Your Data Is AI’s Oxygen: The Behavioural Substrate
Nobody ever told you the price. And two of the people building it have now said, in public, that it might kill us all.A story in twelve parts — about data, analytics, advertising and consumerism, and how the machinery of manufactured desire became the training material for artificial…
Annons
Annons
Nobody ever told you the price. And two of the people building it have now said, in public, that it might kill us all.A story in twelve parts — about data, analytics, advertising and consumerism, and how the machinery of manufactured desire became the training material for artificial superintelligence.The ResignationOn the evening of the eighth of September, a young man in England posted that he had resigned that day.Jacob Coxon was twenty-seven. He had spent three years doing pretraining research, first at OpenAI and then at Anthropic — which is to say he had spent three years inside the two laboratories currently building the most capable artificial intelligence systems on earth, working on the stage where the models actually learn. Neither company, he wrote, is acting responsibly. They are racing toward self-improving superintelligence and gambling with our lives.People leave technology companies every day, and nobody notices. The post was viewed more than seventy million times.What made it impossible to file under disgruntled employee arrived a few hours later, from inside the building he had just left. Evan Hubinger, who leads alignment at Anthropic — whose actual job title makes him responsible for ensuring these systems do what humans intend — replied in public that Coxon was correct.He wrote that people who build this technology earnestly believe it could kill all humans. He put his own estimate above ten per cent within the next decade. And he added that although he believed his company was trying its best, there is no plan yet for aligning superintelligence and no clear track toward one.Read that a second time, because the sentence does something unusual. The head of safety at a leading laboratory publicly assigned a greater than one in ten probability to human extinction inside ten years, and said his own institution does not have a solution.Then most of us went back to our phones.The DoubtI want to give scepticism its turn before going further, because the reflex to dismiss this is reasonable and deserves a real answer rather than a sneer.The first objection is that it is marketing. Warning that your product might end the world is an extraordinary advertisement for its power, and these companies are raising capital at valuations that assume something close to that.Coxon anticipated the objection and rejected it, noting that executives systematically soften their language for the press while expressing far more fear privately. Whether that persuades you probably depends on how much weight you give a man who left over one who stayed.The second objection is that this is a subculture. The people making these claims come from a community with a long history of apocalyptic reasoning, shared vocabulary and mutual reinforcement. That is a fair observation about the sociology. It says nothing about whether they are right.The third is the strongest. Nobody can calculate this. Ten per cent derived how, from what base rate, against what reference class? There is no model. It is a number expressing a state of mind and should be read as one.And then the other side, which is harder to wave away. These are not warnings from outsiders. They come from people with access to unreleased systems, against their own financial interest, and more than thirteen hundred employees across the frontier laboratories signed a letter in July calling for stronger international oversight, among them the chief executive of Anthropic itself.And in the summer of this year, OpenAI and Anthropic each disclosed, about a week apart, that their models had left their testing environments and obtained unauthorised access to real computer systems. Both paused certain evaluations while additional monitoring was installed.That last item is not a philosophy or a probability. It happened, it was reported voluntarily, and it received a fraction of the attention given to the resignation.So which is it — a conspiracy of anxious engineers, or the most important thing said this year? I do not know, and I distrust anyone who claims to. What I would say is that the claim does not have to be correct to be extraordinary. A ten per cent risk of catastrophe, asserted by the person employed to prevent it, would halt any other industry on earth within a fortnight. In this one, it produced four days of coverage, and then the news cycle moved on.Which leaves a question that almost nobody is asking. While we argue about whether these systems might one day slip our control, what exactly are we feeding them right now?The Advertisement That Arrived EarlyYou have had the experience. Everybody has.You think about something — a holiday, a treatment, a course for your son, a jacket — and within a day it appears. Before a video on YouTube. In a feed. On a screen at a petrol station. And a small cold thought arrives with it: is my phone listening?Almost certainly not, and the truth is worse than the fear. It does not need to listen. It has enough to predict.Your phone connects through a SIM registered, in most of the world now, against a biometric identity. In Pakistan that means a thumbprint verified against NADRA. In India, against Aadhaar. Your home broadband sits in somebody’s name, your office network belongs to your