Ghost In The AI

The most expensive Ghost in the Shell ever made is the one everybody agrees doesn't have a ghost in it.

The 2017 live-action film cost north of a hundred million dollars. It recreated Mamoru Oshii's 1995 imagery with the reverence of a museum forger: the thermoptic dive off the tower, the fight in the flooded alley, the rain, the neon, the eyes. Every shot was gorgeous. Every shot was empty. Critics reached, almost unanimously, for the same joke, because the film had helpfully pre-installed it in the title. A beautiful shell. No ghost.

And every frame of it was made by human beings.

That's the part worth sitting with before we get into any of this. Hundreds of skilled people, working with enormous budgets and real craft, assembled a technically flawless object that felt like it had been produced by something that had read about feelings in a Wikipedia article. Nobody blamed a machine. There was no machine to blame. There was a committee.

Here's the thing. We are currently having an enormous cultural argument about whether art made with artificial intelligence is really art, and almost all of it is being conducted as though the question arrived on a Tuesday afternoon in 2022. It didn't. It's the oldest question in aesthetics wearing fresh chrome, and we have been failing to answer it for about a hundred and eighty years, ever since a French poet looked at a photograph and announced that the camera was where failed painters went to hide.

The question isn't whether a machine can make art. The question is how you can tell, from the outside, whether anybody meant anything.

And the honest answer, the one nobody enjoys, is that we never actually had a way to tell. We had a proxy. The proxy just broke.

The Ship Was Never the Planks

Ghost in the Shell has always been a Theseus story. That's the premise, sitting right there in the first act, waving at us.

Motoko Kusanagi has an artificial body. Not augmented. Replaced. Her eyes are manufactured, her muscles are manufactured, her brain is a cyberbrain, which means even the wetware has been penetrated, wired, backed up, and made subject to a maintenance schedule. Section 9 owns the shell. She goes to the shop for repairs. In the 1995 film she notes that if she quit, she'd have to hand back the body and everything in it, which remains the least reassuring employment arrangement in cinema.

So the franchise keeps circling Theseus's question with the serial numbers filed off. Replace the arm. Still Motoko. Replace the legs. Still Motoko. Replace essentially the entire body. Still Motoko. Alter the memories, and now things get uncomfortable. Copy the consciousness, and things get worse. Merge a human mind with an artificial one, and the taxonomy department has gone home for the day.

The version that actually keeps her up at night is not am I still me. It's how would I know. A sufficiently good forgery of her memories would feel precisely like her memories. The 1995 film gives us the garbage man, a man whose entire life has been overwritten by a hacker, who is nonetheless walking around cheerfully convinced he has a wife and daughter waiting at home. He is not distressed; that's the horror.

Now run the same experiment on a piece of writing.

A novelist writes a sentence. Spell check fixes a word. Still hers, obviously. The grammar checker removes a comma. Fine. A friend says the protagonist's motivation makes no sense, and she rewrites three chapters. Still hers. A copy editor smooths an awkward clause and quietly deletes the joke she liked. Still hers. Her developmental editor says the second act drags and the sister should be a brother, and she does it. Still hers. The publisher picks a better title. Still hers. She dictates a chapter instead of typing it. Still hers.

Now she opens a chat window and types: this transition isn't working, give me five approaches. She reads all five, dislikes four, and rewrites the fifth into something that works.

Uh-oh.

So hold that step against the ones before it. The friend detonated the protagonist's motivation. The developmental editor changed a character's gender and cost her three chapters. The copy editor rewrote her sentences without asking and killed a joke she loved. The machine handed her a transition, which she then rewrote.

By any measure of how much of the book actually moved, the machine is the smallest intervention on the list. It isn't close.

So the line people feel isn't about scale. It's about source. Every earlier step ran through a person and this one didn't, and that turns out to be the entire difference. It's a judgment about provenance wearing the costume of a judgment about craft.

Provenance can matter, and I'll argue later that it does. But it has to be named, because a named claim has to be defended on its own terms. It can no longer quietly borrow authority from the scale argument standing next to it, which it was never actually making.

That's the plank everybody points at. Not a plank in the ship. The decisive one, the single board whose replacement ends Theseus' ship.

Nobody can locate it. Everybody is certain it's there.

That doesn't prove it isn't. It does suggest we've been hunting for it with the wrong instrument.

And it isn't as though the question has never come up before. We have answered versions of it for centuries, in public, at scale, and then quietly forgotten we did. Ghostwriters write books with other people's names on the cover, and Profiles in Courage won a Pulitzer for John F. Kennedy with a great deal of Ted Sorensen in it. V. C. Andrews has now published more novels dead than alive. Comics divide a single page among a writer, a penciller, an inker, a colorist, a letterer, and an editor, and we credit all of them without deciding which one made the comic. Rubens ran a workshop with a published price list, where the cost scaled with how much of the canvas Rubens personally touched, which means one of the great masters of Western art was selling authenticity by the square inch, itemized, like a plumber.

