In yesterday’s piece we uncovered the comforting mathematical truth that, left to its own devices, artificial intelligence self-destructs. Without a steady diet of fresh, original human thoughts, it suffers from model collapse, turning into digital photocopy machines, copying their own copies until their logic dissolves into repetitive nonsense and meaningless gibberish. AI, it turns out, cannot survive without us. But this raises the question of what tech companies are doing to fight this decay? They know that the open internet, once a goldmine of training material, is becoming a toxic wasteland of “AI slop.” To keep their systems from eating themselves, companies are building what appears to be an invisible, high-stakes infrastructure made of human labour, ancient data archives, and digital borders.
The first line of defence is a frantic, global scavenger hunt for data that has never been touched by an AI. In the tech industry, the year 2023 is a historic boundary line. Anything written, drawn, or recorded before 2023 is considered pure human gold. Anything created after is treated with intense suspicion, as it might be contaminated by AI-generated content. Because of this, digital archives created before the AI boom have suddenly become valuable property. Tech companies are no longer just scraping random blogs; they are quietly buying up the rights to old, locked data silos and purchasing decades-long archives of local newspapers, academic journals, internal corporate communication logs, and digitised library catalogues.
I can’t help pointing out the strange paradox of the most cutting-edge, futuristic technologies being dependent on looking backward. If an AI company wants to build a smarter model next year, its best bet is not to look at what people are writing online today, but to feed it digitised handwritten letters from the 19th century, old court transcripts from the 1980s, or forgotten forum posts from the early days of the internet. The future of machine intelligence is firmly anchored in the human past!
Even with vintage data, AI companies cannot entirely avoid the modern internet. To find new human thoughts, they must wade through the web, which means they need a way to separate the human gold from the AI dross. To do this they rely on an invisible army of hundreds of thousands of low-wage workers from Kenya and the Philippines to rural parts of the United States. In the tech industry, this is known as “data annotation” or ghost work.
These human workers sit in front of computer screens for hours a day, reading thousands of paragraphs or looking at thousands of images. Their job is to act as the AI’s taste-testers. They flag text that looks too formulaic, identify images with telltale AI errors (like hands with six fingers or melting backgrounds), and manually sort the data. When you use a chatbot and it gives a helpful, grounded response, it is not because the AI is inherently wise. It is because a human being somewhere in the world was paid a few dollars an hour to look at ten different versions of that response, grade them, and teach the machine which one sounded like a real person. This process, called Reinforcement Learning from Human Feedback (RLHF), is the ultimate guardrail. It is the human glue holding the digital mind together. Without this endless, exhausting human labour, the AI would drift into model collapse within weeks (if I’ve understood that right).
This desperate need for pure human data is fundamentally changing how the internet works for the rest of us. For thirty years, the internet operated on a model of openness (political attempts to control it excepted). You could browse websites, read forums, and share information freely. Unfortunately, that era appears to be coming to an end, to be replaced by what looks to me like the digital equivalent of land enclosures. Websites that host authentic human conversations, such as Reddit and, I suppose, FSB – you are all human, aren’t you? – have realised that their data is valuable. They know that AI companies are starving for human words. As a result, these platforms are building massive digital walls. They are blocking AI “spiders” from crawling their pages for free and demanding millions of dollars in licensing fees just to let the machines read what their users are saying. You see my ears pricking up there, at the back?
For everyday users, this means that the internet is becoming more fractured and transactional. Platforms are forcing users to log in to see content, hiding discussions behind paywalls, and aggressively protecting their data. If I’m right, the open, chaotic web is being carved up into private territories, all because human thought has become a scarce commodity (perhaps it always was, on the Left anyhow) that tech monopolies need to steal to keep their AIs alive.
Despite the walls, the filters, and the ghost workers, scientists are warning of a looming crisis: we are running ut of data. Humans simply cannot produce written words or images fast enough to satisfy the insatiable appetite of modern AI algorithms. Some studies estimate that tech companies will completely exhaust the supply of high-quality, available human text online within the next few years. When the pure human data runs out, what happens then? Tech companies are experimenting with a controversial band-aid called “synthetic data generation.” Instead of letting an AI blindly learn from all AI data, engineers are trying to carefully curate “perfect” AI data. They use a highly advanced AI to generate maths problems or logic puzzles, check the answers for absolute accuracy using rigid computer code, and then feed only those flawless answers to a newer AI.
But as I said yesterday, this works for strict, rule-based subjects like mathematics or computer programming, but fails when applied to language, art, human culture, or emotion. A computer can verify if a maths equation is correct, but it cannot verify if an essay about grief is profound, or if a political analysis is nuanced. In the realms of creativity and human experience, synthetic data remains a shortcut to model collapse. So, maybe I’ve misunderstood what the tech plan is, but so far as I can fathom, they have not yet found a way around the basic problem of requiring constant input of subjective human output.
For anyone who looks at the rise of AI with fear, understanding this hidden infrastructure changes everything – I hope. Wishful thinking perhaps, but that’s how I see it now. The mainstream narrative often portrays AI as an independent entity that is slowly eclipsing humanity. But looking behind the curtain reveals a very different reality.AI is not a self-sustaining engine; it is a parasitic technology. It requires a massive, continuous pipeline of human lived experience, human labour, and human history just to maintain its baseline sanity. The moment that pipeline thins, the machine begins to hallucinate, stutter, and break down. This gives humanity an incredible, permanent leverage. We are not the obsolete ancestors of a new digital species. We are the managers, the anchors, and the essential life-support system for a tool that cannot survive a day without our collective intellect. The fear that AI will outgrow us is a mathematical impossibility. The machine can only ever be as vast, as deep, and as varied as the human world we allow it to mirror.
