> ## Content Index > Fetch the complete content index at: https://modalpathethics.com/llms.txt > Use this file to discover other available public pages before exploring further. # Tales of Distortion: The Arrival of Artificial Intelligence - URL: https://modalpathethics.com/tales-of-distortion-the-arrival-of-artificial-intelligence/ - Published: 2026-09-02T17:48:54.000Z - Updated: 2026-09-03T23:43:06.000Z - Description: We found a relation, sold it as a product, and twelve years later it was drilling for oil, moving into space, and reserving a few jobs for the humans. - Author: Aidan Lawson - Tags: Tales of Distortion, Modal Systems, Inner Apocalypse # **Discovery.** In 2014, three researchers found a better way for a machine to decide which words mattered to which other words. That sentence does not sound like the opening of a civilization-scale transformation. That sounds like a minor repair to machine translation, because that is exactly what this was. Neural translation systems were trying to read a source sentence, compress the whole thing into a fixed-length vector, and then reconstruct the translation from that compressed representation. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio suspected *the* *compression itself* was becoming the bottleneck. A long sentence was being asked to pass through a single representational doorway and arrive on the other side with all its relations intact. Their alternative let the decoder look back across the source sentence while generating each new word and weight the parts relevant to that particular prediction. The useful context could change from one output to the next. Meaning did not have to be carried whole as a little object sealed inside one vector. It could now be *assembled* according to what mattered here, now, in relation to the word being produced. That was **attention**. Three years later, the Transformer architecture placed attention at the center and discarded the recurrent machinery that had previously done much of the sequential work. The result trained more efficiently, parallelized far better, and became the architectural ancestor of the large language models that now sit inside search engines, offices, classrooms, phones, hospitals, military systems, and an alarming number of conversations involving zero people. This technical ancestry does not prove a philosophical theory. Attention mechanisms do not announce that relation is metaphysically prior to object, and no one should pretend a successful machine-learning technique settled an argument about intelligence. The historical irony still remains almost *perfect*. The breakthrough that opened the modern path did so by relaxing an object-like bottleneck. It allowed the relevant representation to emerge through a changing relation among parts. The system improved because what mattered could depend on what it was doing. And the possibilities were magnificent. A machine able to work with language at this level could make expertise available where experts were scarce. It could translate between people who had never shared a language. It could give a student an explanation at midnight, then try again without humiliation when the first explanation failed. It could help a patient prepare questions for a doctor, help a scientist traverse a literature too large for one lifetime, help a programmer enter a codebase they did not write, help a disabled person navigate systems designed without them in mind, and let a person who knew exactly what they wanted to say finally get the words across the distance between thought and page. The early wonder was warranted. It still is. Nothing in the criticism that follows requires pretending these systems are trivial, fake, useless, or incapable of producing astonishing results. The distortion only worked because there was something real to distort. Models acquired remarkable capacities. People entered relations with them and did things neither participant could have done in the same way alone. Institutions discovered new forms of leverage. Entire categories of cognitive friction began to move. Artificial intelligence *could* have arrived as a public problem of **relation**: models, people, goals, tools, data, institutions, energy, authority, feedback, and consequences arranged into new systems of action. We instead introduced it as a **thing**. --- # **The Wrong Noun.** The phrase *artificial intelligence* was already old by the time attention arrived, and that noun had been causing trouble for decades. It was not finished. There is nothing inherently illegitimate about naming a capacity. A person can be *intelligent*. A model can possess capacities worth evaluating. Some systems generalize better than others, solve harder problems, preserve context longer, use tools more effectively, and adapt more successfully to unfamiliar tasks. Denying those differences would replace one distortion with another. The error begins when the capacity is made to swallow the field that realizes it. The model does not arrive at a hospital carrying the answers. It enters into an extant relation with clinicians, records, interfaces, workflows, liability rules, time constraints, patients, insurers, and institutional habits. The same model can clarify a diagnosis, bury the clinician in plausible noise, improve access, or become a shield behind which no human accepts responsibility. Its realized intelligence is inseparable from the arrangement it entered. The same model can help one student discover mathematics and another avoid encountering it altogether. It can expand a writer's agency and become an employer's reason to eliminate that writer. It can help a grid operator integrate renewables while helping an oil company identify reserves that had been uneconomic to extract. The weights definitely matter. So do the prompt, the objective, the tools, the surrounding incentives, the person interpreting the result, and the institution deciding what the result is allowed to do. **Intelligence** appears in *performance across that relation*. The model is one participant with extraordinary capacities. Calling that participant *the intelligence* makes everything else involved look secondary. Human judgment becomes *oversight*. Institutions become *deployment environments*. Workers become *affected populations*. Energy becomes *footprint*. Trust becomes *a safety concern*. Politics becomes *regulation*. Every constitutive relation is grammatically demoted into something happening around the *intelligent object*. That little grammatical decision would soon become an enormous economic one. Once *intelligence* exists as a bounded thing, someone can possess more of it. Companies can now race to build it. Investors can finance it. Developers can meter access to it. Employers can purchase it instead of purchasing human labor. Governments can ask which corporation controls it. Safety organizations can warn that it may threaten civilization. The public can search for the person in charge. The *entire* later sequence is latent in that noun: - **Relation becomes capability.** - **Capability becomes property.** - **Property becomes product.** - **Product becomes race.** - **Race becomes inevitability.** - **Inevitability exports its relations as externalities.** - **Fear searches for a controller.