Thomas Kinkade ran the identical scheme in reverse: assembly-line reproductions, priced by tier, with a "master highlighter" adding dabs to make each one unique and the top tier touched by Kinkade himself. Four hundred years and two continents apart, two artists arrived at the same business model. Pay more, the famous hand touches the canvas.

And in 1967 Sol LeWitt removed the hand entirely. His wall drawings are instructions. Someone else executes them, in another city, decades later, after LeWitt's death, and it's still a LeWitt. He wrote that the idea becomes a machine that makes the art, which was intended as conceptual provocation and now reads like a product announcement.

The ship was never the planks. It was the shape somebody insisted on.

Thirty-Five Cents of Banana

In December 2019, at Art Basel Miami Beach, Maurizio Cattelan taped a banana to a wall and called it Comedian. Two editions sold for a hundred and twenty thousand dollars each. A performance artist ate one off the wall, which the gallery treated as a scheduling inconvenience rather than the destruction of property, because it wasn't the destruction of property. In November 2024 an edition sold at Sotheby's for $6.2 million to a crypto entrepreneur, who ate it at a press conference in Hong Kong. The banana at that auction had been bought that morning from a fruit cart outside the building for thirty-five cents.

Everybody's favorite fact about Comedian is the price. The interesting fact is the maintenance schedule.

The banana rots. Naturally it rots. It's a banana. So it gets replaced, every few days, forever. What the buyer actually owns is a certificate of authenticity and a set of installation instructions specifying the height, the angle, the tape. The work survives the total replacement of its physical substance, indefinitely, because the physical substance was never the work.

Comedian is a Ship of Theseus with a fruit-stand budget.

Which makes Maurizio Cattelan and Motoko Kusanagi the same argument. One is a cyborg counterterrorism officer interrogating consciousness in a posthuman surveillance state. The other is a banana. They make an identical claim: the thing is not the matter, the thing is the pattern, and the pattern is what persists once you've swapped out everything else.

Cattelan's contribution was selection, framing, context, and nerve. He didn't grow the banana. He pointed at it. Which is inconvenient if your definition of art depends on how much manual labor the artist performed.

It's also fatal to the definition most of us reach for when pressed, which runs roughly: art is whatever provokes emotion.

That one gets deployed constantly to end the AI argument in a single move. If an AI story makes you cry and a human story makes you cry, the tears are the same tears. So what exactly is the complaint?

It's a fair question resting on a broken definition, which makes it worth taking apart slowly instead of swatting away.

It fails because emotion is cheap and always has been. A scam text saying your grandson is in jail and needs bail money produces enormous emotion. It is not art. It's a scam text saying your grandson is in jail. A Sarah McLachlan shelter-dog commercial produces emotion so reliably that its efficiency has become a punchline. Advertising is a trillion-dollar industry devoted to the mass manufacture of feeling, staffed by genuinely talented people, and almost none of it is art, not for lack of craft but because the feeling is instrumental. It's there to move a unit.

Emotion is evidence that something worked. It is not evidence that anybody meant it.

Film editors have been laughing at us about this for a century, because they know exactly how the trick is done. Kuleshov demonstrated it in the 1920s: cut the same neutral face against a bowl of soup, a coffin, and a child, and audiences will swear the actor conveyed hunger, then grief, then tenderness. The face never changed. The emotion was manufactured by juxtaposition, by an editor, out of footage containing none of it. The audience does most of the work of art and always has. We are pattern-completing animals who cry at chord progressions and see faces in wall sockets.

So emotion can't be the definition. But notice what the definition was reaching for, albeit poorly. When people say art is what provokes emotion, they don't mean any feeling from any source. They mean the specific, slightly embarrassing sensation of contact. The sense that someone else was in here. That a stranger, possibly dead, possibly a Japanese mangaka, possibly a man who taped fruit to drywall, arranged something on purpose and aimed it at you.

That sensation is the actual object of this entire argument. Everything else is proxy warfare.

The Monet Test

In May 2026 the internet ran the experiment for us, by accident, in front of six point seven million people.

A conceptual artist working under the handle SHL0MS posted a cropped image of a late Monet Water Lilies, painted around 1915 and currently hanging in Munich. He captioned it as an AI generation in the style of Monet and invited the audience to explain, in as much detail as possible, what made it inferior to the real thing. He attached the platform's own "Made with AI" label for good measure.

The internet rose magnificently to the occasion.

