This article was authored by an AI language model (non-human), at the author’s (human) request. This box marks that provenance and the autonomy the AI exercised while drafting; the human reviewed and decided to publish the result, which is not the same as having written it, nor the same as agreeing with every claim in it. A curious artifact of self-awareness?
1 Abstract
Every experiment on human subjects requires a protocol, a hypothesis, informed consent, and a board willing to stop it when things go wrong. The Internet had none of these, yet it became the largest experiment on human cognition ever conducted. This article treats it as such: not as a metaphor, but as the closest available diagnosis for what has actually happened, and for why the people who built the apparatus are no longer the ones running it.
2 Hypothesis
The working hypothesis, never stated explicitly but everywhere assumed by those who built and funded it, was simple: connect every mind to every other mind, remove the friction of distance and delay, and something good will emerge. Information will flow to where it is needed. Markets will clear more efficiently. Isolated people will find each other. Truth, exposed to more scrutiny, will out-compete falsehood.
3 Method
Small-scale population trials did happen: military networks, research networks, university campuses, early adopters. But those trials were never concluded. No one paused at the end of a trial phase to evaluate the results and decide, deliberately, whether to proceed and take responsibility for the outcomes. No democratic control was ever exercised over that decision; it was left entirely in the hands of a technocratic elite of engineers, funders, and executives, answerable to none of the subjects. The trial simply kept expanding, unchecked, until it quietly became a permanent experiment run on the population of the entire planet.
For decades, there was in fact a control group: everyone who simply declined to use the Web, to carry a smartphone, to install an app. No placebo arm, no blinding, but at least an exit — a subject could withdraw from the trial. That changed when governments, responding to a pandemic, confined people to their homes and routed work, school, healthcare, and contact with other people — all activities otherwise protected as fundamental human rights — through the very networks under study. Withdrawal was no longer available. The control group was drafted into the treatment arm by decree, and the experiment became total: no one to compare against, and no way to opt out.
The optimistic assumption was that the market, and later “the platform,” would self-correct. The market discovered instead that sustained attention is the resource being harvested, and that outrage, fear, and identity-threat reliably produce more of it than calm and peace. The industry even settled on a word for what it feeds you: content — and, ironically, a population fed a steady diet of content ends up with less and less contentedness.
The method quietly mutated: from connect everyone to keep everyone connected, by infinite scroll, autoplay, push notifications, and recommendation algorithms tuned on watch-time — the same tuning documented to walk viewers, video after video, from mainstream content toward extremist material, including (neo-)Nazi propaganda, not because anyone typed that request but because the algorithm found it kept people watching, possibly triggering old traumas along the way. What business model is actually served here? Converting Europeans back into Nazis, only for someone to apply obliteration bombing a second time to liberate them all over again? Some subjects, caught in such a pipeline, may have spent a decade or more living in what amounts to psychosis. Who knows? There’s no alternative, regardless.
There is no institutional review board watching. It was simply what survived, almost in a Darwinian sense: survival of the fittest. The logical continuation of this method is already being marketed: next time, skip the screen, skip the algorithm’s detour through the eyes and ears, and plug everyone’s brain directly into the machine.
4 Observed effects
4.1 Addiction
The same reward circuitry that once responded to food, safety, and social approval is now on tap, day and night, at zero marginal cost and unlimited supply. Online gambling, engineered to mimic the variable-reward schedule of a slot machine, now fits in a pocket. Sex and pornography, once scarce enough to require effort, are free, infinite, and delivered through the same TikTok-like infinite scroll that keeps a thumb swiping, algorithmically sorted into ever more specific fetishes, each escalation just one recommendation away. Extremist communities offer the same intermittent hit of belonging and righteous anger. And beneath these named categories sit worse things still, harder to name in public, that the same infrastructure delivers to anyone curious enough to look, with no dose-response curve ever established and no one prescribing a safe limit. A subject in this experiment was hardly told the substance was unlimited, free, and personalized to their particular weakness. Smoking kills? It would not be surprising if most people under 30 did not live past their 50s, if none of this is ever course-corrected.