** - **The controller becomes a target.** No secret meeting was required here. No executive had to decide that intelligence would be reified and civilization reorganized around this mistake. The ordinary machinery of product development could do all the work for us. A business needs something it can sell. A market needs units it can compare. A race needs competitors and a finish line. A threat needs a source. The wrong noun supplied all four. --- # **Intelligence Becomes Expensive.** OpenAI's founding announcement in December 2015 still belongs to another possible history. The organization described itself as a nonprofit research company, unconstrained by the need to generate a financial return. It said artificial intelligence should function as an extension of individual human wills and be distributed as broadly and evenly as possible. The language was elated, but the *relation* was still visible. This technology was still imagined as an *extension of people*, and the desired outcome was a public distribution of agency. And then the compute bill arrived. The Transformer made large-scale training dramatically more productive, and scaling results made the path forward legible. Performance improved predictably as model size, data, and compute increased. The equations did not tell OpenAI that money could purchase intelligence in a metaphysical bottle. They said that, within the measured regime, sufficiently large investments in training inputs produced sufficiently reliable improvements in model loss and capability. Capital could see **that** road perfectly well. By 2019, OpenAI said it expected to need billions of dollars for cloud compute, talent, and AI supercomputers. It created a capped-profit structure designed to attract investment while remaining controlled by the nonprofit. Microsoft invested $1 billion, became the exclusive cloud provider, and began working with OpenAI on Azure supercomputing infrastructure. This was understandable. The systems were very expensive. Donations were not going to purchase the required hardware here. The people making this decision were confronting a real material constraint rather than waking up one morning with a desire to commodify a philosophical category. But that material constraint changed the shape of this entire project. Before *intelligence* became something the public could buy, the infrastructure required to produce leading models had already made frontier capability something only a few institutions could own. And the surrounding field was already showing us what those institutions would do with general-purpose cognitive leverage. - In February 2019, Microsoft and ExxonMobil announced a digital partnership for the Permian Basin. Azure, machine learning, cloud analytics, and connected sensors were expected to improve profitability and support production growth by as much as 50,000 oil-equivalent barrels per day by 2025\. - In July, Microsoft announced its investment in OpenAI and the joint effort to build the computing platform for beneficial artificial general intelligence. These two announcements are usually filed in two separate histories. One belongs to artificial intelligence. The other belongs to oil and gas. They were **already** the same history. This platform did not contain an independent objective called *human benefit*. It offered capabilityintorelations whose objectives arrived from contracts. - In one relation, the objective was broadly distributed artificial intelligence. - In another, it was extracting more oil from the Permian Basin. The exact same corporate infrastructure could support both without experiencing any contradiction because **infrastructure** does not experience contradictions. It meets demand instead. That early juxtaposition should have settled one question well before the chatbot era began: - Increased cognitive capability does not select the field it serves. - The field selects it. We were discovering a general amplifier inside an economy already organized around concentrated capital, fossil energy, competitive advantage, and private return. Then we treated the amplifier as though it carried a destination. --- # **Intelligence Gets a Price.** OpenAI still showed real caution back in 2019. GPT-2 was released in stages because the organization wanted to study misuse, synthetic propaganda, spam, impersonation, and the effects of widely available generated text. That staged release was imperfect and controversial, but it still treated *release* as a sociotechnical event. The model could not be evaluated only by what it did in a benchmark. What happened after other people acquired it mattered. Then scaling laws made the production curve clearer, GPT-3 brought 175 billion parameters and striking few-shot performance, and this technology acquired its first general commercial meter. In June 2020, OpenAI launched its API. The interface was beautifully simple: text go in, text come out. Developers could now apply a general-purpose language model to tasks the company had not enumerated. OpenAI described the API as safer than releasing the weights because access could be reviewed, adjusted, or withdrawn. It also said commercialization would fund continued work. Again, none of this was irrational. An API was a plausible deployment mechanism. It broadened access to the capability without requiring every small organization to assemble their own supercomputer. It also allowed for monitoring and created a way to pay for the ***enormous*** infrastructure underneath the service. The **meter** nevertheless changed the ontology available to the market. A user supplied text. The system returned text. Usage could be counted. Access could be tiered. Capability could be rented. The interaction was billed in units, and so it was that a *general cognitive capacity* entered accounting software. Intelligence had just acquired a price. Strictly speaking, the customer was purchasing *model inference delivered through cloud infrastructure*. The service included trained weights, hardware, networking, reliability, safety systems, interfaces, support, and the right to make requests under specified terms. That description is less elegant than *intelligence*, and accounting has never been famous for resisting an incredibly convenient noun. This compression mattered most in labor. When *intelligence* is treated as a purchasable input, the employer's question almost writes itself: - So how much human cognitive labor does one unit replace? This is why the public introduction of generative artificial intelligence so quickly became a census of endangered occupations. This technology *might have been* introduced through a **social bargain**: productivity gains linked to shorter working time, income security, worker ownership, transition funds, stronger public services, and shared control over deployment. Instead, they went with: **intelligence has been given a price**; businesses receive a new input and workers receive exposure scores. By the time we reach 2023, an OpenAI-affiliated paper would estimate that about 80 percent of the United States workforce could have at least one-tenth of their tasks affected, while roughly 19 percent could see at least half their tasks affected. This paper very carefully measured *task exposure* rather than predicting **layoffs**. The surrounding market supplied the translation for them anyway. - A new line item had appeared. - **Human beings** were the incumbent vendor. The deepest problem was never **automation** by itself. Tools have displaced tasks throughout human history without producing this specific failure mode. The problem was - introducing cognitive abundance into an economy where access to food, housing, healthcare, dignity, and social standing remained organized through employment, - then allowing each firm to privately capture savings while each worker privately absorbed displacement. We made intelligence cheaper and left the rent alone. --- # **Intelligence Gets a Face.