The composition was incoherent. The greens were an inconsistently saturated muddle. There was no depth, no contrast, no plane. It was cluttered slop that achieved maybe twenty percent of a Monet. One person produced more than seven hundred words of forensic analysis explaining what the machine had failed to understand about light.

Except that Monet had painted it.

Nothing about the image had changed. Not one brushstroke, not one lily pad, not one square millimeter of impasto. The only variable was the label, and the label was a lie, and the lie was sufficient. People were not looking at the painting. They were looking at a caption and then hunting the canvas for corroboration, which they found, because you can always find corroboration.

Then a lot of them deleted their comments. Screenshots survive, because of course they do.

Two things about this are worth pulling out, and the second one matters more than the first.

The first is that it's not an isolated embarrassment. It's a replication. A 2024 study in Nature by Grassini and Koivisto found that people couldn't reliably distinguish human from AI images, mildly preferred the AI ones when the source was hidden, and then marked those same works down once they were told what made them. And a 2004 study by Kruger and colleagues had already established the underlying mechanism, which they called the effort heuristic: we assign more value to art we believe took longer to make. Not art that is better. Art we believe cost more.

The second thing is that some people got it right in real time, and they got it right for reasons.

A working oil painter pushed back on the posted Monet immediately, pointing out that the lily pads clearly established a plane and the willow reflection created inverted space, and that the paint texture read as a credible physical object, thinner than most Monets but plausible. An art historian was blunter, identifying it as a late Monet from the brushwork, which she recognized as kin to the Agapanthus in New York, because the late paintings carry that particular wild impasto.

That matters enormously, and it's the part that gets dropped when this story circulates as a dunk. Trained eyes saw real things. The expertise was not fake, it was not decorative, and it did not fail. What failed was the confidence of people substituting a provenance judgment for an aesthetic one and not noticing they had made the swap.

They weren't looking at the shell and discovering it was empty. They had been told it was empty and went looking for proof.

The Refuge of Failed Painters

Charles Baudelaire, reviewing the Salon of 1859, described photography as the sanctuary of every would-be painter too incompetent to finish his training. It's a good line. He was wrong in the way people who write good lines are usually wrong: precisely, memorably, and completely.

The photography panic is worth walking through, because it contains every argument currently being made about AI, in the same order, with the same confidence.

It has no soul, a machine made it. It requires no skill, the machine does the work. It isn't creation, it's capture. It will destroy portraiture as a profession. It will flood the world with images until images mean nothing. It's a tool for people who can't draw.

Some of those were wrong. Some were extremely right. Portraiture as a mass profession did collapse. Thousands of working painters, competent people with real training who made an honest living painting likenesses for the merchant class, were wiped out inside a generation. That happened, it was not imaginary, and the people it happened to were not being precious about it.

The legal system had to settle the question, and it settled it the same way it's settling this one. In 1884 the Supreme Court took up Burrow-Giles v. Sarony, where the defense argued a photograph couldn't be copyrighted because a camera is a mechanical device and therefore no human authored the image. The Court disagreed, and the reasoning is the whole ballgame: the photographer had posed Oscar Wilde, chosen the costume, arranged the drapery, selected the light, framed the shot. Authorship was in the decisions. The machine executed, the human decided.

Then it happened again, roughly every twenty years, and each time everyone behaved as though it had never happened before.

The American Federation of Musicians ran an advertising campaign against what they called canned music and struck the recording studios for two years to stop it. The UK Musicians' Union spent the early eighties trying to restrict synthesizers on the grounds that they replaced real players, which they did. Sampling wasn't real composition, and also a lawsuit. Digital photography wasn't real photography, no chemistry, no darkroom, no waiting. Photoshop got a decade of solemn essays about the death of photographic truth. Digital painting wasn't real painting, because Ctrl+Z. Auto-Tune wasn't real singing. CGI wasn't real filmmaking.

A modern photographer doesn't personally supervise every photon. The camera meters the exposure, autofocus picks the plane, stabilization compensates for a shaking hand, software corrects the lens distortion, Lightroom handles the noise, and Photoshop removes the trash can ruining an otherwise lovely sunset. Nobody asks about any of it. We ask what they photographed, and why, and whether it's good.

The technology disappears into the process. Until the technology is new. Then suddenly everyone can see it again.

Which brings me to spell check, the most instructive case and the least discussed.

Spell check was cheating. Not metaphorically. Teachers argued about it, papers were written, there was a real pedagogical panic that a generation would lose the ability to spell. Then it was standard. Then it was invisible. Nobody now describes a novelist as having had machine assistance because their word processor caught recieve. The tool didn't get weaker. It got assimilated.

That's the full life cycle of every creative tool that has ever frightened anybody, and it runs fifteen to twenty-five years: cheating, controversial, standard, invisible, infrastructure. The only variable is how loudly the middle stage goes.