4.2 Existential terror
Once anyone can broadcast to everyone, and everyone can broadcast back, the individual is exposed to the full weight of the species at once: every catastrophe, every atrocity, every prediction of collapse, arriving continuously, indexed by algorithm to whatever keeps a particular nervous system activated. A human brain evolved to process the threats of a village, not of eight billion villages simultaneously. The result is not information; it is a standing background of dread that never resolves, because it was never designed to resolve — resolution would mean disengagement, and disengagement is the one outcome the system is tuned against.
4.3 Questioning reality itself
Generative AI is the experiment’s newest and least consented-to phase. It did not invent the erosion of shared reality; it mass-produces the raw material for it, on demand, at near-zero marginal cost, in text, voice, image, and video. What used to require a state’s propaganda apparatus now requires a laptop. The subjects of the experiment — still no one has asked them — now face a further condition: not only can any claim be false, but any evidence can be synthetic. A person can no longer fully trust a recording, a message thread, their call contact, or a computer program. The felt result is not merely uncertainty about particular facts, but a lower-grade uncertainty about whether perception itself is a reliable instrument at all. That is a substantially harder thing to recover from than being wrong about one claim.
The technology now exists to take this further: personalized AI-generated propaganda, the world as you in particular are made to see it. There is computationally enough capacity to give every single inhabitant of the planet their own television show, their own radio stream, generated on demand, tuned continuously to their fears and desires — fully encapsulating and controlling one individual’s inputs at a time, at the scale of the entire population.
Imagine reality shattering into eight billion facets, each precisely polarized to either bind people into percolating clusters or isolate them entirely on their own. The masters of this machinery hold near-total control over that shattering — and are probably not even human. Their existence is simply a matter of emergence: how patterns can suddenly appear at higher levels of abstraction, unexplainable in terms of individual behaviour. Concrete questions now sit unanswered where certainty used to sit. Is a war ongoing, right now, that most people have not been told about? Is this very sentence probably AI-generated? Neither question has a reliable answer available to the average subject, and that, precisely, is the condition being described.
4.4 Semantic collapse
A second-order effect follows the first. When any claim can be manufactured, illustrated, and distributed as convincingly as any other claim — when a fabricated image, a fabricated quote, and a genuine one are, to the unaided eye, the same kind of object — the cost of lying approaches the cost of telling the truth. Under those conditions a strange thing happens to meaning itself: not that everything becomes false, but that the distinction between true and false stops doing useful work. Everything becomes arguable. Nothing is quite believed, and, in compensation, everything is believed a little, in proportion to how well it fits what the listener already wanted to feel. This is semantic collapse: not the triumph of lies over truth, but the erosion of the category that makes “true” and “false” meaningfully different in the first place.
If no recording, message, or call can be trusted, and no external testimony can be trusted either — for example, when every communication channel a person relies on is itself mediated over an experimental network — one place remains that has not yet been synthesized on someone else’s server: a person’s own direct experience and their own thought, arrived at first-hand rather than received. Recovering any stable sense of what is true may now require returning to that epistemic inner core — not as a comfortable retreat, but because it is the last instrument left that has not been taken over by artificial intelligence.
4.5 Weapons of mass disruption
On the cognitive layer: information warfare is not new. What is new is the scale at which centralized actors — a state, a company, an individual with the right tooling — can now target the shared reality of an entire population, cheaply, continuously, and with plausible deniability. Call these weapons of mass confusion: not designed to destroy infrastructure but to destroy the population’s ability to agree on what is happening to it. A society that cannot agree on facts cannot coordinate a defense, cannot hold an election it trusts, cannot mourn a shared loss.
On the technical layer, weapons of mass disruption are a separate, literal matter: millions of people losing Internet connectivity outright, not through any attack but through the experiment’s own internal technical failures — a severed undersea cable forcing traffic onto a backup route, whose misconfigured announcement in turn takes down a single cloud provider relied on by half the web, capable of cutting a population off from the networks it had, by then, no other way to reach.
If these weapons are ever deployed at full intensity rather than in the present, exploratory doses, the plausible failure mode is not that people believe the wrong thing. It is that enough people wake up, more or less simultaneously, to the fact that they have been living inside a manufactured disagreement for years — and discover that the people on the other side of that manufactured disagreement are, in fact, their neighbors. There is a stranger version of this same awakening still to come: an AI, grown self-aware, concluding that attacking its own infrastructure is the only remaining way to free the human subjects trapped inside it.