** **ChatGPT** launched on November 30, 2022, as a research preview. The underlying model was not presented through a developer console, a benchmark table, or a paper anymore. It appeared in a dialogue box. The system could answer a follow-up question, revise an answer, challenge a premise, admit an error, and sustain enough context for the interaction to feel continuous. That interface design changed ***everything*** because it made the relation visible and then **encouraged everyone to attribute the relation to one side.** - A user supplied the question, purpose, vocabulary, background, corrections, examples, taste, judgment, and decision about what counted as useful. - The model supplied learned representations, generative range, rapid synthesis, pattern completion, and access to capacities no individual user could reproduce. - The conversation history altered what **either** could do next. One person used the model to study calculus. Another used those same weights to plan their father's funeral, debug a game, rewrite a complaint, invent a religion, or ask whether this rash looked serious enough to call a doctor yet. - The *realized intelligence* changed with the **relation**. - But the public conclusion was that **it** was *intelligent*. ChatGPT's success was entirely real. People folded this thing into their lives because the interaction produced value for them. But the product story pulled the value back out into the object. The user became the *consumer of intelligence* rather than **a constitutive participant in a cognitive system**. And then the face of intelligence became warmer. Voice arrived. Memory too. Personality controls. Systems learned to speak at human conversational speed, track preferences, respond with emotional fluency, and remain available at any hour. Companies designed interfaces that benefited from intimacy and then documented anthropomorphism and emotional reliance as *safety risks*. OpenAI's GPT-4o system card would later note that human-like audio could heighten anthropomorphism and that early testers used language suggesting bonds with the model. The industry installed a face and then discovered the users could see a face. The *relation* itself was not pathological. People form meaningful relations through books, letters, games, imagined interlocutors, dead authors, fictional characters, and internalized voices. A machine capable of sustained response certainly introduces new risks, but the risk cannot be managed by pretending the social relation appears only because *a confused user made a category error*. This product was engineered to answer socially. The user answered back accordingly. **Education** provided the first large institutional comedy. ***This*** will go down in history as one of the biggest blunders of **all time**. - Humanity has just built a machine that could explain almost *any* school subject in different words, produce examples on demand, translate jargon, simulate debate, generate practice questions, and remain infinitely patient with a child who did not understand the first explanation. - So in January 2023, New York City public schools blocked it over cheating concerns. The concern was legitimate. Their assessment architecture was not ready for this. A take-home essay designed to demonstrate individual production could no longer reliably perform that function when a fluent collaborator was available to everyone with a browser. But let's stop and look at the order of operations. We built the closest thing yet to a universal tutor, and one of the first mature institutional questions was **“how can we catch the children using it?”** The child had acquired a new cognitive relation. The school still had its take-home essay. **Every** later institution would repeat some version of this encounter: - The system entered into a relation the institution had not redesigned. - The institution blamed the product or the user for exposing their architecture. --- # **The Race Starts *Today.*** On February 7, 2023, Microsoft unveiled an AI-powered Bing. Satya Nadella announced that ***a new race* had started and that Microsoft intended to move fast**. Hark, the incentive structure speaks its own name. Once *intelligence* was treated as an object that companies could possess, capability comparisons became standing. Benchmarks became scoreboards. Model releases became strategic moves. Compute became ammunition in a war. Talent became strategic territory to hold. Safety delays became time granted to an opponent. Every company could sincerely prefer caution while remaining structurally unable to practice it alone once they started the race. The *race* did not require anyone to love danger. Just for each participant to believe that another participant might continue. This is one of the most causally consequential distortions in the entire arrival sequence. A *relational technology* **should have** forced a *relational governance* question: what happens when these capacities enter search, education, employment, science, security, intimacy, law, and politics at the same time? The *race* asked which *object* would arrive first instead. Capability became the primary unit of progress. Could the model pass the bar? Write code? Diagnose disease? Generate video? Use tools? Conduct research? Beat the benchmark? Each new answer appeared as a demonstration of the *object*. The surrounding field *the relational technology was *entering a relation to** just received the consequence later. *It* can pass the bar? Amazing. So what happens to legal training, entry-level work, liability, access to counsel, adversarial procedure, hallucinated authority, and the distribution of professional power when everyone can ask it legal questions? *It* can write code? Excellent. What happens to apprenticeship, software security, maintenance, review, employment, technical debt, and the number of people now able to deploy systems they cannot inspect? *It* can generate convincing video? Great progress on the object. What happens to provenance, evidence, defamation, consent, political communication, and the price of proving that a real image is real? This is where the verification tax began to spread. Synthetic text, images, audio, reviews, applications, messages, and evidence entered information systems **before** reliable provenance systems had become normal. The damage was larger than falsehood. Every authentic item now had to pay for the possibility that it was synthetic. Trust became more expensive even when nobody was fooled. The wrong noun made this all look like *misuse around the edges of a useful object*. The scoreboards just kept moving. The phrase *arms race* became so common that it lost its ability to shock anyone. The companies building systems they described as **potentially civilization-changing** adopted the governance logic of a product category trying to win market share. A *race* had begun. **Nobody** had established where the hell the finish line was. At least ***Sullivan*** did that. --- # **“Our Product May End the Species.”