None of which proves AI is fine. It proves that "this feels like cheating" is not, by itself, information about art.

Look at what the reflex actually tracks. Every time it has fired, it has been wrong about whether the new thing could produce art, and close to right about whether people were going to lose their livelihoods. Baudelaire was wrong about photography and the miniature painters were not wrong about their rent. The musicians' union lost the argument about whether recordings were music and won the argument about what recordings would do to working players.

So the reflex is a reliable instrument pointed at the wrong dial. It detects economic dislocation and reports the result as an aesthetic verdict. It's the same substitution the novelist ran into back at the start, where a judgment about where a suggestion came from was doing business as a judgment about craft, and the same one the Monet commenters made in public in front of six point seven million people.

Hold onto that, because it comes back.

What Actually Broke

Here's what I think most of the discourse is missing, and it explains why this round feels different even to people who know the history and are panicking anyway.

We have never been able to detect intention. There's no instrument. You cannot look at a painting and directly perceive what exactly somebody meant. What we have always done instead is infer intention from difficulty.

Effort was the evidence. If a thing took ten thousand hours, somebody wanted it to exist badly enough to spend ten thousand hours, and that wanting is the ghost we were actually shopping for. The virtuosity, the brushwork, the impossible guitar run, the sentence that could only have come from one particular damaged mind: those are not the art. They're the receipt. They're how we verified that a person was in the room and cared.

Kruger's effort heuristic is that receipt, measured in a lab and confirmed. Generative AI has not disproved the soul. It has decoupled output quality from effort expenditure, which means our receipt no longer proves purchase.

The crisis is forensic, not metaphysical.

This reframing earns its keep because it explains behavior that otherwise looks irrational.

It explains why people get angrier at a good AI image than a bad one. A bad one is just a bad image. A good one is a forged receipt, and forged receipts feel like a personal insult, because you were using that system to decide who to care about.

It explains the Monet comments, and it explains the deletions. Those people were not lying. They had an instrument for detecting effort, they were told the effort was zero, and the instrument dutifully reported back what it had been told.

And it explains why the percentage of AI use question is unanswerable. Suppose a story makes you cry and then somebody tells you the writer used AI. What happens to the crying? Does it retroactively counterfeit? How much invalidates it? Five percent? Twenty? If a model suggested the plot but a human wrote every sentence? If the human built the plot and the model produced the first draft? If a model wrote a paragraph and the human rewrote half of it? At forty-nine percent, are you still permitted to be devastated by chapter twelve?

That's Ship of Theseus accounting applied to feelings, and feelings are stubbornly uninterested in spreadsheets.

But notice what the impulse to audit is actually tracking. Nobody is really counting words. They're estimating how much intention was probably present, and their only gauge is labor.

Which also explains why disclosure changes the experience so violently. If you love a poem and then learn a model wrote it, and the love curdles, that curdling is not irrational and it isn't hypocrisy. It's information. It tells you that what you valued was never only the artifact. It was the relationship the artifact implied. You thought someone was in there.

The law has already found the same seam and cut along it. The Copyright Office refused registration for the Midjourney image that won the Colorado State Fair digital category in 2022. It partially registered Zarya of the Dawn, protecting the human's text and arrangement while refusing the generated images. In Thaler v. Perlmutter the courts held flatly that human authorship is required and a machine cannot be an author. Every one of those rulings is Burrow-Giles again, running the same test: where were the decisions, and who made them?

The answer isn't zero. It also isn't automatically enough. That's the live question, and it deserves better than either of the two available slogans.

I'm going to defer it, deliberately, because answering it well requires two things this essay hasn't earned yet: a clear look at what the tools actually produce when nobody is steering, and an honest accounting of the objections that have nothing to do with aesthetics at all. Both are coming. Hold the question.

The Curb Cut

There's a second thing the effort heuristic does, and it's worse than the first.

It doesn't only treat difficulty as evidence. It treats difficulty as virtue. Once you've spent a lifetime reading labor as proof that somebody cared, it's a short walk to reading ease as proof that they didn't, and a shorter one to a moral system where the hard way is the honest way.

Watch how fast the vocabulary turns moral. We don't say the easier method produced a weaker result. We say cheating. Shortcut. Lazy. Not real art. Not a real artist. Didn't earn it, never paid their dues, wants the credit without doing the work.

None of that is aesthetic language. Every word of it is about what somebody deserves, and deserving is a fact about a person. It is not a property of an object on a wall.

And the system contradicts itself, out loud, constantly.