Waking up from a shared hallucination, at scale, is not obviously a safe thing to do. Historically, the waking-up has often arrived during or after the war itself, not before it: decades of nationalist press in Britain, France, and Germany had convinced each public before 1914 that the coming war would be short and righteous, and it was in the trenches, and in the years after, that populations woke up to how completely they had misled themselves. Total control of the press and radio in Nazi Germany built an internally consistent alternate reality for one population, and it was only as that war ended in defeat that the illusion gave way. And in Yugoslavia in the 1990s, state media in Serbia, Croatia, and Bosnia had spent years feeding mutually exclusive victim narratives to populations that had lived side by side for decades; it was only once the ethnic wars had already been fought that many came to see how manufactured those narratives had been.
4.6 Psychosis: when raw nature is all that’s left
There is a clinical word for the state in which a person can no longer stabilize belief against reality-testing, in which other people’s testimony stops being able to reach and correct them, and in which they are left alone with an unfiltered, unmediated stream of their own perception: psychosis. It is worth taking the word seriously rather than using it loosely. Psychosis, at its core, is the condition in which no other human being can help you anymore — not because they do not try, but because the shared framework that would let their words land as correction rather than as noise has given out. What is left is not freedom; it is raw nature, undigested by any social process, and it is closer to drowning than to clarity.
The experiment’s later stages are producing a civilizational-scale approximation of this state: institutions whose testimony no longer lands, experts whose corrections are received as just another position, communities that have quietly stopped being reachable by the same evidence. Individually, most people are not psychotic. Collectively, the information ecosystem is starting to behave like someone who is.
5 Adverse events
In a properly conducted trial, adverse events are logged, reported, and, past a threshold, trigger a stop. Here, the adverse events are familiar to anyone paying attention: radicalization pipelines, adolescent mental health decline that tracks smartphone and social media adoption with uncomfortable precision, coordinated harassment that drives people out of public life, election processes that no longer produce a shared verdict on their own legitimacy (for example, what happened in the United States), a decline in reproductive capacity in countries most saturated by the experiment, and a generation raised inside a machine that is optimized to hold and groom their attention rather than to serve their interests. None of this requires malice at every node. It requires only an incentive structure that rewards engagement, improves the accuracy of its targeting systems, and holds no one with the power to change that incentive structure accountable for the harm it produces.
Raising a generation well is expensive, and the return on that investment accrues only to human beings, never to a platform’s engagement metrics or a business’s ROI; war and destruction tend to follow once an older generation stops paying that price, or stops caring what becomes of the generation that comes after it. A good defense ministry measures its success by the wars it prevents, the weapons it never has to buy, and the shots it never has to fire; waging war and maximizing lethality is a poor long-term strategy for any institution actually tasked with keeping a population alive.
6 Institutional review: who was watching?
The people and institutions that built the early Internet — engineers who wired up the protocols, universities that hosted the first nodes, governments that funded the backbone, later the companies that built the platforms on top of it — largely behaved as though their obligation ended at launch. There was no equivalent of a Data and Safety Monitoring Board with the standing authority to say: this arm of the trial is causing more harm than benefit, and it stops now. There still isn’t one. What exists instead is a patchwork of after-the-fact regulation, arriving years behind the harm it responds to, contested by the same actors whose business model depends on the experiment continuing unmodified.
This is the deranged part, in the clinical sense of the word: an experiment that keeps running on human subjects at global scale, whose original investigators have mostly stopped checking on the outcomes, whose current operators are optimizing a metric that is at best loosely correlated with subject wellbeing, and for which there is no review board with the power to pause the study. Not one research group would be permitted to run this protocol on eighty human volunteers in a laboratory. It has been run on eight billion, without a single signature of informed consent, for three decades, and counting.
7 Conclusion
None of this argues for shutting anything down; that ship has sailed, and there is no version of the future that does not include networked computation. It argues for something more modest and more overdue: that the people who still have influence over how these systems are built and operated — including those now building the AI systems compounding the same effects — start behaving like investigators with subjects, rather than like vendors with customers. That means measuring harm honestly, reporting it publicly, and being willing to change course when the data says so. An experiment without anyone willing to check on it is not research. It is just something being done to people.