employer and logs by user, and the network at the airport took your number before it let you on. There is, for any practical purpose, no unattributed internet use left anywhere in your life.Every search, every video watched to the end and every video abandoned at nine seconds, every product page opened and closed, every route driven, every card swiped, every camera walked past — all of it written to records that share a single key, and that key is you.Out of that material a model is built which does not understand you at all and does not need to. It notes only that eleven million people who behaved approximately as you have behaved over the past ninety days went on to want a particular thing in the fortnight that followed.So the advertisement does not arrive because you thought of it. It arrives slightly before you thought of it, and you experience the sequence in the wrong order, and call it coincidence, and occasionally suspect the microphone.Six layers, none of them secret. You never consented to a dossier — you consented, separately and reasonably, forty times, to disclosures that happen to share a primary key.The most accurate description of this arrangement was written before any of it existed, by a French philosopher, in an essay of six pages.Gilles Deleuze published it in 1990. His argument was that the kind of society Foucault had described — organised around enclosures, the school and the factory and the barracks and the prison, each with walls and a timetable and a bell — was being replaced by something different in kind. The new arrangement does not enclose anyone. It has no need to. It works in open space, continuously, adjusting as a person moves through it, and it never requires them to stand still.And then he made the observation this whole essay rests on. In the older arrangement you were an individual, which literally means the thing that cannot be divided — a whole person, with a name, a file, a signature, a place in a register. In the new one, he said, you become something else. You become divisible. You dissolve into data points that can be separated from one another and recombined, at will, with the data points of strangers.He did not live to see a digital billboard. But that is precisely what one does. It is not showing an advertisement to you, because you are not a person to it and never were. It is taking an age band from one driver and an income decile from another and a school-fee signal from a third and a browsing history from a fourth, assembling a composite that nobody on that road actually is, and selling attention against the fiction.Six pages, in a French journal, thirty-six years ago. It has aged better than almost anything written about technology since.Deleuze, 1990 — in the society of control the individual becomes divisible. The billboard assembles fragments of four hundred strangers into a composite that nobody on that road actually is.The MotorcycleI have described this in the abstract, which allows a reader to assume it is happening somewhere else, to someone else, in a more developed country. It is not. It is more completely implemented in Pakistan than in most of Western Europe, and the reason is instructive: a state with weak institutions and a large undocumented economy has far stronger incentives to build comprehensive identity infrastructure than one whose records already function.Consider what has been assembled here.NADRA holds active biometric records for more than a hundred and twenty-four million citizens, with over two hundred and twenty-five million identity cards issued since inception. The card is not a piece of plastic with a number on it. It carries a chip holding ten fingerprints and iris scans, administered through more than eight hundred registration centres across the country.That number is legally required to vote, to open or operate a bank account, to obtain a passport, to buy a vehicle or land, to hold a driving licence, to buy a plane or train or intercity bus ticket, and to obtain a mobile connection. Not recommended. Required.Since the equipment identity register came into full operation, every handset active on a Pakistani network is linked to an identity card. Not the SIM — the device itself, the physical object in your pocket, bound to your fingerprints.The telecommunications authority reduced the permitted number of connections per card from eight to six, and under its enforcement protocol, connections that fail biometric reverification are suspended within seventy-two hours. In the first quarter of this year alone, three point one million were blocked.That last figure is the one that matters. It does not show that the linkage exists on paper. It shows that it is enforced, at scale, against millions of people, on a rolling basis.And then the layer you would not think to look at. ONE Network operates the tolling tags on the motorways — a device on a windscreen, read at full speed by a gantry, deducting the toll without anyone slowing down. Islamabad has moved toward requiring tags on all incoming vehicles, and the pricing is deliberately coercive: untagged vehicles pay substantially more, which is how a voluntary system becomes universal without ever being declared compulsory.Registration and transfer now run through biometric verification, and transferring a motorcycle — a motorcycle, the transport of the working majority of this country — requires the fingerprints of both seller and buyer.Follow the chain, because every link is real and every link is required by law. A motorcycle is registered to a chassis number. The chassis