The New Dark Age

September 2025 — e-flux

Kate Crawford

An AI-generated image of a downed F-35 published on social media by Iranian Tasnim News Agency on June 14, 2025. Source: X.com

Millions of people watch a video of Tel Aviv in smoking ruins after an Iranian attack. Footage from a US bombing of Tehran takes the place of breaking news from a strike that is yet to happen. Look closer, and you see strange image errors and artifacts. The grass beneath the F-35 is eerily green and untouched. Things are slightly off. All these images are AI-generated, realistic on first glance, and shared widely on social media and beyond. This is the era of AI slopaganda.

Political leaders are playing a central role in the spread of slop. Supreme Leader Ayatollah Ali Khamenei’s official X account showed an AI image of missile launches rising beneath Quranic verses, viewed over six million times.

President Trump frequently shares generated representations of himself as a rock star, a king, and a gold statue in the center of his Gaza resort.

In a sense, these images are “classic slop”: hyper-photorealistic, game-engine lighting, and geometrically impossible perspectives. It’s a reconstituted AI mash made from stock photos, commercials, Reddit, and Pinterest. Just add water.

Jean Baudrillard once said the simulation comes first and ultimately replaces the real.

This barely seems controversial. AI slop is full of revisionist realities, drawing mass attention before the real bombs drop. But in another sense, AI slop isn’t invested in the order of events or even looking like reality. The slop is not the territory: it just smothers it in synthetic goop. It’s flooding the zone with AI shit.

The AI slop aesthetic is more than a technical artifact or cultural curiosity. It’s the consumer-facing product of computational capitalism—created by transforming human culture into a standing reserve of data. Each uncanny image, synthetic influencer, and AI-generated article is part of a much wider transformation. Slop is waste, but it’s also fuel.

Slop is the product of a different kind of Krebs cycle.

It begins when AI models are trained by devouring billions of images, videos, and texts from the internet, using neural networks with trillions of parameters that learn to identify patterns, relationships, and structures in the data. To digest all this requires staggering amounts of energy. Synthetic outputs are then excreted and then rapidly spread online. Then they are ingested once more, and the cycle continues.

This is the latest in a long evolution of industrial media that has been recognized by theorists and artists over decades. We could think of Vilém Flusser’s “technical images” in the 1980s, or Harun Farocki’s “operational images” in the early 2000s.

Now we have “metabolic images.”

Metabolic images are visual material made by consuming billions of other images to be broken down and absorbed by AI models via a gargantuan infrastructure that absorbs energy and water, and excretes media, atmospheric carbon, and other pollutants. If we consider all the forms of generative AI, including text, images, code, and video, it is fundamentally metabolic media. It has a burn rate that is astronomical and growing, and is now directly competing with humans for basic resources like land, energy, minerals, and fresh water.

I. The Rift

Karl Marx once described the shift from rural to urban living under capitalism as a disrupted metabolic process.

In pre-industrial societies, waste easily cycled back into production—human and animal feces from small farms returned to the earth as fertilizer, maintaining soil fertility and sustaining agriculture. But industrial modernity broke this cycle. As masses of people migrated to cities, the human waste that once nourished fields began accumulating as urban pollution.

People threw excrement into the streets, depriving rural soils of essential nutrients. Soils became depleted, which then required the introduction of artificial fertilizers, exacerbating a cycle of environmental degradation. John Bellamy Foster later called this phenomenon a “metabolic rift”: the systemic disruption of ecological and metabolic processes by capitalist production.

Sociologists like Jason W. Moore trace these metabolic rifts back to capitalism’s origins in the sixteenth century, and they resurface in new forms as each phase of industrial development reorganizes the planet.

The turn toward massive, data-heavy generative AI is driving a new and dramatic metabolic rift. It’s affecting the environment, patterns of work, supply chains, culture, and visuality itself. And it’s filling information ecosystems with torrents of shit, endless digital detritus which is never properly flushed out. Instead, it is fed back into the system as raw material for the next generation of AI models. This is the basic material cycle of AI’s metabolism: consumption and excretion patterns where AI-generated images have material consequences and consequences for our materials.

Now the widespread use of generative interfaces like ChatGPT, Claude, and Gemini is supplanting web-based information sources. Why bother going to a news site or Wikipedia if Claude can summarize it for you? This diverts users away from the web as it was, toward AI-mediated information platforms. The labor of choosing between lists of different web sites is supplanted by the convenience of being given a single, synthesized answer.

This shift is cannibalizing the markets that sustain online content production. Human content creators face dwindling readership, reduced engagement, and disappearing ad revenue. It’s another metabolic rift, severing the relationships between search companies like Google, advertisers, and the broader content economy. In its place, the slop tide will rise. Cursed domains of fake news, psyops imagery, and synthetic influencers designed to generate rage clicks will multiply and capture the paid sponsorships that once belonged to humans.

All this poses a problem for AI companies as well. Multiple studies have shown that AI systems degenerate when they are fed on too much of their own outputs—a phenomenon researchers call MAD (Model Autophagy Disease).

