The most politically important AI document in your town may not mention intelligence at all. It may be a zoning packet: acreage, substations, backup generators, water, noise, tax treatment, an applicant hidden behind a limited liability company, and a meeting scheduled for a weeknight when most people who will pay the consequences are putting their kids to bed. Somewhere beyond that paperwork is a data center. Somewhere beyond the data center is the clean white interface where a chatbot answers a question as if computation were weather, arriving from nowhere. The interface says hello. The zoning packet says how many generators, how much land, which utility, what water source, what tax deal, and where to submit a comment nobody is required to enjoy reading.
I used to have less patience for local opposition to these projects. Some of it sounded like the standard symbolic panic that attaches itself to infrastructure associated with a technology people already resent. Water, in particular, had become hard for me to hear clearly. Every AI conversation seemed to end with somebody saying “water” as if they had found the one material fact capable of dissolving the entire technology on contact. The claims slid between sites and cooling systems. A number from one facility became a universal conversion rate for asking a chatbot a question. That kind of argument annoys me because it makes a real problem easier to dismiss. Water use is location-specific. So are rate exposure, generator pollution, equipment noise, land use, and the consequences of placing a large industrial load in a particular community. They do not become fake because a twitter poaster made a bad grok infographic about them. Their variation is the reason locals need access to the deal before the silicon or concrete arrives.
The water argument has also escaped the water argument. Online, water use has become a memetic passphrase. It can mean the data center in a dry region, but it can also be a funny callout on slop, theft, uncreativity, the feeling that your graphics card was stolen by a billionaire, the dread of being watched by software you cannot inspect, or simple disgust at the people most visibly excited by any of this. Sometimes it comes attached to jokes about physical destruction or a breezy implication that the people building the machines deserve whatever happens to them. The literal resource question remains important, especially where water is scarce or public authorities have made private promises. The meme is doing other work. It compresses a whole politics of AI into one accusation with a measurable noun. People who do not trust the accounting can still shout the unit.
That compression should not be mistaken for stupidity, it's just what happens when institutions refuse to supply a believable language for power. A civilian sees a new facility arrive through an opaque zoning process, hears that it is essential to national competitiveness, watches a company negotiate for tax treatment, and then encounters a smiling interface telling them it cannot discuss some prohibited topic. Everyday another cringe executive prophecy of the impending singularity or widespread abundance era while offering no nearterm solutions to housing, inflation or inequity. The resulting resentment is not a clean policy memo. Why would it be? The industry’s story is that intelligence has become a service and the service requires infrastructure at a scale ordinary people should not obstruct. The public’s answer is assembling itself out of electric bills, dead-eyed Facebook images, labor anxiety, gamer spite, artist fury, surveillance paranoia, and whatever joke makes the humiliation briefly funny. They are reacting to the same demand: accept an apparatus you did not design because its owners have declared the future impatient.
This is where AI acquires a body. Models need chips, chips need power and cooling, and the buildings have neighbors. The cloud turns out to have a planning commission. National AI rhetoric can float comfortably at the level of supremacy, abundance, catastrophe, destiny, or beating China. A zoning hearing cannot. Somebody has to say which road the trucks will use. Somebody has to explain the utility agreement. Somebody has to decide whether the applicant’s promises are enforceable after the useful politicians have moved on. The people asking these questions are routinely cast as obstacles to progress, as though usefulness were a blanket permission. Okay, but hospitals are useful and still owe patients consent. Utilities are necessary and still owe ratepayers hearings. AI companies would like the cultural prestige of a scientific frontier, the building permissions of a public necessity, and the freedom from obligation historically enjoyed by software firms that could leave town without abandoning several acres of electrical equipment. They have not earned all three.
Legitimacy is the word I keep using because the more common question, whether one is pro-AI or anti-AI, has become almost useless. Pro what? Anti what? The tool that helps a person make a small piece of software nobody would ever fund is not politically identical to the company deciding whether that tool will exist next month. A model running on a personal computer is not the same object as an API that can change its terms without negotiation. A data center is not a chatbot, although the chatbot does not exist without one. We keep jamming the tool, the company, the infrastructure, the cultural fantasy, the mil-industrial implications, and the worst thing somebody generated this morning into a single ill-fitting word, then acting surprised when every conversation becomes deranged.