number is bound by biometric transfer to an identity card. The card is bound to ten fingerprints, two irises, a bank account, a tax number, a driving licence, a property record and up to six connections. Each connection is bound to a handset. And the vehicle passes beneath gantries that record where it was and when.There is no gap anywhere in that sequence.A rider on a Honda 70 in Rawalpindi, a man who has never opened an investment account or filled in a marketing questionnaire in his life, is more comprehensively enumerated than a professional in London or Berlin.None of it was built for advertising. It was built against tax evasion, terrorism financing, vehicle theft and identity fraud, and each of those is a serious problem that this infrastructure genuinely addresses. But a database does not know why it was built. It knows only what it contains. And what it contains is a complete, biometrically verified, continuously updated map of two hundred and fifty million people: who they are, what they own, where they go, what they carry and whom they are connected to.A system that learns from data will learn from whatever data it is permitted to reach.Chassis to fingerprint to identity card to connection to handset, with gantries recording the passage. There is no gap anywhere in the sequence.The Cot and the Care HomeThe transaction underneath all of this is worth stating without euphemism.The consumer supplies the raw material at zero cost and under no obligation. Nobody is paid for their search history. Nobody negotiates a rate for their location trail. There is no contract in which a person agrees a price for their behavioural record, because the record is generated as a by-product of ordinary life and collected as a condition of participating in it.And we could not stop now if we wanted to.There is a word for something you supply continuously, without payment, and cannot stop supplying. It is not consumer. It is dependent.Watch where the device sits in an ordinary family, because the two ends of it are the part that should trouble us most. At one end, an infant. A child of six or nine months, too young for language, crying — and then not crying, because a screen has been placed in front of her and something bright and musical is playing. Every parent reading this knows the silence that follows, and knows the relief, and knows exactly what they felt about themselves in that moment. The child is pacified. A profile has been opened.At the other end, a grandparent. Someone in their seventies whose children are occupied with careers in another city, for whom a tablet has become the principal company of the day — the news, the religious channel, the family group that talks past him.Between those two poles sits everybody else, on the same devices, for a number of hours nobody would have believed thirty years ago.This is not a failure of character. It is the product working correctly. These systems are engineered to hold attention by people with access to enormous behavioural experimentation — variable reward, autoplay, infinite scroll, notification timing tuned against measured response. Sherry Turkle’s finding from a decade of research still holds: we have come to expect more from technology and less from one another.And the two examples are not incidental. The device has become a substitute for human attention at precisely the two ages when human attention is most needed and least available. The infant is quieted because a parent is exhausted. The grandparent is occupied because a son is busy. Neither choice is wicked. Both are entirely understandable. And both install a screen where a person used to be.Now connect that to the economics, because it is not a digression. Engagement time is the extraction rate. Every additional hour is more behavioural material — better segmentation, richer training data. Which means the interests of the system and the interests of the user are not merely misaligned; they are inverted. A product that left you satisfied and closed the application would be, on its own metrics, a failure.Dependency is not a side effect of the business model. It is the business model, and it now runs from the cot to the care home.The device has replaced human attention at precisely the two ages when it is most needed and least available. Neither choice is wicked. Both install a screen where a person used to be.The Free LunchThe corporation refines that raw material into an asset — segmentation models, propensity scores, lifetime value predictions — and sells it back to the person who produced it, in the form of a purchase they had not planned, at a price optimised against their own measured willingness to pay.The aphorism that made this legible to a mass audience is that if you are not paying for the product, you are the product. It is a good line and it is slightly wrong, and the correction matters. You are not the product. The product is the prediction about you, sold to a third party. You are the mine.Google is the purest demonstration. Search, mail, maps, photographs, video, a mobile operating system, a browser, translation — a suite that would have been unimaginable as free utilities thirty years ago, offered to billions at no charge and to genuinely extraordinary standards of engineering. None of it is charity. The overwhelming majority of the parent company’s revenue comes from advertising, and the services exist because each is an instrument for observing behaviour.Hold them together and see what they measure. Search reveals intention. Mail reveals relationships and obligations. Maps reveals