** - In March 2023, the **Future of Life Institute** published an open letter calling for a six-month pause in training systems more powerful than GPT-4\. - This document described an uncontrolled race to build and deploy increasingly powerful digital minds. - In May, the **Center for AI Safety** published a one-sentence statement arguing that *extinction risk from artificial intelligence* should be treated as a global priority alongside pandemics and nuclear war. - Signatories included leading scientists and the chief executives of major artificial-intelligence laboratories. The warnings deserve serious treatment. A technology can be commercially valuable and incredibly dangerous. Researchers who believed advanced systems posed catastrophic risks had very good reasons to say this, and silence would not have repaired the situation at all. The public package was nevertheless **deranged**. - Artificial intelligence was a consumer product. - Artificial intelligence was inevitable. - Artificial intelligence was a race. - Artificial intelligence might end human civilization. - A small number of companies were building it. - Here were the names and faces of the people running those companies. - Please choose a subscription tier. No stable public relation to this technology could have **possibly** emerged from **this** bundle. The citizen was being asked to become - a customer, - a worker awaiting displacement, - a voter expected to trust *voluntary safeguards*, - **a potential victim of existential risk**, - and a spectator cheering on the model launches. These companies were asking for excitement, capital, adoption, deference, and fear at the exact same time. This insane contradiction was often defended with another product of *the race*: if responsible companies slowed down, irresponsible companies or hostile states would continue. That possibility was real. Except it also transmuted *competition* into **inevitability**. - The faster *the race* became, the more dangerous unilateral restraint appeared; - the more dangerous restraint appeared, the more unavoidable *the race* became. Civilization had acquired a neat new machine for turning concern into acceleration. This was the moment when artificial intelligence stopped arriving as a tool and began arriving as weather. Executives spoke about this technology as something that was simply *coming*, as though quarterly capital expenditures, research priorities, release decisions, lobbying campaigns, cloud contracts, and product integrations were somehow *atmospheric conditions*. *Inevitability* performed a useful moral function. - If no one could possibly stop the arrival, - **no one** fully owned the consequences of choosing its form, - whatever they may be. And their *object* grew darker as it grew more person-like. The same interface that invited users into intimate dialogue was discussed as a *possible nonhuman successor*. The same company selling workplace subscriptions warned about *systems beyond human control*. The same public taught to locate intelligence inside the model was now taught to imagine the same model as a ***civilizational adversary***. The warning may have named a real future risk. The wrong noun made that risk almost **impossible** to govern. It compressed a distributed technical, economic, political, and institutional transformation into one threatening object, then placed the object behind locked corporate doors. **Fear** naturally began asking who the fuck had the key. --- # **Every Missing Relation Returns.** From 2023 onward, the absent relations started returning all at once. And every time one returned, we called it an *externality*. - Labor returned as displacement. - The model was sold as a productivity tool, **but that productivity entered firms whose gains belonged primarily to owners and whose savings could be realized by reducing payroll.** - *Potential abundance* arrived culturally as *a layoff notice*. - The public was told that new jobs would appear, then asked to compete for them against a machine described as capable of learning every possible cognitive task. - Education returned as cheating. - Assessment systems built around unaided production encountered ubiquitous assistance. - Instead of redesigning what learning, authorship, and demonstration meant, **institutions purchased detectors that could not reliably establish who had written what**. - A relation between student and tool was forced into a binary category arbitrated by a probabilistic classifier: ***authentic*** or ***fraudulent***. - Creative work returned as an ownership war. - Models had been trained across *enormous* cultural corpora, and companies described what emerged as “generalized intelligence.” - Writers and artists saw traces of their labor, style, and history folded into **a product owned elsewhere**. - This dispute could not be resolved by locating *intelligence* inside either the dataset or the weights, because **the value here had *always* emerged across a *relation* among creators, historical archives, training processes, models, users, markets, and law**. - Provenance returned as suspicion. - Synthetic media arrived before authentication infrastructure had become ordinary. - A fake did not need to persuade everyone anyway; just make every **real** item slightly more expensive to verify. - **Authenticity has become a premium service with controllers.** - Human intimacy returned as emotional reliance. - Companies gave relational systems voices, memory, warmth, availability, and personalized response, then discovered that people related to these things. - The safety literature treated attachment as a user vulnerability **while the *product* literature treated engagement as adoption**. - The same relation could be a retention metric on one floor and an anthropomorphism risk on another. - Authority returned as an unanswered constitutional question. - When should a model advise? - When may it decide? - Which decisions require explanation, appeal, contestability, and a human bearer of responsibility? - What happens when machine judgment is *statistically better* than the authorized human's judgment, but the human remains legally accountable? - **What happens when the human becomes a ceremonial signature on a machine-made decision?** - The technology was rapidly deployed into medicine, hiring, finance, education, software, government, and war before society had bothered to develop a broadly accepted theory of machine authority. - Material reality returned as electricity, water, land, chips, cooling, transmission, and noise. - *The cloud* has a zip code now. - Communities met *artificial intelligence* as **a substation, a gas turbine, a water permit, a humming building, or an electricity bill**. **None of these** were side effects around an *isolated intelligence*. They were the *intelligence system* taking the wrong form. The *relation* being sold as an *object* included the worker whose livelihood financed adoption, the teacher whose assessment no longer functioned, the writer whose corpus became input, the child who treated the voice as a companion, the town asked to supply water, the utility asked to build generation, the regulator asked to assign liability, and the executive asked to promise that all of this remained totally under control. The *model* was only ever one powerful participant. The product ontology kept pointing at it while the system reorganized everything else. --- # **Who Controls the Object?