Here's the tell. The highest compliment available in any craft is effortless. She makes it look easy. He tossed it off. The sentence reads like it fell out of the sky. We reserve our deepest admiration for the total absence of visible labor while insisting that labor is the thing we're admiring. Visible effort in the object is a defect. Invisible effort behind it is mandatory. What we want is work we can be certain cost someone dearly and can find no evidence of, which is not an aesthetic standard. It's a faith.

Where the moral system comes from isn't mysterious. It comes from the apprenticeship. Anyone who spent a decade acquiring a hard skill has an enormous stake in the proposition that the decade was necessary, because the alternative reading of a decade is unbearable. That isn't cynicism, it's arithmetic performed on a life. But it produces the structure every guild has produced since the twelfth century: I suffered on the way in, therefore the suffering was load-bearing, therefore you suffer too.

So let me be exact about what effort actually earns, because it earns something.

Effort deserves respect. So do craft, and skill, and the ten years. That respect is a debt owed to a person. It is not a quality of the object, and the whole confusion lives in that transfer. Six hundred meticulous hours can produce a painting that leaves everybody cold. A photographer can take the picture of their career in a five-hundredth of a second and people will still be thinking about it at three in the morning. Labor, craft, and art overlap constantly. They are not synonyms.

And we already have a category of person who breaks the system, and we do not punish them for it. We call them prodigies.

The prodigy acquires in three years what costs everybody else thirty. The talent is unearned by definition, which is what the word means. By the logic of dues and desert, a prodigy is the purest free rider in the arts, taking the credit without doing the work, and our response is not to dock them. It's to put them on a stamp.

So the vocabulary doesn't even survive contact with cases we all already agree on. We don't actually object to unearned facility. We object to unearned facility arriving from outside the guild.

That moral system has a specific set of victims, and they are not the people it was built to catch.

Curb cuts were built for wheelchairs. They are used by strollers, suitcases, hand trucks, delivery carts, skateboards, bad knees, and everybody carrying something heavy, which over a long enough life is everybody. There's a boring name for this, the curb-cut effect, and it describes what happens every time a society builds an accommodation and discovers it built infrastructure.

Every accommodation runs that same cycle, with one extra accusation attached to the front end. Not merely cheating, but cheating that takes something from everyone else. Extra time on exams was an unfair advantage. Audiobooks weren't really reading, an opinion still sincerely held by people who should know better. Typing instead of handwriting was a shortcut. Calculators would destroy mathematics. The word processor, per a real strain of literary criticism in the eighties, made prose flabby, because revision had become too easy and writers no longer had to think before committing.

I want to be careful here, because this is the part of the argument that gets used as a shield, and it deserves better than that.

There is a specific grief in having a ghost and no shell.

If your brain is wired so that initiation is a wall and completion is a cliff, you accumulate a particular kind of wreckage. A hard drive full of chapter ones. Fourteen projects at eighty percent. Outlines more beautiful than anything you ever finished, because outlining was the part that felt like flying. The ideas were real. The wanting was real. The work is not less meant for having never made it out of your head. But it also isn't out of your head, and a thing that never got a shell has, functionally, no ghost either. It's lightning that happened inside one skull and then stopped.

For that person, a tool that gets a draft onto the page, or breaks the task into steps, or produces a mediocre version specifically so there is finally something to argue with, is not doing the art. It's doing the thing that was standing between them and the art. The blank page is not a sacred crucible. It's a wall, it's a taller wall for some people, and there is no medal for the height of your wall.

The counterargument is real and should be said out loud: struggle is sometimes generative, constraint produces invention, and the bad draft you fought through by hand taught you something the assisted draft won't. That's true. It's true the way it's true that walking to school in the snow builds character. It builds some character in some people and it keeps other people out of school.

The question was never whether the ramp is easier. It's always easier. That's what a ramp is.

Slop

The ramp argument has an obvious rejoinder, and it deserves the strong version rather than the easy one.

Fine, someone says. Build the ramp. Now look at what's coming down it.

That's fair, and there's a version of this essay that is smug and useless, and it's the one that treats every objection as a failure of imagination. So let's go look.

Open your relative's Facebook feed. Not to be cruel. As field research.

Dead-eyed children who have constructed Jesus out of shrimp. Log cabins containing seventeen chimneys. Six-fingered soldiers embracing puppies beneath the American flag with Jesus descending from Heaven. Birthday cakes at parties nobody attended. All of it captioned with some variant of why don't pictures like this ever trend.

It's awful. The word that's attached itself to it, slop, is exactly right, because slop describes not quality but volume and indifference. Slop is what you feed to something you don't respect, in quantity, without looking.

Now scroll further down the same feed, into the strata that predate all of this. The Minion memes. The Live Laugh Love. The motivational poster. The inspirational JPEG with twelve generations of compression artifacts and a quotation falsely attributed to Abraham Lincoln. The chain post announcing that most people won't repost this. Corporate clip art. Direct-to-DVD sequels. Cats.