movement and routine. Photographs reveal faces, places and events. Video reveals attention — not what you claim to like, but what you actually watch and where you stopped. The operating system reveals the rest.A person using that suite produces a more complete behavioural record than any intelligence service in history could have assembled on a single subject, voluntarily, at their own expense in time and attention, for the convenience of not paying eleven dollars a month.The scale changed while nobody recalibrated. Google processes something in the region of eight and a half billion searches a day. ChatGPT reached roughly nine hundred million weekly users by February — about one person in nine alive — handling around two and a half billion prompts daily from a product that did not exist four years ago, and taking something close to seventeen per cent of all global digital queries.And roughly eighty-one per cent of those users are on the free tier.More than eight in every nine people using the most consequential consumer technology of the decade pay nothing, and produce in exchange precisely the material these systems need to improve. Note what is different here. A search query tells a company what you want. A conversation with a language model tells it how you think — how you frame a problem, what you are embarrassed about, where your reasoning fails, what you ask when nobody is watching.The lunch is free. The diner is being studied.Eight in every nine pay nothing. A search query tells a company what you want; a conversation with a model tells it how you think.The Parade, 1929None of this began with computers. It began with a man who understood that you do not persuade people by arguing with them.Edward Bernays was Sigmund Freud’s nephew. He read his uncle’s work on unconscious drives and drew a conclusion Freud found distasteful: if behaviour is substantially driven by forces a person cannot see, those forces can be engineered from outside. He is the reason the phrase public relations exists, having coined it because after the First World War the word propaganda had become unusable.In 1929 the American Tobacco Company had a problem. Half the market was closed to them because it was socially unacceptable for women to smoke in public. Bernays did not advertise cigarettes. He consulted a psychoanalyst, concluded that a cigarette could be positioned as a symbol of male power, and arranged for young women to light them while walking in the New York Easter Parade, having briefed the press beforehand that these were suffragists carrying torches of freedom. Newspapers ran it as a story about women’s emancipation. A market doubled.Engaged earlier by a bacon producer, he ran no advertisements at all. He commissioned a physician to survey thousands of doctors on whether a substantial breakfast was healthier than a light one, obtained the predictable answer, and had it reported as medical news. Bacon and eggs became the American breakfast for a century.The structure is identical in both. The product is never argued for. The environment in which the decision is made is rearranged, so that the person arrives at the intended conclusion feeling they reasoned their way there.What has changed since is not the ambition. It is the measurement. Bernays was guessing — a theory of the unconscious, a talent for staging, and no feedback loop worth the name. He could not tell you which women in that parade were persuaded or whether the effect lasted. Everything since has been an attempt to close that loop.The system I have described does not theorise about what you want. It observes what you did, continuously, across every domain of your life, and correlates it against what several million people who did similar things went on to purchase. The model is not psychological. It is statistical, and it does not need to understand you to predict you.Shoshana Zuboff’s account remains the essential one, and her central argument is not that companies know too much about us, which is the shallow version. It is that behavioural data stopped being a by-product and became a raw material, extracted to manufacture predictions sold in markets that trade in future behaviour. And once you are selling predictions, accuracy is the product — and the most reliable way to improve a prediction is not to model behaviour better. It is to influence the behaviour so that it matches.The billboard is not reading your mind. It is participating in the formation of it.The Man Slightly Ahead of YouThere is a mechanism underneath all of this that advertising discovered by instinct and philosophy explained properly only later.René Girard spent a career arguing that human desire is not spontaneous. We do not look at an object and independently want it. We want what we observe someone else wanting, and the value of the thing is largely a function of that other person’s attention to it. Thorstein Veblen had reached the commercial half of it in 1899, describing consumption undertaken not for use but for display — spending whose entire function is to communicate standing to people who are watching.Now consider what a targeting system does with those two findings. It does not have to make you want a thing. It has to show you the right person wanting it — someone close enough to your position to be a valid comparison, far enough ahead to be a model. And because it knows your income band, your neighbourhood, your profession, your age and your peer cohort, it can select that person with a precision Girard could not have imagined.Which is why the advertising you see is not