118e9c1d-013a-4574-9777-0b62a9b88296_1456x816

One of my friend’s mother posted a video on WhatsApp status last month with a question I have been hearing more and more: she had seen a video of a tall building collapsing into a street, with a roaring dust cloud and a crowd screaming in the foreground, and she wanted to know whether it was real.

I recognised the fake within seconds, because the debris fell in smooth, unnaturally even slabs and the shop signs in the background were made of letters that belonged to no language, but I also understood why she was shaken.

She belongs to a generation raised on the assumption that a moving picture means something happened in front of a camera, and nothing in her experience prepared her for a world where a stranger can type one sentence into a machine and collect millions of views for a disaster that never took place.

That is what people now call AI slop: low-quality content produced in huge quantities with generative tools and published without fact-checking or any sense of responsibility for what it does to the people watching.

The defining feature is the priority given to volume over truth, which is why a carefully edited, honestly labelled AI-assisted piece does not belong in the category while a lazy human-written spam article does.

The reason such content keeps appearing is plain arithmetic. Content farms were once limited by labour, since every video needed a person and every person needed a wage, but generative tools lifted that ceiling at the same moment platforms were paying for views rather than for accuracy, so whoever floods the feed fastest earns the most.

The producer keeps the advertising revenue while the cost is spread across everyone else, including the aunt who now cannot tell a real catastrophe from a rendered one, the teacher who finds fabricated facts in a student’s favourite educational video, and the honest creator whose real work disappears under thousands of near-identical imitations.

Economists call this a negative externality, though a more honest description is a tax on public trust, collected by people who never asked permission and who have no plans to pay anything back.

The real-world price of artificial intelligence

If slop were merely irritating, an essay would be wasted on it, but the machines behind it are physical and hungry, which is how a fake building collapse ends up on the price tag of a laptop in Dhaka.

Modern AI accelerators depend on a stacked, specialised memory called HBM, and the three firms that control roughly 92 percent of the DRAM market, Samsung, Micron and SK Hynix, are cutting production of conventional DRAM substantially to prioritise HBM for AI infrastructure.

The consequences for ordinary buyers have been severe: TrendForce reported that DRAM contract prices rose 90 to 95 percent quarter over quarter in the first quarter of 2026, the steepest single-quarter increase in memory pricing history, and PC makers responded by raising retail prices 15 to 25 percent while unit sales declined.

Micron even retired its consumer-facing Crucial brand in February to focus on far more profitable data-center and HBM products, an admission that the everyday customer had become the least interesting person in the room.

A student who simply wants sixteen gigabytes of RAM for a thesis is therefore subsidising a data-centre boom she never joined, and import duties, exchange rates and distributor margins make the bill larger by the time it reaches a counter in Elephant Road.

Fairness requires one clarification, because the argument is stronger when it is precise. Most of this demand comes from training frontier models, enterprise software and scientific computing rather than from spam, and a chip cannot tell whether it is running cancer research or a fabricated disaster.

That indifference is exactly what makes the model so troubling, since the seller of compute earns the same revenue whether the output heals someone or deceives them, and no part of the business ever has to ask which one it is.

When the AI industry starts financing itself

The money behind the boom deepens the concern.

In August, Nvidia agreed to guarantee up to $105 billion backing an OpenAI lease at the PORTS-Pike campus in Ohio, a ceiling it set after cutting its earlier discussed figure of roughly $250 billion by more than half once its shares fell and investors objected to the exposure.

The structure is worth reading slowly: Nvidia guarantees the value of the buildings, invests $1.5 billion in the developer, supplies all of the AI compute at the site, and has discussed financing OpenAI’s purchase of the chips.

Its chief executive has acknowledged the upside, estimating that each generation of systems deployed in Ohio represents about 1.5 million GPUs worth up to $200 billion.

A supplier that pays for the building, backs the lease and then sells everything inside it is effectively writing the demand for its own product, which is why investors from Michael Burry to market commentators have compared the arrangement to the self-reinforcing loops that preceded earlier crashes.

Nvidia replies that OpenAI pays the lease and that the capacity could be re-let to other tenants if OpenAI left, and that defence has some force, yet it does not remove the central problem: when suppliers, investors and customers are this tightly knotted together, headline demand stops measuring what real people want and can afford.

This is where the pieces lock together, and where the polite version of the story loses its credibility.

Platforms profit from engagement whether it is earned or manufactured, the AI labs profit from subscriptions and compute whatever that compute produces, memory makers earn more from scarcity than they ever did from abundance, and the chip seller profits from financing the very buildings that need its chips.

Each link is rewarded for looking away from the damage, and the damage lands on those standing at the end of the chain: the older relatives who cannot verify what they watch, the buyers paying a memory surcharge, and, should the loop ever snap, the investors and public institutions who were told the demand was genuine.

I do not need a villain in a boardroom to call that evil, because a system that pays every participant to ignore its harm has already made a moral choice on their behalf, and the absence of a single mastermind is little comfort to the people who receive the bill.

We’re becoming the raw material

The loop does not stop at the checkout counter, because the machine must be fed, and more and more it is fed by us.

Researchers have shown, including in a 2024 Nature paper, that models trained repeatedly on synthetic output can degrade in a process known as model collapse, so the flood of slop pollutes the material that future models learn from, and the companies need fresh, authentic human data more than ever.

That should make anyone pause before the next trend asks them to upload a face and become a cartoon, a movie poster or a person from another decade.

Consider what changes hands. You give away a clear photograph, often with family, friends or the inside of your home visible behind you, in return for ten seconds of amusement you will forget by next week, while the company receives the image, its metadata and a measurable sign that the trend is working.

Policies vary by provider and product, but OpenAI’s published policy as of its 2026 updates says content from individual consumer services may be used to improve models unless the user opts out, whereas business and API products are off by default, so your protection depends on a setting most people never open.

The people in your photographs never agreed to anything, your child’s face never agreed to anything, and once an image leaves your device you cannot easily recall it or control how long it is stored, who can reach it, or what happens during a breach or a legal demand.

The company gets a durable asset and you get a toy.

What we can do about it

This is not an argument for fearing every tool, since blanket rejection would punish students, accessibility users and honest artists who use these systems responsibly.

It is an argument for the pause that the business model is designed to prevent, which means asking whether a service uses your upload only to answer your request, how long it keeps it, whether you can opt out of training, and whether you would be comfortable if it leaked; anyone who cannot answer all four should close the app.

At the system level the remedies are well within reach: platforms can reward quality over raw volume, provenance tools can show what is real, deceptive content can carry consequences regardless of whether a human or a machine made it, and the financing behind the boom can be examined through revenue against spending, real usage against announced capacity, and an honest account of who carries the risk.

Until those changes arrive, the most effective resistance is also the least glamorous, which is to verify before believing, to teach the older people around us how to check what they watch, and to keep our faces and families out of someone else’s training data.

The building never fell, and the trend asking for your photograph is not a favour, since both are products and you are the raw material.