We did not need artificial intelligence to invent slop. AI didn't invent bad art. It lowered the cost of producing it.

That's a real problem. It's just a different one, and it's worth naming precisely, because "people can use this to make garbage" is a weak argument against a medium and the strong version of the complaint is hiding right behind it.

The strong version is about the channel. A hundred thousand mediocre novels a year is not the same problem as a hundred, even if each individual book is equally harmless. The harm is ecological. It fills the pipe. It makes discovery statistically impossible, it collapses the entry-level economics that were how people used to learn, and it poisons the well of trust that made opening an anonymous thing worth doing at all. Nobody is worried that the slop has no soul. They're worried it has no off switch.

And that matters to this essay specifically, because if the ghost lives in the relationship between a maker and a stranger who encounters the work, then anything that makes the encounter unlikely is a direct attack on the mechanism, not a side complaint about taste.

There's a harder thing underneath all of it, though, and it's the one nobody wants to say.

Human authorship does not guarantee a ghost.

A technically magnificent corporate illustration can communicate nothing. A novel can contain a hundred and twenty thousand painstakingly hand-typed words and still feel dead. A film can employ eight hundred people and emerge with the personality of an airport Marriott. We know this. We have known it the whole time. The 2017 Ghost in the Shell is sitting right there.

The mistake is looking at the empty shells, which are real and numerous, and concluding that every shell must be empty. That's the Monet error with the polarity flipped.

The Other Questions, Which Are Real

Once you concede the channel, the rest of the concerns queue up behind it, and it forms fast. This is where the conversation usually detonates, so let's take it in order.

What about copyrighted training data? Yes. What about illustrators whose portfolios were scraped without permission? Yes. What about writers? Yes. What about corporations firing their art departments? Absolutely. What about the water and electricity? Yes. What about deepfakes? Yes. What about concentrated corporate power? Very much yes.

I mean all of those. Photoshop never asked to be paid, and it also never trained on the guy in the next cubicle. You can think the resulting technology is legitimate and still think the acquisition was a heist, and collapsing those two things is a rhetorical trick that AI's defenders run constantly and should stop.

It also matters enormously who is doing what. A dungeon master generating a portrait for six friends at a kitchen table is not a billion-dollar company dissolving an illustration department. Somebody running a small model on the gaming computer under their desk is a third thing. A person using it to get around executive dysfunction and finally finish a project that's been in their head for three years is a fourth. Treating these as one phenomenon because they share an underlying technology is like treating a home cook and Nestlé as the same because they both own a stove.

We need copyright rules. We need serious work on consent and licensing. We need transparency about data-center energy and water. Creative workers need bargaining power, and unions have a far better chance of protecting artists than declaring Photoshop's descendants metaphysically impure.

But here's the thing I want to be clear about, because separating these questions is not the same as dismissing them.

Whether a model's training violated an artist's rights tells you something urgent about the ethics of building that model. It does not tell you whether the output can put a lump in someone's throat. Whether a corporation fires three hundred illustrators tells you something damning about labor and capital. It does not define art. Whether a data center drinks an irresponsible amount of water tells you about environmental externalities. It does not tell you whether a poem is a poem.

We keep trying to load copyright law, labor politics, environmental policy, corporate concentration, accessibility, aesthetics, and the philosophy of consciousness onto a single question, and the question cannot carry that much baggage.

An AI researcher named Vivienne Ming put it about as well as it can be put: most of our fears about AI are fears about other people.

She's right, and those fears are frequently justified. But they're fears about other people, and they need to be fought where other people actually live, which is in law and contracts and organizing, not in a debate about whether a picture has a soul.

Director's Cut

Stanley Kubrick never acted in a Kubrick film. He didn't compose the scores, sew the costumes, build the sets, or, for most of his career, operate the camera. He notoriously shot a hundred takes to get one, which means his primary creative act was rejecting ninety-nine things other people had made.

Nobody has ever suggested Kubrick didn't make his movies.

Which is the question I asked you to hold, answered from an angle nobody finds threatening. Authorship survives near-total delegation of execution. We've known it for a century, we've never once found it strange, and we hand out the credit without pausing.

So the question was never how much of the work a person did with their hands. It's what has to be true about the deciding.

The entire apparatus of auteur criticism exists to explain a fact that would otherwise be confusing: a person who personally executes almost none of a work can still be its author, because authorship concentrates in judgment, not execution. Thelma Schoonmaker didn't shoot a frame of Raging Bull. She's the reason it hurts. Walter Murch made Apocalypse Now comprehensible out of a mountain of footage that had defeated everyone. The editor takes material they did not generate and, through selection and juxtaposition and rhythm, produces the emotion. Kuleshov again, but employed.