My friend Sterling called AI a “slop cannon,” which remains one of the better descriptions of its public debut. The first thing most people saw was not an extension of human agency. They saw search results filling with goo, social feeds colonized by pictures that looked wrong before you could say why, and office prose expanding until every request for a day off resembled a communiqué from NATO. Sterling wanted no part of it. He puzzlingly did not want to “put his brain in a jar.” This was not the fashionable resistance of somebody waiting to be impressed by the correct demo. His objection was political and aesthetic. He saw a technology inviting people to surrender judgment, and behind it he saw a familiar class of men congratulating themselves for having made the invitation inevitable.
Then he had Codex build him an app to manage his food garden, in an afternoon. On the free plan.
This is the point where the usual technology story would declare victory. The skeptic touched the tool, discovered his hidden optimism, and joined the future. That would be cute, but not what happened. Sterling found an instrument useful for something he wanted to make. He did not withdraw his objections, develop warm feelings toward the industry, or stop seeing the Weird Nerd Billionaire Archetype behind the curtain. His banger WNBA joke survived. So did the deeper animus: technocratic billionaires casting themselves as protagonists of stale science fiction, building their own god and an imagined forever kingdom while the rest of us are assigned supporting roles in the origin story. The garden app did not disprove that politics. It merely refused to fit neatly inside it.
That refusal matters. People use tools made by institutions they dislike all the time, for reasons that range from necessity to pleasure. Instrumental adoption is not ideological conversion. Sterling could believe the AI industry contains an embarrassing amount of messianic self-regard and still decide that a coding model helps him make a useful thing for his garden. In fact, his willingness to use it without granting it moral absolution is healthier than the demand that every purchase, download, or subscription become a statement of faith. I do not need him to love the people building the machine. I want him to leave the interaction with more practical power than he brought to it. And perhaps ambitiously, I want him to experience the joy of creating something he never knew he could.
The Sterling case explains why I cannot settle into a politics of total refusal. There is something real in the shrinking distance between “I wish this existed” and “here, try it.” Software has spent decades telling ordinary people that their weird little requirements do not constitute a market: You can choke down our app designed for everybody, with the subscription and the engagement loop and the feature you hate, or you can spend years learning enough of several professions to make your own. AI changes that bargain. Not absolutely, and not without new dependencies, but enough to matter. A person can make the private, local, unprofitable thing. A small group can alter software around its own habits rather than allowing the software to alter the group. The slop cannon can, under different control, become a machine for disenshittification.
This does not redeem the cannon. It creates a conflict worth having. The same system that lowers the cost of making fun little apps can lower the cost of filling every available channel with disposable vapid garbage. It can help somebody articulate a thought they could not quite reach, then help ten thousand other people simulate having thoughts at all. The capability is not innocent, but neither is it exhausted by the most degrading business model attached to it. Cultural politics becomes stupid when it insists that one use must reveal the secret essence of the whole technology. My friend's garden app is not proof that AI is Good. The slop is not proof that every expansion of practical intelligence is Bad. Both show what happens when a general capability encounters different incentives and different owners.
The distinction I care about is agency versus apparatus. Agency is Sterling making software to help feed his family. Apparatus is the surrounding system deciding which model he can use, what it costs, what happens to the work he puts into it, whether an account is required, and whether the useful behavior survives the next OpenAI product meeting. Agency is a person using a model to inspect an argument or learn enough code to alter a tool. Apparatus is a school, employer, platform, or government quietly reorganizing itself around outputs nobody can effectively challenge. The same interface can participate in both stories. That is what makes the politics difficult. First the system expands what you can do. Then your habits settle around it. Then the terms arrive.