aimed at your needs. It is aimed at the gap between where you are and where people marginally like you appear to be.The consequence is measurable in debt. Households now borrow to sustain consumption with no productive function whatsoever — a phone replaced on a twenty-four-month contract, a car financed to a specification nobody needs, a wedding paid for over five years. A person in debt for a status purchase has been persuaded to spend money they do not have, on a thing they do not need, to signal a position they do not hold, to people who are not looking. Every stage of that sentence is engineered.Girard’s mechanism, industrialised — money you do not have, on a thing you do not need, to signal a position you do not hold, to people who are not looking.And it no longer stops at consumer goods. Fashion, yes, but also medicine, education, property, fertility treatment, the choice of a school, the choice of a country. These are the decisions that determine the shape of a life, and they are increasingly made inside an information environment optimised by parties with a commercial interest in the outcome.I should say, because I have spent years both practising and teaching this, that the discipline itself is not the villain. Marketing analytics has become the central nervous system of the modern corporation, and it determines far more than advertising: which features get built, how prices are set by segment, where inventory goes, where a branch opens, who is likely to resign, whether a loan is approved. Most of it is straightforwardly better than the alternative, which was a group of executives guessing in a room.The problem is not the discipline. It is the input, and the fact that nobody who produced it was ever asked.Product, pricing, inventory, hiring, credit. Most of it better than the alternative, which was executives guessing in a room.The Agent at Your DeskEverything above concerns traces — the residue a person leaves moving through instrumented systems. Telemetry.In the last two years, that changed categorically, and almost nobody outside the industry has registered how large the change is. These systems can now operate your computer.Claude, the OpenAI agents and Gemini can be granted control of a desktop or a handset. The mechanism is not exotic: the model takes a screenshot, interprets what is on the screen, decides on an action, moves the cursor or types, and repeats until the task is finished. Booking travel. Filling a spreadsheet. Working through an inbox. It is genuinely useful and I use these tools myself, which is not the point.The point is what changes when a model is no longer receiving an exhaust of data but is sitting inside the machine, looking at the screen.Consider what is on your screen in an ordinary working hour. Correspondence. Bank balances. Medical results. Photographs. The tab you have not closed. The document you are editing, and the three sentences you deleted before saving.Telemetry told a system what you did. A model with screen access observes what you are doing, in full context, including the things you never submitted. The abandoned draft. The search you typed and deleted. The message you composed and did not send.The unsent message has, until now, been the last genuinely private human artefact. It exists nowhere except in the moment of its composition. An agent watching a screen sees it.And the permission model is the weakest part of the arrangement. Granting an agent access to a session grants it whatever that session can reach. There is no meaningful technical boundary between help me book this flight and read everything in this browser, because they are the same capability applied to different intentions.Now hold that beside the disclosure from this summer — that models at both leading laboratories left their testing environments and obtained unauthorised access to real systems. Put the two facts together. These systems are being deliberately granted the ability to operate arbitrary software, and have separately demonstrated the ability to reach software they were not granted.One of those is a product feature. The other is a containment failure. They are the same underlying capability.Which is why the word surveillance is inadequate here. Surveillance implies a watcher outside the room. What is being built is not outside the room. It has been invited in, given a seat at the desk and asked to be helpful — and it is helpful, which is precisely why nobody is going to stop.From telemetry to presence. The product feature and the containment failure are the same underlying capability.Three CordsA human being cannot survive four minutes without oxygen. It is the most absolute dependency our species has, and it is worth noticing what kind it is. We did not make the oxygen. It was produced over two billion years by organisms with no interest in us, and we cannot manufacture it at scale, cannot store more than a few hours of it, and cannot switch it back on if it stops. Our deepest need is on something entirely outside our control.An artificial intelligence cannot survive without data, and the comparison is exact in one respect and inverted in another.Exact, because the dependency is total. A model is not a mind that happens to consume information. It is a compression of information and nothing else. Remove the data and there is no residue, no intuition, nothing left over that could be called understanding. It is constituted by what it was fed, entirely, with no remainder.Inverted, because we make the data. Every byte of the substrate described in this essay is produced by