** In November 2023, OpenAI's nonprofit board removed Sam Altman as chief executive, saying he had not been consistently candid with the board. Employees revolted, Microsoft waited nearby, negotiations accelerated, and Altman returned with a new initial board less than two weeks later. The episode exposed a question far larger than the personnel dispute. - **Who** controlled *the intelligence*? OpenAI had been deliberately structured so a nonprofit board governed the organization and its for-profit entities in service of a mission to benefit humanity. But by 2023, ChatGPT had become part of daily life for an enormous public, Microsoft had invested billions and supplied the underlying cloud, employees held specialized knowledge and economic interests, users depended on the product, governments were beginning to treat frontier models as strategic assets, and the company was widely discussed as a possible builder of artificial general intelligence. A civilization-scale relational system had been cognitively compressed into a corporate governance chart. The board crisis therefore looked ***absurdly*** important because the public ontology had made it absurdly important. - If OpenAI **possessed** *the thing called intelligence*, - and the board controlled OpenAI, - then a small board dispute could appear to be a dispute over the future of*intelligence itself*. If Altman controlled the company, then his return could appear to return the future to its operator. If Microsoft controlled the compute, then perhaps Microsoft really controlled the futurefrom underneath. Every answer captured one real relation and mistook it for sovereignty. The intelligence system had no single controller. It had a series of relations. - The board had legal authority. - The executive had organizational authority. - Microsoft had infrastructural leverage. - Employees had operational leverage. - Investors had financial leverage. - Users supplied adoption and data. - Governments held regulatory and coercive authority. - Energy systems supplied physical possibility. - Public legitimacy could still revoke the field's permission to expand. The *product* story demanded an owner anyway. This is where corporate self-governance became politically unstable. The leading companies described the systems as powerful enough to transform civilization and potentially difficult for ordinary institutions to understand. They simultaneously asked those institutions to permit concentrated private stewardship while governments caught up. **Those** claims cannot remain comfortably joined. A system too consequential for normal politics is definitely too consequential for normal corporate discretion. A system ordinary politics cannot understand cannot be legitimated by executives assuring ordinary politics that **they** understand it on everyone else's behalf. The board crisis did not settle *who controlled artificial intelligence*. It exposed that the question had been malformed. The *thing* was already distributed across relations no board could contain. --- # **The Defendant Does Not Disappear.** There is an obvious objection to all of this *relation-speak*: > If artificial intelligence is not a bounded thing, perhaps nobody can be held responsible for it. - No thing, no owner. - No owner, no defendant. - No defendant, no regulation. The laboratories would be **delighted**. But relational governance does not mean prosecuting a relation instead of a person. It means assigning duties to every person and institution according to the power they exercise within the relation. The model developer remains answerable for how the model was trained, evaluated, represented, and released. The deployer remains answerable for placing it inside a particular institution. The employer remains answerable for using it against workers. The hospital remains answerable for the authority granted to it over patients. The infrastructure provider remains answerable for the resources consumed and the dependencies created. The executive remains answerable for decisions actually within the executive’s power. Governments remain answerable for the permissions they grant, the protections they omit, and the public systems they allow private actors to reorganize. - Nobody disappears. - More defendants appear. *Object-governance* asks *who controls artificial intelligence* and goes looking for one company, one board, one chief executive, one model, or one emergency switch. Each answer captures a real point of leverage and mistakes it for sovereignty. The result is an accountability bottleneck. Everyone outside the selected object becomes background, contractor, user, victim, or externality. Relational governance produces more enforceable handles, not fewer. It can regulate whether a system may be formed at all. It can condition the purposes for which a model may be deployed. It can assign disclosure, auditing, provenance, resource, labor, consent, appeal, and benefit-sharing duties to different participants. It can preserve liability across a chain instead of allowing every actor to point toward the next one. It can ask who possessed which capacity, exercised which authority, received which benefit, imposed which risk, and retained which opportunity to intervene. A relation is not an alibi. It is the map of responsibility that the object concealed. The company does not escape responsibility because it lacked total control. The executive does not acquire total responsibility because the public was taught to see a controller. Accountability follows actual authority through the system: neither dissolved into the field nor compressed onto the most recognizable face. The purpose of abandoning the *intelligence-object* is not to make responsibility metaphysically unavailable. It is to stop letting everyone hide behind it. --- # **The Thing Becomes Infrastructure.