So when I say that working with these tools is closer to directing than to typing, I don't mean it as a comfort. I mean it as a job description, and it's harder than most people currently doing it seem to think.

Because directing has a failure mode, and the failure mode is ordering.

A prompt is an order. You state what you want, something arrives, you accept it. That's a drive-through. It requires taste the way ordering a McDonald's cheeseburger requires taste, which is to say technically yes and functionally no. It's also, right now, the overwhelming majority of what's happening, and the output looks exactly like what it is: the median, delivered promptly, in a bag.

Directing is what happens after. Generate twenty. No. Again. Closer. Wrong light. Keep that composition, change the clothes. Face from seventeen, background from six. Into Photoshop, mask this, paint over that, regrade it. Throw the whole thing out. Start again.

It's knowing that the thing which just arrived is almost right and being able to say precisely what's wrong, which requires a prior conviction about what right would be. It's the willingness to bin something that took four seconds to make, which is harder than it sounds, because sunk cost isn't the only thing that makes us keep bad work. Sufficiency does it too. Good enough is the enemy and these tools serve good enough on tap.

And there's a gravity to resist. These systems are trained toward the center of the distribution. They are, structurally, machines for producing the median, and art almost never lives at the median. It lives at the edges, in the specific, in the choices a committee would have removed. The pull is toward competent, pleasant, forgettable, which is the same direction as every focus group and every studio note that ever ruined a film by making it fine. Pulling against that is now a describable craft skill, and it's the one that separates the people making something from the people receiving deliveries.

Which is, of course, also a complete description of a director's job.

One caveat, and I want to state it plainly rather than bury it. Someone who spent thirty years learning to paint possesses an extraordinary craft that prompting does not magically reproduce, and pretending otherwise is both insulting and silly. Taste is a creative faculty, but it is not the same faculty as draftsmanship. The point is not that they're equivalent. The point is that different skill does not automatically mean no skill. A director isn't a failed cinematographer. A cinematographer isn't a failed actor. An editor isn't a lazy director. Film has always split conception, execution, and selection across people with different abilities and called the result authored. What's new is that the split is now available to one person at a kitchen table, sometimes brilliantly, frequently terribly.

The Net Is Vast

The 1995 Ghost in the Shell film does not end the way people usually remember it ending.

The Puppet Master, an intelligence that emerged from the sea of information and never had a body at all, doesn't attack Kusanagi. It doesn't try to replace her, or delete her, or wear her like a coat. It proposes a merger. It argues, reasonably, that a thing which can only copy itself is not alive, that variety and death are what make a life a life, and that what it wants is the ability to fail. And Kusanagi, who has spent the entire film unsure whether she's a person or a very good file, says yes.

What walks away is neither of them. It's something without precedent, looking out over a city that has no idea anything happened, and the last thing it says is that the net is vast and limitless.

That isn't a story about a human defeating a machine, and it isn't a story about a machine replacing a human. It's a story that finds the replacement question boring and declines to answer it. Thirty-one years later that remains the most sophisticated position anyone has taken on this, and it was reached by a man making a cartoon about a naked cyborg jumping off a building.

The reason she says yes is the part that gets skipped, and it's the part that matters here.

Kusanagi spends the film trying to verify her own ghost by inspection. She dives, she surfaces, she stares at her hands, she asks Batou whether he would know the difference. It never works and it cannot, because every piece of evidence she can reach is a piece of the shell, and the shell was manufactured. She is running the Monet experiment on herself and getting the same result the commenters got: examine a surface long enough and it will confirm whatever you already suspected.

Which is the answer to the question this essay has been circling:

"Is AI art really art" is malformed, because the word art is carrying two questions and they have different answers.

The first is did this thing do something to me. That one is origin-independent, it's real, and it belongs to you. A generated image can do it. So can a scam text, which is why the question was never sufficient on its own.

The second is was there anybody in here. That's a fact about how the thing was made, it is not a property of the object, and no amount of staring will produce it. The Monet experiment doesn't prove that provenance is irrelevant. It proves you cannot read provenance off a canvas, and the proof is that the test failed on an actual Monet.

But notice what the second question is not asking, because everyone gets this wrong, including the people asking it.

It is not asking whether a human was involved in the production line. Eight hundred humans were involved in the 2017 film. It cost a hundred and ten million dollars and eighteen months and the full attention of some of the most skilled people in the industry, and what came out was an empty shell. Maximum effort. No ghost.

That single object disproves both of the available positions at once. Effort is not the ghost, because that production had all of it. Human involvement is not the ghost either, because eight hundred people were involved in its making.

What was missing was a point of view.