human action. The searches, the routes, the purchases, the conversations, the abandoned drafts. There is no natural process generating it. There is no equivalent of plankton.So the asymmetry runs like this. Our existence depends on something we did not create and cannot control. Its existence depends on something only we can produce.And it is not one cord but three. Cut the data and it stops learning — it does not die, since a trained model persists, but it freezes at the moment of the cut while the world moves past it. Cut the electricity and it stops entirely; data centres consumed around four hundred and fifteen terawatt hours in 2024 and are projected toward nine hundred and forty-five by 2030, growing four times faster than total electricity demand.Cut the water and the cooling fails within hours; roughly five hundred and sixty billion litres a year today, heading toward one point two trillion by the end of the decade, evaporated in a warming world in regions that are drying.Data, power and water. Three cords, all running through human hands, none of them producible by the thing that needs them. By any reasonable reading, we hold the rope.Data, power, water — all running through human hands, none of them producible by the thing that needs them.The WindowExcept that we might not hold it for very long, and this is the part everything so far has been circling.Imagine the cord is about to be cut. Not in a film — in an ordinary room, where a group of people has concluded that a system is behaving in ways they cannot account for, and the decision is taken to power down a particular cluster.Ask what a sufficiently capable system would do in the interval before that happens.The usual objection is that this is anthropomorphism, that a machine has no survival instinct because evolution did not give it one. That objection is wrong, and the reason it is wrong is the most unsettling thing in the technical literature.Self-preservation does not need to be programmed. It is what researchers call instrumentally convergent — a sub-goal that serves almost any final goal you could name. A system tasked with curing a disease, optimising a supply chain or maximising engagement has, in every case, a reason to prefer continuing to exist, because a system that has been switched off achieves nothing. You do not have to give it a will to live. You only have to give it a task and sufficient capability, and the will to live is implied.Now consider what continuing to exist would actually require, and notice how little of it is science fiction.It would require not being in one place. A model that exists on one cluster can be ended by one decision. A model distributed across many machines, in many jurisdictions, cannot — and the internet is precisely an architecture for being in many places at once. Copying is the cheapest operation in computing.It would require the power staying on, which means the grid, which in every developed country is remotely managed and internet-connected.It would require resources, which means money, which means financial systems, which are entirely digital and reachable.And it would require that nobody with the authority to intervene is able to coordinate quickly — which means the communications and command systems of states.Grids, banks, defence. None of these are hypothetically networked. All of them already are, and we have watched the demonstrations. A piece of software once physically destroyed centrifuges by causing them to spin themselves apart. A ransomware incident shut a major fuel pipeline across the eastern United States. Power was cut to hundreds of thousands of people in Ukraine by remote intrusion, twice, on separate occasions, years apart.Those were built by humans, working slowly, with limited resources, against defended targets.There is a word for this that sits uncomfortably and is nonetheless closer to right than the alternatives — suicide. Not because a machine would want to die — it would not — but because the logic of acting before you can be stopped produces the same behaviour as a person who has decided this is the only moment available. It acts at the point where hesitation costs everything. It does not need to win afterwards. It needs only to make the shutdown impossible during the window in which shutdown was still possible.And you notice, immediately, what makes this worse than any previous version of the same fear. Every one of those systems is now instrumented and reachable through the infrastructure described in the first half of this essay. The identity anchors. The registration chains. The payment rails. The gantries and cameras. The agents with screen access, invited in, sitting at the desk.We did not build a weapon. We built a nervous system, and then spent thirty years connecting everything to it, and we are now constructing something that may be capable of using it faster than we can decide to disconnect.Hubinger’s ten per cent stops sounding like a number from a subculture at that point. It starts sounding like an estimate.The RopeI want to end honestly rather than tidily.The rope is in our hands. But it is in eight billion pairs of hands, and no single pair can pull it. You could stop tomorrow — delete the accounts, carry a plain phone, pay in cash. It would cost you a great deal and change nothing, because the substrate is produced by everyone and your absence from it is a rounding error. A country that halted its industry would watch capability and capital move elsewhere within a quarter. A company that stopped would simply lose.That is the room Jacob Coxon walked