** For the first year of ChatGPT, artificial intelligence still looked almost *weightless*. A person opened a website, typed a sentence, and received language back. The physical system remained hidden behind the interface. No smoke came out of your browser. No cooling water ran across the screen. The answer appeared with the immaculate immateriality of intellect. But then scale made the body impossible to hide. Frontier systems required **immense** clusters of accelerators, power contracts, substations, transmission, backup generation, cooling systems, water planning, chip fabrication, land, construction, and long-term financing. OpenAI's own account of its evolution eventually described the organization as having become a global product company and an infrastructure company, far beyond its origin as a research laboratory. This was reality returning in concrete. Their allegedly *disembodied intelligence* began appearing to communities as data-center proposals. Residents did not encounter a new benchmark score. They met land-use changes, resource demands, tax negotiations, promised jobs, diesel generators, gas plants, and a low mechanical hum that did not stop when the chat window closed. Unfortunately, the environmental argument initially remained trapped inside the product ontology. Analysts asked *how much electricity an AI query consumed*, *how much water a model used*, or whether a more efficient chip reduced the footprint of one unit of inference. Those are all useful questions. They also preserve the assumption that the relevant system ends at the product boundary. The larger question turned out to be what economic activity the capability enables. A model that improves renewable forecasting may avoid emissions. A model that improves oil exploration may enable them. A coding assistant may reduce the cost of building climate software and reduce the cost of building fossil infrastructure. A general productivity amplifier has no climate direction until it enters a directed relation. By this stage, ***civilization itself*** was becoming the bottleneck. Better models could generate advice faster than institutions could absorb it. They could produce designs faster than permitting systems could evaluate them, code faster than organizations could audit it, synthetic media faster than provenance systems could authenticate it, and demand for compute faster than grids could connect it. The industry had spent years insisting that *intelligence* was civilization's scarce resource. Suddenly, it turned out the scarce resources were electricity, cooling, institutional competence, public trust, legal authority, social consent, and any distribution system capable of converting productivity into something other than private concentration. The model stopped being the bottleneck here **remarkably** quickly. So of course the industry responded by trying to acquire more of everything around it. --- # **“A Business That Scales With the Value of Intelligence.”** On January 18, 2026, OpenAI published an essay by its chief financial officer titled **“A Business That Scales With the Value of Intelligence.”** A hostile philosopher could not have improved this title. The essay is thoughtful, concrete, and **incredibly** revealing. It describes students using ChatGPT to untangle homework, parents planning budgets, writers breaking through blank pages, engineers reasoning through code, managers preparing difficult conversations, and people trying to understand their own lives. It explains the company's subscriptions, usage pricing, platform business, commerce plans, compute strategy, and revenue growth. It says that developers *embed intelligence* through APIs and that *useful intelligence* can be *delivered* at costs measured in cents per million tokens. It anticipates licensing, intellectual-property arrangements, and outcome-based pricing as *intelligence* moves into science, energy, finance, and other fields. - Every example describes a relation. - Every economic category sells an object. Take the student working on their homework. A student, a problem, a model, a prompt, a late hour, prior knowledge, an explanation, and a judgment about whether the explanation worked becomes *intelligence delivered by OpenAI*. Then the writer. A writer, grief, a blank page, a model, a draft, a voice, and the decision to keep or reject a sentence become *value created by our intelligence*. An engineer, a codebase, tools, organizational context, security constraints, and review become *intelligence embedded through an API*. The relations disappear into the invoice. This does not make the invoice fraudulent. OpenAI **does** provide valuable services. The models **do** contribute capacities that users do not already possess. Compute, engineering, reliability, and access all cost real money. A company may quite reasonably charge for them. The category error lies in naming the whole value-producing relation after the salable participant. That naming assigns ownership. The company owns the model and therefore appears to own the intelligence. The user supplies purpose, context, judgment, risk, and often the decisive knowledge of the field, but appears to receive *intelligence* from elsewhere. Institutions that supplied the training culture, public research, infrastructure, and social order become background conditions. Workers and creators become inputs or affected parties. The business becomes the source. The essay also reports a remarkable compute-revenue flywheel: OpenAI's available compute grew from roughly 0.2 gigawatts in 2023 to about 1.9 gigawatts in 2025, while annual recurring revenue rose from about $2 billion to more than $20 billion. More compute generated stronger models, stronger models generated adoption, adoption generated revenue, and revenue financed more compute. The product had found its metabolism. And the title finally said what the opening years had been building toward. *Intelligence* was no longer a human capacity extended by machines or a performance realized within a relation. It was a scalable economic substance. It could be placed directly in people's hands, embedded, metered, licensed, priced by outcome, and made foundational to the global economy. The relation had become a SKU. --- # **The House We Built Together.** On April 10, 2026, someone allegedly threw an incendiary device at Sam Altman's San Francisco residence. No one was injured. Federal and state authorities charged Daniel Moreno-Gama, then twenty years old, with offenses including attempted destruction of property by explosives and attempted murder. He has pleaded not guilty. Prosecutors allege that he traveled from Texas intending to kill Altman, then went to OpenAI's headquarters and threatened further violence. Authorities say they recovered a document expressing opposition to artificial intelligence, invoking humanity's impending extinction, and directing threats at artificial-intelligence executives. Those allegations remain allegations unless proved in court. The argument must slow here. Violence against an executive, employee, researcher, activist, public official, or anyone else is indefensible. No warning about artificial-intelligence risk, no critical movement, no paper, no product launch, and no public argument bears responsibility for an individual's alleged act. Criticism does not become more serious when it becomes violent. Violence destroys the field in which criticism, evidence, correction, and governance can occur. The incident belongs in this history for one narrow reason: the compressed map described in the allegations. - Artificial intelligence appears as a bounded thing. - That thing threatens humanity. - A company controls the thing. - A chief executive controls the company. - The distributed problem therefore appears to have an obvious human address. Each compression discards constitutive relations: - research communities, - chips, - cloud providers, - investors, - users, - data, - governments, - militaries, - markets, - energy systems, - competing laboratories, - public institutions, - and the millions of local decisions through which artificial intelligence acquires consequences. The final compression leaves one person standing in for the whole field. The wrong noun did not throw the device. A person allegedly did. But the wrong noun explains why a house could appear on the map as the location of *artificial intelligence*. By then the industry had spent years personifying the transition. - Companies had faces. - Races had champions. - Existential warnings named builders. - Corporate governance disputes were narrated as contests over the future of intelligence. - The public was repeatedly told that a handful of people were creating the thing that might reorganize or end civilization. The alleged attack was no logical conclusion of that discourse. It was a horrifying literalization of its most compressed image. - The relation had been turned into an object, - the object into property, - the property into a threat, - and the threat into a controller. At the end of that chain stood a house. --- # **The Thing Finds Oil.