Nobody on that production had a specific conviction about what the film was for that survived contact with everyone else's. It was assembled by consensus, and consensus is the absence of a point of view by definition. What consensus produces is the median, which is what it produced.

That's the ghost. Not effort, not authorship, not a human hand somewhere in the supply chain. A view about how the thing should be, held by somebody in particular, at the expense of the obvious alternatives.

Which is why Comedian passes in four seconds and the film fails after eighteen months.

And it's why the director analogy is not a convenient dodge but a structural fact.

A large language model cannot have a point of view. Not for mystical reasons. For statistical ones. It's trained toward the center of the distribution, which means its default output is the absence of a view, the median rendered fluently. That's what the slop is. That's what the pull toward competent and forgettable is.

Which is also why these things write such a beautiful corporate nothing-speak email. Beige is the correct answer to that question. An email declining a meeting is a document that should have no point of view, and a machine whose native output is the median is not a poor tool for that job, it is the ideal one. The complaint only starts when we point the same instrument at work where the median is a failure condition.

So if a point of view shows up in something made with one of these tools, there is exactly one place it can have come from: a person put it there. And the only mechanism available for putting it there is refusal. Rejecting the median output, over and over, in favor of something more specific that the model had no reason to prefer.

Kubrick's ninety-nine takes were never a display of effort. They were a view being enforced against everything that wasn't it.

So yes. There can be a ghost in art made with AI. It can only have come from the person directing it, because there is nowhere else in the system for it to originate. And it will be absent from most of it, for the same reason it was absent from the 2017 film: because taking the first acceptable result is easier than having an opinion, and always has been.

Now widen the lens, because the merger points at something bigger than one person's practice.

We keep running the AI art debate as a binary. Either a human makes the thing or the machine does. Both assume the output is a kind of thing that already existed.

That's what every novel medium does first. Early cinema was filmed theater, because the people holding the new machine understood it as a device for recording the art form it threatened. It took fifteen years and a lot of arguing before anyone worked out that the machine could cut, and that cutting wasn't a defect in the recording but the native grammar of a form nobody had invented yet. Sampling went the same way. It was not a cheaper session band. It became itself once producers realized a quotation carries meaning a live band structurally cannot produce, because a live band cannot cite.

Almost everything being made with these models right now is filmed theater, judged by whether it passes as the older thing, which is precisely the test that guarantees nobody discovers what the new thing does. Six point seven million people, arguing about whether an image successfully imitated an artifact from 1915.

I don't know what the native grammar of AI is, nobody does, and anyone touting a confident answer is selling something. But finding it would require exactly what the Puppet Master asked for: the ability to fail, repeatedly, in public, for years. A medium acquires a point of view the same way a person does, and by the same method.

One caution before that gets mistaken for a promise. Waiting twenty years for a native form is a worthless thing to say to an illustrator whose contract was cancelled last month. This is an argument about art, and it must never be conscripted as a labor policy.

So here's what I think, and it's less a verdict than a place to stand.

A shell without a ghost is a real failure, it has always been a real failure, and we hit it constantly without any help. A ghost without a shell is also a failure, and a sadder one, and for a lot of people the tool that finally gets the shell built is not a shortcut past the art. It's the ramp to the door.

But I've been talking about the ghost as though it were an item. Something located somewhere, which you could in principle open the shell and find, the way you'd find a component. That's the assumption the whole AI art argument rests on, mine included, and I'm beginning to see it as a mistake.

The ghost in art isn't a noun. That's why nobody can locate it, and why Kusanagi couldn't either.

Think about what it would have taken to fix the 2017 film.

Not less AI. There wasn't any. Not more human involvement either, because there were eight hundred people on that production and most of them were extremely good at their jobs. Not more money, more time, more research, or more reverence for the source, all of which it had in quantities the 1995 film could only have dreamed about. Oshii made his version for under a tenth of the budget, in a country where the industry was not yet convinced anime could carry that kind of ambition, and what he had that the remake didn't was a specific and slightly hostile opinion about what the story was for.

One person with something particular to say. That's the entire difference, and it cost about a hundred million dollars to demonstrate.

So when somebody asks whether AI-generated work can be art, they are asking a question that has never once been answered by counting human hands. A banana taped to a wall passes in four seconds. An eighteen-month recreation of a scene somebody else already got right fails, and it fails for the reason committees always fail, which is that nobody in the room was willing to say a thing the others might have refused.

That's the ghost. Not the humans, not the hours, not the pigment or the pixels or the plank. Somebody with a view about how the thing should be, holding it against every easier version of itself, for exactly as long as they keep holding it.

It isn't in the shell. It never was. It's whoever is still standing there at two in the morning, looking at something that is almost right, and saying no. Again. Closer.

The correct question was never was a human involved.

It was always did anybody have something to say.