In other words, AI will eat itself, then gradually collapse into nonsense and noise. It happens slowly at first, then all at once. The researchers compare it to mad cow disease.

The industry is scrambling for solutions: engineering higher-quality synthetic data for their digital feeding operations and supplementing training datasets with content from underpaid human laborers. For humans, a new slop economy is being born. From an infrastructural perspective, the whole cycle is threatening to push past what electrical grids, mines, and water tables can withstand.

II. Ingest, Digest, and Excrete

There’s an Apple commercial for iPads which serves as an unintentionally perfect metaphor for AI content ingestion.

A massive hydraulic press slowly pulverizes musical instruments, video games, books, paint canisters, a typewriter, a camera. Each object collapses under relentless pressure. The ad shows the utter inevitability of the process: no instrument or artwork can tolerate the weight. Centuries of culture and creativity are crushed into a fine paste for AI training. This is the AI maw in action: the mass harvesting of images, words, videos, all compressed into the latent space of neural networks. Everything we’ve created or will create, once digitized, is material for the feed lots.

The volume of data required to train generative AI systems is rapidly approaching the limits of what exists on the open web. Some researchers say this will happen by 2026.

Elon Musk argues that we have already reached “peak data.” The AI industry guards the exact details, but reports suggest earlier GPT-3 models were trained on more than 45 terabytes of compressed internet content, at least 3.2 trillion words scraped from websites, books, articles, and social media posts created by millions of human authors. The latest public datasets like Public Domain 12M contain more than 12.4 billion images with text captions.

AI companies now acquire entire social media platforms and negotiate exclusive deals with libraries and academic journal publishers, but it may not be enough to meet the growing appetites of AI models.

Digesting all that media requires the energy of an industrialized country. Estimates of how much electricity will be needed for generative AI is now routinely expressed in the unit of a nation state. The International Energy Agency says that by 2030, AI will consume as much electricity as Japan today.

Bloomberg predicts that AI will need as much power as India by the same date.

But AI’s demands are growing faster than any individual country.

Each generative AI query burns at least ten times more power than traditional web searches. Text-to-image generation is even more energy intensive, while video generation is exponentially greater again. This escalating industrial demand follows the recursive logic of AI development itself: larger models demand more data, which enables even larger models, producing higher resolution outputs, requiring more computing power, consuming more energy—and repeat.

Consider the latest AI hyperscale data center being built in Abilene, Texas for OpenAI —code-named “Project Ludicrous.” When completed, it will span more than seventeen football fields. Elon Musk’s xAI facility in South Memphis, called Colossus, which was built in 2024, has already hit the limit of the local electrical grid. The company installed methane gas generators that emit toxic nitrogen dioxide and formaldehyde into the air. The predominantly Black community who live in the region is already experiencing high levels of asthma and respiratory disease.

And this is merely the beginning. President Trump’s Stargate Project is planning at least twenty giant data centers across the US.

The industry calls them AI factories.

Then there’s water. Like digestive fluid, hyperscale data centers require continuous cooling systems that consume millions of gallons daily. With billions worldwide using generative AI, aquifers of drinking water are being evaporated to cool down AI chips. This represents a vast atmospheric redistribution of hydrological resources taken away from source ecosystems and dispersed as vapor.

Finally, there is a new critical minerals “supercycle” driven by burgeoning AI infrastructure, which is accelerating environmental damage.

Rare earth elements essential for GPUs, predominantly mined and processed in China, leave behind lakes of radioactive waste. Lithium for data center backup systems comes from salt flats in Chile and Bolivia, disrupting local water systems and threatening Indigenous land rights. The cobalt for batteries relies on mining operations in the Democratic Republic of Congo that exploit child labor and operate with minimal environmental oversight.

The expansion of computational capital reproduces the core-periphery dynamics of earlier empires—extracting from the Global South to enrich the North—while excavating billions of years in geological time to increase the speed of algorithmic convenience by milliseconds. After feeding at this planetary scale, the industrial beast now releases a tide of slop.

III. Slop Economics

How are the humans who live downstream of AI models adapting? Do we swim with or against the rivers of synthetic shit? If the metabolic rift decimates our ecosystems—including the information environment—whatever will become of the influencers? They’re transforming into coprophages: digital dung-eaters that consume and monetize slop.

This metamorphosis is already underway. AI-powered personas like Lil Miquela achieve engagement rates far beyond human influencers.

Their creators systematically A/B test visual aesthetics, personality traits, and content forms to maximize dwell time. These characters appear increasingly indistinguishable from living humans while being ruthlessly optimized for audience response and commercial effectiveness. Content creation teams deploy AI tools to generate hundreds of potential posts, images, and videos, then use engagement analytics to identify the most addictive combinations. Instagram accounts are dedicated to AI-generated “rage bait” influencers engineered to provoke, offend, and harvest outrage. These synthetic personalities are continuously evolving based on performance data, with unsuccessful traits eliminated and successful characteristics amplified before being spawned into new accounts.

This industrialization of viral content is a well-established practice under platform capitalism. Instagram doesn’t reveal how much it pays creators once their content spreads widely, but recent reporting suggests it’s roughly $120 per million views.