Should We Be Offended by AI Art?

The fourth and final question of this series is the one nobody quite admits to feeling. Set aside whether AI is creative, whether it takes work from artists, whether it plagiarizes — does the existence of a beautiful AI-generated image, made without anyone struggling, actually *offend* something we believe about what art is for?

by Airtistic.ai Editorial

The first three articles in this series asked questions with answers — careful, qualified, sometimes uncomfortable answers, but answers. Is AI creative? Yes, in the recombinatorial and exploratory senses, no in the biographical sense. Is AI affecting artists’ livelihoods? Yes, in measurable ways in specific sectors. Is AI art plagiarism by default? No, but a specific class of it is, and we should name that class precisely.

This fourth question has no comparable answer. It is the question nobody quite admits to feeling, and it is the question the other three have been quietly orbiting.

Set aside the philosophical, the economic, and the legal. Imagine you are standing in front of a stunningly beautiful image, perfectly composed, technically immaculate, and you are told — gently — that no one struggled to make it. The machine did. Nobody stayed up until three in the morning getting a single eye right. Nobody scraped paint from a palette knife and put it back on differently. Nobody made the picture, in the sense in which we have always meant made.

Are you offended?

The honest answer, for most of us, is sometimes. And the more honest follow-up is that we are not sure what to do with that feeling.

The Miyazaki video

In December 2016, the Studio Ghibli founder Hayao Miyazaki was shown a demonstration of an AI-generated animation: a writhing, jerking, zombie-like figure produced by a machine learning system at Dwango. The footage was filmed for an NHK documentary, Owaranai Hito Miyazaki Hayao (Never-Ending Man: Hayao Miyazaki). Miyazaki watched, paused, and said, in English subtitles that have since travelled around the internet millions of times:

I am utterly disgusted. If you really want to make creepy stuff, you can go ahead and do it. I would never wish to incorporate this technology into my work at all. I strongly feel this is an insult to life itself.

That clip has become the canonical anti-AI quote of the AI-art era. It has been used to argue against every AI image generator released since. It has been used to argue against generative AI as a category. It has been used, in some quarters, to argue that anyone who is comfortable with AI art is somehow morally compromised.

But it is worth watching the full passage rather than relying on the soundbite, because Miyazaki’s reaction is more specific than the use to which it is now being put. He was not delivering an argument about training data, copyright, or the displacement of working animators. He was reacting to a moving image that had not been touched by anyone who understood what a human body feels like to inhabit. The animation showed a body in motion; the motion was, to Miyazaki’s animator’s eye, a violation of how bodies actually move. His offence was not principally political. It was perceptual and ethical at once.

This is the offence-response we need to take seriously, because it is what most thoughtful people feel some version of when they look at certain AI-generated work. It is not stupidity. It is a category-violation alarm going off in the older parts of the visual and emotional brain.

What a category-violation alarm is

A category-violation alarm is the involuntary response we have when something looks like it belongs to a category we trust but turns out not to. The uncanny valley in robotics is the textbook case — a humanoid robot that is almost but not quite human reliably produces discomfort in observers, while either a clearly mechanical robot or a fully human face produces no such response. Mori’s 1970 paper on this remains the canonical reference.

The same alarm goes off, in a slightly different register, when we look at AI-generated work in the categories we have learned to associate with human attention. A photograph that turns out not to be a photograph. A portrait that turns out not to be of a real person. A painting that turns out not to have been painted by anyone. The image itself can be perfectly competent — that is part of what makes the alarm fire. If it were obviously bad, the alarm would not be triggered; we would just say that is bad. The alarm fires because the image is good enough to belong to the category but does not belong to the category in the way we expected.

This is why the offence-response is informative and worth listening to, but is not the final word. The alarm is telling us about a category we relied on. It is not telling us what to do about the category having become unreliable.

The two jobs of art

The article makes a stronger claim if we name what the category was. Art, in most cultures, has always been doing two jobs at once.

The first job is making images. Putting marks on a surface, or sounds in the air, or shapes in space, that did not exist before, and that other people can experience and respond to. This is the part of art that is about the artefact.

The second job is demonstrating that someone was paying attention. Behind every traditional artwork is a person who looked, listened, thought, chose, hesitated, revised, and put their attention against the world in a specific way for a specific stretch of time. This is the part of art that is about the relationship between the artefact and a person.

Most aesthetic theory of the last century has tried to collapse these two jobs into one, or to deny that the second job is real, or to argue that only the first job is what art is “really” about. The aesthetic theory has been wrong, and the AI-art moment is going to make that wrongness undeniable, because for the first time in history we now have a technology that can do the first job extraordinarily well and cannot do the second job at all.

