The Four Horsemen of the Knowledge Apocalypse

Conquest does not announce itself as conquest. It arrives wearing the clothes of progress: a better model, a larger benchmark, a more capable system.

By the time the field recognizes the constraint, the infrastructure, the capital, and the imagination of an entire industry have been organized around it.

Yann LeCun, Turing Award laureate and founder of AMI Labs (Advanced Machine Intelligence), and co-authors named this in March 2026: human intelligence is not general. The appearance of generality is a product of cognitive blindness. We are only aware of the tasks we can perform. The vast landscape of what we cannot do is invisible to us by design.

Building AI toward a human-cognitive ceiling is the ornithopter mistake. Before the Wright brothers, aviation engineers built machines that flapped because birds were the only visible model for flight. The logic was sound. The model was wrong. No amount of additional engineering would allow an ornithopter to escape the structural constraints of its own design premise.

The first horseman has already ridden. The field's imagination, capital, and infrastructure have been organized around a biological ceiling that no one thought to question because it was the only visible model for intelligence. That is not failure. That is conquest.

War is not a single battle. It is an accumulation of individual engagements, each locally rational, whose cumulative outcome is defeat.

Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar formalized this dynamic in NBER Working Paper 34910, demonstrating that under specific modeled conditions, welfare is non-monotone in agentic AI accuracy. As AI crosses an accuracy threshold, the incentive for humans to engage in costly learning can drop sharply. Every individual who lets the machine answer a question they could have worked through themselves makes a locally rational decision. The answer is faster, more convenient, and often more accurate than what they would have produced through their own effort.

Each individual engagement is a win for efficiency. The cumulative outcome is the erosion of the capacity to generate genuine knowledge, the kind that comes from the friction of real problem-solving, the kind that produces general principles that society accumulates and builds upon.

War does not require a single decisive battle. It requires only that the accumulation of small surrenders adds up to a loss that no individual surrender could have predicted. That is precisely what the Acemoglu equilibrium describes: a self-reinforcing threshold, once crossed, with no natural exit.

The second horseman rides through every prompt box, every AI-generated answer accepted without verification, every hard problem handed to the machine before the human attempted it.

Famine is the cruelest horseman because it does not look like what it is. The shelves are full. The volume of content keeps growing. Every search returns more results than the last. The internet has never contained more words.

Ilia Shumailov and co-authors proved what is actually happening to that abundance. They call it model collapse. When AI models train on AI-generated data, they lose the rarest, most unusual, most creative parts of the original human data first.

The internet is filling with AI-generated content. Blog posts. Articles. Reviews. Comments. Social media. AI companies scrape that internet to train the next generation of models. Which means the next generation of AI trains on the output of the current generation. Each cycle loses information. Not randomly. It loses the tails of the distribution first: the weird ideas, the unexpected perspectives, the things that made the internet feel human. What remains is the average. The safe. The expected. The bland. Then the next generation trains on that. And loses more.

The shelves look full. The nutritional content of what is on them is collapsing with each generation. That is not abundance. That is famine wearing the costume of plenty.

And unlike the other horsemen, Famine leaves damage that Shumailov et al. found to be largely irreversible under the studied recursive-training conditions. Once the tails are gone from the training data, recovery requires reintroducing human-generated content at the margins. The weird ideas, the unexpected perspectives, the things that made human knowledge worth having; once gone, they cannot be reconstructed.

Death is the most feared horseman but not the most dramatic. The Book of Revelation describes his horse as pale, the color of a corpse already cooling, not of fire or blood. Death does not arrive loudly. It arrives quietly, into a silence that has already been prepared.

Researchers quoted by Axios call it cognitive surrender: when people defer to AI outputs without fully evaluating them. 130 million U.S. adults already read below a sixth-grade level. AI is masking that, not addressing it. Workers produce outputs they do not fully understand. Teams look productive on the surface. The invisible drag on productivity does not show up in data. It shows up in what teams cannot do when the AI is unavailable, when the situation requires genuine judgment, when the stakes are too high to accept an answer that sounds right.

Not only are skill levels going down, researcher Stephen Reder told Axios. Among people at the lower end of the skill spectrum, the amount that they use the skills they have is going way down.

Skills that are not practiced atrophy. Knowledge that is not generated disappears. Judgment that is not exercised erodes. The fourth horseman does not kill with a sword. He kills by making the thing he destroys unnecessary, until the moment it is desperately needed and no longer exists.

The traditional four horsemen are not independent. Conquest creates the conditions for War. War creates the conditions for Famine. Famine creates the conditions for Death. The sequence is not coincidental. It is structural.

The same structure governs the four mechanisms of the knowledge apocalypse.

The conquering premise (building AI toward a biological ceiling) creates the infrastructure for the war against human knowledge generation. The war against human knowledge generation produces the impoverished content that feeds the famine of model collapse. The famine of model collapse accelerates the cognitive surrender that is already death's quiet work in the workforce.

And because the first horseman has already ridden (because the wrong premise has already captured the field's imagination, capital, and infrastructure), the remaining three become increasingly difficult to address by working harder within the current frame. You cannot defeat War, Famine, and Death by improving the ornithopter. The frame itself is the first horseman. The frame has to change.

Individually, each pressure could in principle be managed through incremental adaptation: better training data curation, targeted retraining programs, architectural refinements. Together, they reinforce one another. Infrastructure investment increases dependence on AI, dependence reduces human knowledge production, reduced knowledge weakens future models, and the resulting cycle increases pressure for still more infrastructure. The significance of the four horsemen is not any single trend but the feedback they create when considered together.

In some interpretations of Revelation, the four horsemen are preceded or followed by a fifth figure: the one who opens the seals and makes the hidden visible.

The word apocalypse does not mean catastrophe in the original Greek. It means unveiling. A revelation. The moment when what was hidden becomes visible.

The four mechanisms of the knowledge apocalypse have been hidden in plain sight: published in peer-reviewed journals, flagged by Nobel laureates, documented by journalists, named by Turing Award winners. What has been missing is not the evidence. It has been the frame that makes the convergence visible.

That is what this series of papers has attempted: not to be the horsemen, but to open the seal. To name what is riding. To make visible what has been happening in plain sight.

The machines are faster. The outputs are more abundant. The benchmarks keep improving. And the knowledge base that all of it depends on is being quietly consumed by the horsemen that no benchmark measures, no capability claim addresses, and no ornithopter can outrun.

The fixed wing does not flap. It does not inherit the constraint it was built to transcend. And the intent economy does not optimize for output at the expense of the human process that generates it.

That is not a technology argument. It is a civilizational one.