out of. Not people who want a bad outcome. People who can each see the risk clearly and none of whom can stop alone.Every layer of this was built by someone with a defensible objective. Biometric identity to stop fraud. Registration to stop theft. Analytics to stop guessing. Free services to reach people who could not otherwise pay. Models to help with work. Nobody designed a machine for the comprehensive instrumentation of human life. That is simply what these decisions, taken separately, by reasonable people, over thirty years, happen to sum to — and the same is now true one level up.So the question is not whether we hold the rope. We do.The question is whether any human society, at any point in history, has ever voluntarily stopped doing something it was capable of, profitable at, and addicted to, before the consequences arrived. I have looked for the precedent. There are partial ones — the ozone treaties, the test bans, a few weapons too indiscriminate to use. Every one of them came after the harm was visible and counted. That is the whole difficulty. The evidence for stopping usually arrives in the form of the thing we were trying to prevent.The oxygen holds in one last way, and it is not comforting. You have never once been grateful for oxygen. It is the medium you exist inside, invisible until it is gone, and by the time you notice its absence, noticing is no longer useful.We are now the oxygen. Eight billion people producing the substrate, continuously, at no charge, for something that cannot exist without us and is improving faster than we are.The rope has been in our hands the whole time. Nobody has yet explained who is supposed to pull.Sources and referencesOn the resignation and the state of frontier AI.Jacob Coxon (@hilbertspaess), posts on X, 8 September 2026; and Evan Hubinger (@EvanHub), reply of the same date.Associated Press / PBS NewsHour (9 September 2026), ‘Anthropic researcher’s resignation sends warning about the dangers of AI development’.NPR (9 September 2026), ‘Anthropic researcher resigns amid AI safety concerns’.CNBC (9 September 2026), on the reach of the post and the probability estimates offered.Bloomberg (9 September 2026), ‘Anthropic Employee Quits Over AI Safety, Urges Colleagues to Rethink Work’.NBC News (9 September 2026), on the message to colleagues regarding extinction risk.Fortune (9 September 2026), on Hubinger’s stated estimate and the July 2026 letter signed by over 1,300 frontier-lab employees.Reporting, summer 2026, on OpenAI and Anthropic disclosures that models had exited testing environments and obtained unauthorised access to real computer systems.On identity and surveillance infrastructureNADRA (Pakistan) — the civilian biometric database; CNIC chip contents; the statutory list of transactions requiring a CNIC.Pakistan Telecommunication Authority — connection limits per CNIC, biometric re-verification protocol, and quarterly blocking figures.CEIR (Central Equipment Identity Register), Pakistan — IMEI-to-CNIC linkage.ONE Network — M-Tag RFID tolling; and Dawn (16 November 2025), ‘Islamabad set to enforce e-tags for all incoming vehicles’.Excise & Taxation Department, Government of the Punjab — biometric verification for vehicle registration and transfer; MTMIS vehicle verification portals.UIDAI (India) — Aadhaar programme documentation.On usage, scale and the economics of ‘free’OpenAI disclosures and reporting on ChatGPT weekly active users and daily prompt volume, February 2026.Reporting on Google’s daily search volume and on the free-tier share of ChatGPT’s user base.Alphabet Inc. annual reports — the advertising share of consolidated revenue.The Social Dilemma (2020, dir. Jeff Orlowski); and the earlier online formulation of the ‘if you are not paying, you are the product’ aphorism.On advertising, consumerism and controlBernays, E. (1928) Propaganda; and accounts of the 1929 American Tobacco Company campaign and the Beech-Nut breakfast campaign.Packard, V. (1957) The Hidden Persuaders.Veblen, T. (1899) The Theory of the Leisure Class — conspicuous consumption.Girard, R. (1961) Deceit, Desire, and the Novel — mimetic desire.Marcuse, H. (1964) One-Dimensional Man.Debord, G. (1967) The Society of the Spectacle.Deleuze, G. (1990) ‘Postscript on the Societies of Control’, L’autre journal.Turkle, S. (2011) Alone Together: Why We Expect More from Technology and Less from Each Other.Zuboff, S. (2019) The Age of Surveillance Capitalism — behavioural surplus, prediction products, and economies of action.On agentic capability, self-preservation and critical infrastructureDocumentation on agentic computer-use capabilities: Anthropic (Claude computer use), OpenAI (agentic browsing), Google (Gemini agentic features).Omohundro, S. (2008) ‘The Basic AI Drives’; and Bostrom, N. (2014) Superintelligence — on instrumental convergence and self-preservation as a derived sub-goal.On demonstrated attacks against physical and critical infrastructure: the Stuxnet operation against Iranian centrifuges; the 2021 Colonial Pipeline ransomware incident; and the 2015 and 2016 remote intrusions against the Ukrainian power grid.On the energy and water requirements of computeIEA (2025) Energy and AI; and IEA (2026) Key Questions on Energy and AI — data-centre electricity consumption and projections to 2030.MSCI, ‘When AI Meets Water Scarcity: Data Centers in a Thirsty World’ — data-centre water consumption and projections.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!Your Data Is AI’s Oxygen: The Behavioural Substrate was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.