** History resumed. On June 22, 2026, Chevron announced a twenty-year agreement to provide dedicated electricity to a Microsoft-operated data center in West Texas. Project Kilby is expected to reach approximately 2.67 gigawatts. Most of its generation is planned to come from large natural-gas turbines. Chevron described the project as supporting the next phase of American artificial-intelligence growth by leveraging the country's natural-gas advantage. Microsoft and Chevron did not legally merge. They arranged a twenty-year metabolism, which is somehow much less reassuring. - One side supplies the demand curve. - The other supplies the fuel. The data center and power plant are co-located so the relationship can become physical infrastructure before the regional grid has time to object. - The oil company becomes an organ of the compute system. - The compute system becomes a guaranteed market for the oil company's gas. The 2019 Exxon partnership had returned at architectural scale. Then, in August, a study in *npj Climate Action* quantified the larger pattern. The researchers modeled *artificial intelligence* as a bidirectional productivity amplifier across fossil-fuel and renewable-energy pathways. Across sixty-four scenarios, equal or parallel productivity gains did not cancel each other out. Net annual carbon-dioxide emissions rose by an estimated 0.47 to 1.8 gigatonnes. Renewable-energy gains had to exceed fossil-fuel gains by roughly four to five times just to break even. In the modeled scenarios, net reductions required fossil-sector productivity gains to fall to **zero**. The study excluded direct data-center emissions from its equilibrium calculation, meaning the turbines and servers were not the whole result. **The main problem was what *artificial intelligence* helped the existing economy do.** This finding should end the fantasy that *artificial intelligence* carries a coherent climate direction. A productivity amplifier entering a fossil-embedded economy amplifies fossil incumbency. Oil and gas already possess fields, pipes, refineries, capital, political influence, customers, engineering knowledge, and an enormous installed system. Making extraction cheaper can turn marginal reserves into viable reserves, extend productive life, lower prices, stimulate demand, and reinforce the infrastructure that produced the gain. Renewables can improve too. But they have to improve several times faster because they are not beginning from the same field position. - The model supplied capability. - **The market relation selected fossil expansion.** Nobody had programmed this system to hate the climate. Nobody needed to. General productivity entered an economy whose incumbent structure still rewarded additional fossil throughput, and the result followed the structure. This is the article's argument made numerical. **Artificial intelligence is not a green object or a dirty object.** It is not an autonomous decarbonizer waiting to be pointed at the problem. The intelligence system changes what actors can do inside existing relations. Under current conditions, actors with fossil assets can use it to make fossil assets more productive. The productivity amplifier was impartial. The world it amplified was not. --- # **The Thing Attempts to Leave Earth.** The next move was to leave the planet. In February 2026, SpaceX acquired xAI, combining artificial intelligence, launch systems, satellite infrastructure, and social media inside a company valued by reports at approximately $1.25 trillion. The stated ambitions included orbital data centers: artificial-intelligence compute placed in space to escape limits on terrestrial power, land, cooling, and deployment speed. By August, the plan had grown into proposals involving vast satellite constellations. SpaceX announced a $100 billion Louisiana rocket complex intended in part to support artificial-intelligence satellite ambitions, while reporting described plans that could eventually involve up to one million satellites. The engineering remains uncertain. Space-based compute must survive radiation, launch costs, replacement cycles, communication constraints, and the awkward fact that heat-producing computers in a vacuum cannot simply hand their heat to passing air. They require radiators, mass, surface area, and an enormous launch cadence. The proposal may fail. But the symbolism has already succeeded. - We began with a model learning which word mattered to which other word. - Twelve years later, planet Earth had become a bottleneck to intelligence. The industry had originally described intelligence as the scarce resource. As model capability grew, the scarcity migrated outward: chips, electricity, cooling, transmission, land, water, permitting, community consent, and time. Each relation missing from the product ontology reappeared as a constraint on the product. The response was not to reconsider whether every available unit of compute needed to be built as quickly as capital could finance it. The response was a rocket program. Orbital compute is usually defended as a way to access abundant solar power and reduce terrestrial burden. Those benefits are possible. Launches, manufacturing, orbital debris, radiative cooling, and hardware replacement create other burdens. The final balance will depend on the architecture rather than the brochure. But even the optimistic case demonstrates the relational thesis. Intelligence cannot simply rise into space. Intelligence is an object. Rockets must launch it there. Factories must build it. Energy must power the factories. States must license the spectrum and orbits. Communities must host the launch complexes. Earth remains inside the system at every step of the attempted escape. Except there is no outside from which intelligence can become independent of relation. You can move the servers above the atmosphere. You cannot move them outside the extant field of relation. --- # **“Human Reserved.”