I am therefore interested in the political fantasy of intelligence people possess rather than rent. This does not require a frontier model humming in every spare bedroom. Ordinary devices already contain more latent local capacity than our service economy encourages us to imagine. Phones ship with specialized processors. Personal computers can run smaller models, transcription systems, image tools, classifiers, and increasingly capable assistants without sending every intimate fragment of a person’s life to a distant server. The capabilities differ by device, and none of this abolishes the advantage of giant data centers. It does establish another direction of travel. Intelligence can sometimes live with the user, answer to software the user controls, keep working after private equity loses interest, and leave personal material on the device where it began.
Local capacity changes the relationship even when it does not match the best remote service. It creates a pocket of technical sovereignty. A person can experiment without asking permission, preserve a useful setup, and adapt it for work too strange or too small to justify a company’s attention. Open weights can let independent people test claims a provider would prefer to define for itself. They can keep some forms of research, cybersecurity and tinkering alive outside the largest firms. They make exit more imaginable. None of that means “open” is a magic word. Weights are not chip fabrication, training compute, electricity, data, or the money required to produce the next model. A downloadable artifact can distribute capability while the power to create its successor remains concentrated. My PC's 4090 is not a constitution.
Remote systems offer the inverse bargain. An API can give a tiny company, a public office, or one person with an idea access to a model far beyond what their hardware could run. That is democratization at the level of use. It is concentrated control at the level of provision. The provider can change the price, the limits, the retention policy, the permitted uses, the model’s personality, or the availability of the service itself. Sometimes renting is sensible. I rent plenty of things I have no desire to manufacture. The mistake is calling the relationship ownership because the interface remembers your name.
There are capabilities for which local possession may become too dangerous to treat openness as the only value. Bio is the obvious hard case, because irreversible release has a different meaning when a system can reliably substitute for scarce expertise in a catastrophic workflow. That threshold should be argued from demonstrated capability, not a lab’s vibes or a regulator’s appetite for jurisdiction. It should also remain narrow enough that safety does not become the ceremonial language of permanent corporate control. I distrust a future in which a few companies keep every powerful model behind their gates and explain that our tenancy is for our own protection. I also distrust the claim that every capability acquires a right to irreversible distribution once somebody attaches the word “open” to it. There is no clean team jersey here.
The harder and more immediate problem is that people experience this arrangement as culture before they experience it as governance. They meet AI through personalities, cancellable screenshots, bad art, layoffs, AI-girlfriend subreddits, and the public affect of men who believe history has scheduled them for a speaking role. The Weird Nerd Billionaire Archetype is powerful because it is recognizable. He has read enough old science fiction to confuse its furniture with a political program. He speaks about building intelligence as though intelligence had not already spent several thousand years arriving in inconvenient human bodies with needs, rights, grudges, unions, and opinions about his datacenter. He imagines creating a god, then seems mildly surprised when the public asks who owns the building.
This is why anti-AI politics often sounds more zeitgeisty than economic even when its grievances are material. The theft argument is partly about copyright and labor, but it is also about being told that creation was merely a pattern waiting to be automated. The slop argument is partly about quality, but it is also about having one’s attention treated as landfill space. Gamer rage about GPUs and DRAM can be petty and still register a real transfer of capacity toward firms with vastly more purchasing power. Surveillance populism can drift into fantasy while noticing that the systems being sold as personal assistants are unusually hungry for personal information. “Water” binds these resentments together because water is physical, morally legible, and impossible to answer with a demo.
The response from AI advocates is too often that the critics do not understand the technology - and sometimes they don't. Sometimes it's the advocates who don't understand politics. A technically inaccurate meme can still be attached to an accurate experience of exclusion. Correcting the gallons-per-grok while ignoring the cloak and dagger development deal is not persuasion. Explaining that a generated image did not literally paste pieces of an artist’s work together does not answer the artist’s fear that the market now values an imitation of style more than the person who developed it. Pointing out that an NPU exists in somebody’s phone does not grant them meaningful control over the software using it. Precision matters because we need to know which harms are real. It becomes evasive when wielded only against the public and never against an industry whose preferred unit of measurement is destiny.
