidian lives, the aesthetics of generation are taking on an inevitable currency that art strives for.

The slopification of art is not about art about slop but art that has internalised the aesthetics of slop. Performing the slop is a performance that awakens similar affects as general slop does for its consumers, in particular producing shock and lurid fascination. Similar to how slop-producing influencers work, slop artists give “people what they want: the creepiest videos, the most over-the-top.” One of the most recent examples of performing the slop was at the Venice Biennial 2026, at the Austrian Pavilion. In contrast to the shock effect created by the performance art in its origins, in the slop timeline, no other feeling is intended to evoke, such as disgust, discomfort, threat or provocation. Abjection may be there, but existing in a more passive way, as in an endless loop. There is much more to come from this genre, it seems.

Many have already written that AI slop recapitulates fascist aesthetics; some have compared it with Italian futurism, which celebrates acceleration and war machinery. Slop’s connection to trans-humanism, however, is largely absent in this discussion. Similarly absent is its connection to cyberfeminism, although this is not a clear path as the one to futurism. Looking from today’s perspective, cyberfeminism cannot be seen as a separate category of its own but must take its place within historical relations, where technologies were lionised and positioned as saviours, then “cutified”, and then no longer treated as the mere tools (or products, services, or infrastructures) as they should have been. Without asking infrastructural questions, any attempt at reckoning with critical understanding of technologies can only feed into the dominant optimistic narrative of technology as Big Tech bros define it. Girl-bossing capitalism only serves to add sparkle to the system itself.

(Break)

A data scientist is not a scientist, but a person who analyses and categorises collected data.

The archiving practices of past centuries were pursued under the banner of science, since science enjoys an exalted status in Western society. Although science, like other tools, crucially advanced the colonial arsenal through land grabs and the immense exploitation of people at scale. The dominant narrative still positions science as serving a greater common cause whenever the role and history of science and technology are considered.

Manufacturing consent through science-ability is a method that can be utilised whenever it is needed. The narrative must be that there is a larger cause at stake, bigger than the prerogatives of a handful of companies wanting to multiply their wealth and power. In the case of AI, the familiar refrain goes: the use of technologies means progress, and progress will bring good for all. Science has always been an integral part of the capital accumulation system and does not exist in a completely independent realm.

After World War 2, many senior Nazi party members continued to teach in universities through the 1950s. Their ideologies were not an aberration but a fundamental building block in the basis of post-war Western science. Computing history in the West is directly entangled with military spending, extractivisim, and colonialism, machine learning and automation technologies are no exception.

Today, tech companies work very closely with universities; they need universities’ facilities and personnel to develop their products in the first place. University campuses and research centres under capitalism are playgrounds for the corporations and start-ups that make up the current tech landscape.

As art enjoys its own form of science-ability by being included in these funding structures, you can receive public or private funding (almost exclusively) if your work uses the technologies of extractivism in an affirming way, illustrating how arts and culture function as forms of “soft power”.

Most of the funding programmes related to AI and art start from a heavily weighted question: how AI can be useful, and follow now-familiar right wing narratives of victimisation, inverting the reality and depicting the dominant as the oppressed. According to such funding calls, negative critical discussions on AI are abundant, but the ones embracing AI are very limited.

The following provides an example of the genre:

“While critical discussions [of] technologies like AI, blockchain, and immersive media are well established, attention to the operational and infrastructural conditions that enable AxAT practices has been limited.”

Slopification is not just an aesthetic proposition. The slop logic is now becoming procedural to AI-assisted production. The first wave of the AI-generated images, a prompt saying “a person doing X” generates an image. When you look at the image, what you see is an image performing a prompt.

In contrast to other traditions of antromorphic image-making, for example puppetry or animation, there is no extra layer to the process other than performing the prompt in a convincing way. In puppetry, the intervening aspect of embodiment enables the de- and reconstructing of the human body. In the words of Matthew Isaac Cohen, “Puppets are alien others and closely associated with the person. They are ‘not me’ and also ‘not not me’.”

With animation, however, there is no connection between the human and the anthropomorphic anymore. Still, there is an extra layer of textures and 3D calculation instead which calculates movements based on the mechanics of the human skeleton.

In a generated video, there is no calculation, simply a mimicking of human movement that is not based on a structure or skeleton. There is simply an antromorphised automation, one could say. Similarly, brainrot can be regarded as the abject of rationality, or the obsession with a perfect prefrontal cortex.

In that sense, slop is the true aesthetics of what remains of humans under capitalism.