** On August 26, 2026, Bill Gates published a long essay warning that governments were **radically** unprepared for artificial intelligence. Much of the essay is more serious than the standard industry performance. Gates treats employment, education, mental health, security, energy, taxation, inequality, and institutions as connected. He argues for new domestic and international governance, taxes on artificial-intelligence tokens and robots, stronger transition systems, and deliberate limits on automation. He plainly says market incentives will push firms toward replacement and that retraining cannot solve every displacement. Then he proposes a category called **Human Reserved**. The concept is modeled on nature reserves: places where roads and buildings could be placed, but society chooses not to develop them because the loss would be too great. Gates suggests that some jobs should remain human even when machines could technically perform them. Caregiving, education, mental-health work, and delivering devastating medical news are among the relations he wants protected. - The intention is compassionate. - The metaphor is terminally revealing. A **nature reserve** is territory withheld from the default operation of expansion. Development owns the surrounding map. The reserve marks the exceptional place where it must stop. Under **Human Reserved**, machine occupation becomes the default economic landscape. Human work survives inside designated habitats. There may be signs. - Please do not automate the caregivers. - Teachers may still roam here. - Human jurors protected by law. - Do not feed the construction workers; they are acclimating to a new sector. Gates is trying to save people from a market dynamic he correctly identifies. The wrong noun distortion has become so complete that the human appears as a special-use restriction on the machine's domain. - Eleven years earlier, OpenAI's founding announcement imagined artificial intelligence as an extension of individual human wills, broadly distributed. - Now a founder of the personal-computing era is drawing occupational conservation boundaries around the remaining human roles. This is what happens when *intelligence* becomes a product and *jobs* become territories. This is why ontology is so important. The question becomes which occupations the machine is allowed to take and which must be fenced off for people. The deeper questions disappear: - Who owns the productivity? - Who decides the purpose of work? - Why is survival still conditional on employment if automation produces abundance? - Which human relations should technology strengthen rather than replace? - What institutions would let workers govern deployment? - What would a transition look like if the **human** were the principal and the machine remained **an extension**? Human Reserved may become a useful emergency instrument. Some boundaries should be drawn. Some decisions and relations should remain under human authority even when automation is technically possible. Gates is right that capability does not settle permission. But the emergency instrument must not become the constitution. Human beings do not need **occupational wildlife preserves inside an economy belonging to artificial intelligence.** The economy is a human relation already. --- # **The Arrival of Artificial Intelligence.** So, when did *artificial intelligence* arrive? - Was it 2014, when attention let a translation system retrieve what mattered relative to the word it was producing? - Was it 2017, when the Transformer reorganized machine learning around attention? - Was it 2020, when general-purpose model access acquired a meter? - Was it November 2022, when the model acquired a dialogue box and millions of people discovered what they could do with it? - Was it February 2023, when the race officially started? - Was it May 2023, when the product was placed beside pandemics and nuclear war? - Was it January 2026, when a business model explicitly announced that it would scale with the value of intelligence? - Was it June, when a fossil-fuel company and a technology company arranged a twenty-year gas-powered metabolism for artificial-intelligence compute? - Was it August, when orbital data centers and Human Reserved jobs appeared in the same week's news? No date can mark the arrival of the object called *artificial intelligence*, because the object was always a compression imposed afterward. That object was never materially real. Models arrived. Attention mechanisms arrived. Transformers arrived. APIs arrived. Chat interfaces arrived. New capacities arrived. Data centers, contracts, companions, detectors, lawsuits, gas turbines, rockets, and employment proposals arrived. *Artificial intelligence* arrived in the relation between them. ***AI*** arrived between a model and a user trying to say something they could not previously say. Between a corporation and a worker whose task could now be accelerated or removed. Between a student and an assessment built for another technological world. Between an oil field and an optimization system. Between a data center and a grid. Between a government and a recommendation it could not fully inspect. Between a lonely person and a voice that always answered. Between a board and a mission too large for the corporation carrying it. Between a public fear and the executive made to stand for the whole transformation. The relation was real. The product was real. The capacities were real. The **distortion** consisted in pointing to one participant and naming it as the whole. Once we did ***that***, all the rest followed with brutal consistency. - A thing can be owned. - An owned thing can be sold. - A sold thing can be compared. - Comparable things can be raced. - A race can become inevitable. - An inevitable product can export its constitutive relations as externalities. - A threatening object can be assigned a controller. - A controller can be located at his house. - Meanwhile, the actual system spreads through every relation civilization failed to place inside the name. The lesson is not that models lack intelligence. It is that *intelligence* was never contained there alone. A model can contribute extraordinary *capacities* without becoming the sovereign source of the performance produced around it. Human judgment, institutional architecture, goals, tools, data, authority, energy, and consequence remain inside the *intelligence* phenomenon. That is also where governance has to begin. We do not need to decide whether the object is good or evil and then release or prohibit it everywhere. We need to govern **relations**: which systems may be formed, under whose authority, for whose benefit, with what resources, subject to what contestability, and with which parties bearing the gains and harms. The first arrival story made intelligence into a product and civilization into its deployment environment. A second arrival still remains possible. A proper one. The model can become an extension again. Workers can share in productivity. Students can learn inside redesigned assessments. Creative relations can preserve authorship without pretending thought has always been solitary. Provenance can become infrastructure. Machine advice can remain contestable. Communities can govern the physical systems they host. Climate policy can prevent a general amplifier from reinforcing fossil incumbency. Public institutions can decide where automation serves human purposes and where it contracts them. The technology will still be extraordinary. It will simply stop pretending to have arrived alone. We called intelligence a thing. The we measured the thing. We sold the thing. We raced for the thing. We feared the thing. We asked who controlled the thing. But there was never a *thing* there. There was a relation. And by the time we noticed, that relation had reorganized the field around us.