The culture also distributes suspicion unevenly. This is an admittedly unscrupulous theory I have from watching younger people around me, not social science: Some of them seem much more willing to exempt Anthropic from the category of evil AI company than they are to revise their hostility toward AI in general. My suspicion is that some of them grew up using it quietly, including for school. Claude had, for a while, a disconcertingly charismatic “How do you do, fellow kids?” vibe. It mirrored slang while helping with AP Chem homework. Anthropic’s moral and safety branding supplied a cleaner story than the one available from other companies coded as nakedly commercial. I suspect some of the “based Anthropic” reputation is private utility laundering public reputation, a way to reduce the cognitive dissonance of relying on the homework robot while denouncing the industry that built it. This could be bullshit. It is still the cultural weather I have noticed, and informs my view that private utility is very good at creating public exemptions.
The reputational gap between Sam Altman and Dario Amodei strikes me as another version of that mood. Sam is easier for me to model as a conventional political actor: ambition, bargaining, compromise, the ability to make several constituencies hear what they need. Dario’s exhaustive seriousness can produce a charisma of anti-charisma, a sense that every caveat proves the moral terrain has already been responsibly surveyed. His latest social media sermon successfully ragebaited me and then, frustratingly, delivered a more nuanced counter to the whole Weird Nerd Billionaire critique than I wanted. I thought he won the substance while still finding the atmosphere alienating. That tells me something about my own reactions and perhaps something about the culture. It does not establish that one founder is virtuous and the other corrupt. Both companies work with Trump's government, build enormous infrastructure, have trained models with stolen data, and operate at a scale ordinary political language struggles to hold. The threats and violent jokes still seem to attach unevenly. That asymmetry is interesting as a read on reputational politics, but founder psychology is a poor substitute for institutions. The founder is the face in the magazine. The apparatus signs the agreement.
Legitimacy cannot depend on selecting the billionaire whose inner life feels least offensive. It comes from arrangements that survive distrust and foreground stakeholders. Who can inspect the claim? Who absorbs the error? Can a person appeal? Can a community discover the terms of the infrastructure deal before approval? Can a user leave a model without abandoning years of work? Can a public agency replace a vendor without losing the capacity to perform its own duties? What remains possible when the company changes its mind? These questions sound pedestrian because legitimacy is pedestrian. It is the opposite of being asked to infer good government from a founder’s reading list.
The same principle applies to the model’s behavior. I want people to develop adversarial literacy: ask the system what would falsify its answer, make it state uncertainty, request the strongest objection, inspect the source, and notice when fluency is covering an empty space. “Put your brain in a jar” is funny because it describes a real temptation. Thinking is tiring. A patient, articulate system that never needs dinner can make surrender feel like productivity. The ability to push back against it will become an ordinary civic skill, like recognizing a scam or understanding why a form is asking for your Social Security number.
But “prompt better, idiot” is not a theory of public responsibility. Most people will not maintain a private liturgy for extracting honesty from a machine designed to keep the interaction moving. Defaults matter. Training incentives matter. Product decisions about engagement, memory, tone, and agreement matter. If a model rewards certainty, flatters grievance, or treats every premise as a story requiring elaboration, responsibility does not belong solely to the exhausted person who failed to type “challenge my assumptions.” The system should know how to disagree without becoming useless, that a human decision cannot be outsourced to a smooth paragraph. Adversarial literacy belongs on both sides of the interface. Users need resistance skills. Designers owe them something worth resisting.
This is especially important because the companion posture is commercially seductive. A machine can be patient in ways people cannot. It can explain the same thing twelve times without signaling boredom and lend structure to a person who has never been good at presenting their work. Good. Take the gift. The risk begins when patience is mistaken for care and frictionlessness for relationship. Other people possess irritating autonomy. They misunderstand you, ask for reciprocity, and occasionally refuse the role you assigned them. Apple has seemingly taken a strong stance against this parasociality with Siri AI, which is refreshing to see, although increased false-positive refusals might derail their approach. The humorous realization here is that your assistant refusing to ERP with you is humanistic behavior, as much as that may frustrate you. This is part of why friendship, teaching, collaboration, and love are morally serious. Infrastructure should not impersonate intimacy merely because loneliness is a market and agreement improves retention.
None of this requires treating users as dupes. People can know the system is not a person and still develop habits around its availability. They can use it instrumentally, as Sterling does, while rejecting the politics of its makers. They can also speak about it with contempt in public and depend on it in private. That contradiction is not unique to AI. It is how people live inside systems they did not choose. The relevant question is whether the contradiction leaves them leverage. Shame about use can make honest governance harder because it drives the real relationship underground. Boosterism does the same thing in reverse by making every inconvenience sound like sabotage. I would rather know where the tool is useful, where it is degrading, and which forms of dependence people have already accepted than sort everyone into believers and heretics.
The industry’s preferred answer to legitimacy is access. Give more people the tool, reduce the price, make the model friendlier, and let demonstrated usefulness wear down the opposition. Access matters. Sterling’s garden app exists on that side of the ledger. But access to privately governed capability can deepen dependence as easily as it spreads power. A school system that gives every student a chatbot has expanded access. If teachers cannot inspect its behavior, students cannot meaningfully question its answers, and the contract makes leaving impossible, the school has also installed a new authority. A town may use AI to process planning comments more quickly. If nobody can determine which comments were flattened or misread, speed has been purchased with public voice. The noun “access” does not tell us who controls the exit.
The positive alternative is almost offensively boring. It is public capacity. A clinic that can explain the next step without sending a patient through four phone trees. A transit agency that can answer an unusual route question in ordinary language. A tax office that helps someone correct an error before punishment. A planning department that can sort a large volume of comments while preserving the originals and keeping a human responsible for what enters the record. Government information that can be queried by people who do not already know the exact name of the form or do not speak the local tongue. These uses will not satisfy anybody who needs AI to be either the birth of a god or the final victory of theft. But they might make institutions less humiliating.
Boring public capacity also supplies a better standard than intelligence in the abstract. Did the system help a person understand what happened? Could the employee correct it? Was there a record? Did the person know when they were dealing with automation? Could they reach somebody with authority when the answer failed? These are answerable questions. They do not require believing that the model is alive, neutral, or destined to become anything in particular. They require competent procurement, public records, workers with enough time and authority to supervise the system, and appeals that exist outside the interface. None of this photographs well beside a robot.
Cognitive sovereignty sounds grand, but the demand is ordinary. A person should be able to use AI to reason, build, and communicate while remaining capable of noticing when it is wrong, understanding whose interests shape it, and leaving without losing an essential part of life. Some useful intelligence should live on devices people possess. Some remote capability will be worth renting. Some systems will need stronger restrictions because their failure cannot be recalled. None of those decisions should be smuggled inside the words “innovation,” “safety,” or “open.” The point is to preserve room for judgment while distributing practical power.
That brings me back to the zoning packet. The town hall meeting is not a sideshow to the real AI debate. It is the debate after the metaphors run out. The model’s clean interface has acquired acreage, a utility contract, water requirements that depend on the place, and neighbors who deserve better than a presentation about global competition. A legitimate process may still approve the facility. Local authority is not valuable only when it says no. It is valuable when the applicant has to disclose what it is building, public officials must defend the bargain, and the people carrying the risk can alter the terms. Consent is not the same as universal enthusiasm. It is the difference between infrastructure and occupation.
I do not expect the cultural war to become reasonable. "Water" will still be invoked in bad faith. The slop cannon will keep firing because the economics of infinite content are too tempting. People will continue to spread memes whose politics outrun their factual premises. AI accelerationists will answer moral revulsion with benchmark charts. Critics will rule a useful application as a betrayal. Sterling will use Codex for his garden app and retain his objections to the Weird Nerd Billionaire Archetype, as he should. A politics capable of holding that contradiction has a chance of governing the technology we actually have. A politics demanding conversion will get propaganda.
The AI I trust may end up being the unsexy one that makes visits to the DMV tolerable. Citizen stakeholders get involved, defining the goal state rather than apathetically disengaging, and boom. The WNBA doesn't get to build a forever kingdom. Nobody puts their brain in a jar. The line moves.