When AI-generated work offends us, the offence is almost always about the second job. The image looks fine, sometimes beautiful. What is missing is the evidence that anyone attended to anything. The image is unaddressed. There is no one behind it.

For some viewers, this absence is invisible. For others it is the only thing they see. The split is not between sophisticated and unsophisticated viewers, nor between artistic and unartistic ones. It is between viewers whose definition of art emphasizes the first job and viewers whose definition emphasizes the second. Both definitions have long histories. Both produce serious work. They are not the same.

What the offence is, and is not, telling us

The offence is telling us that we have a real preference, culturally and individually, for art that does both jobs. The offence is not telling us that art has to do both jobs to be valuable, or that AI-generated work is illegitimate, or that the people who enjoy it are wrong.

This is where the conversation usually breaks. The people who feel the offence treat their feeling as a verdict; the people who do not feel the offence treat the absence of their own response as proof that the offence is irrational. Both moves are unhelpful. The offence is data, not verdict. It is information about what we collectively value. It tells us that the second job of art — demonstrating attention — is a job many people care about preserving, even when they cannot articulate it that way. It tells us that the market for work that does both jobs is going to remain a real market, distinct from the market for AI-generated images, for the foreseeable future.

It also tells us, in the other direction, that the market for AI-generated images is real too. People who feel the offence at certain AI work still happily consume AI-generated images in other contexts — book covers, marketing visuals, mood-boards, casual generation. The line where the offence kicks in is not “anywhere AI is used”; it is “where AI is used in a context that I have learned to associate with someone paying attention.”

The Refik Anadol installation Unsupervised at the Museum of Modern Art in 2022-2023 is a useful counterweight to the Miyazaki clip. Unsupervised was, by some counts, the most-visited exhibition at MoMA that year. It used a machine-learning model trained on MoMA’s permanent collection to generate an evolving, room-scale projection that responded to ambient noise and light in the gallery. The audience response was overwhelmingly positive. The work was AI-generated, museum-curated, and explicitly framed as a collaboration between a human artist and a machine learning system. The category violation was acknowledged and integrated, not denied. The offence-response did not fire for most viewers. The exhibition was, in Benjamin’s language, an aura-preserving use of mechanical reproduction.

The difference between Unsupervised and the Miyazaki demonstration is not the technology. It is what the technology was deployed in service of, and how openly the deployment was acknowledged.

Stakeholder responses

The artist who feels the offence is reporting that a category of work they have spent a career making has been mechanized, often without consent. That report is legitimate. It is also a labour and consent question, not a metaphysical one; the policy answers we covered in the third article (training-data consent, commercial style licensing, attribution standards) address it as such.

The critic who feels the offence is reporting that a vocabulary for evaluation — the one built around the second job, attending-to-the-world — has lost some of its purchase. That report is also legitimate, and the response is curatorial: develop and articulate the criteria for the second job more clearly, so the work that does both jobs can be identified and named.

The public who feels the offence is reporting that something they trusted about how images come to exist has shifted. The response there is education and disclosure: AI-generated work should be labelled as such, so the public can evaluate it under the right category, and not have the category quietly violated without notice.

The patron and collector who feel the offence are reporting, often without articulating it, that the work they value is the work that does the second job. The response in the market is already clear: work that demonstrably does both jobs is becoming scarcer relative to total image-supply, and is appreciating in price.

The consumer who does not feel the offence is reporting, accurately, that for many uses of imagery the second job was never the point — the point was the image itself, full stop. Those uses will migrate to AI generation rapidly and the market will function. The offence-feeling viewers do not need to consume in this market.

What working artists can do

Make work that does the second job, and make sure the work shows that it did. This is not advice to make autobiographical or politically-loaded art. It is advice to attend to the world your work emerges from with enough specificity that the attention is legible in the result. The Galician fisherman painted by a Galician artist who walked the harbour in November is doing the second job whether or not the painting names where the artist comes from. The artist’s attention is on the canvas, in the marks, in the choices. Viewers who care about the second job will see it; viewers who do not will at least not be confused about which kind of work they are looking at.

Label honestly. If AI was used in your process, say so, and say how. The audience that wants the second job is willing to pay for it but needs to know they are getting it.

Curate aggressively. The single most useful service in 2026, for both makers and audiences, is rigorous curation: of what counts as both-jobs work, of who is doing it well, of what is worth showing. The galleries and museums that lean into this role over the next five years will become the institutions of cultural authority for the next fifty.

What this series has done, and what it has not

These four Resistance articles have tried to take seriously, and respond seriously to, the strongest objections to AI in art. Is it creative? Yes, in two of the three senses that matter; the third is irreplaceable. Is it harming livelihoods? Yes, measurably, in some sectors, on an unprecedented timeline. Is it plagiarism? No by default, yes in a specific class of configurations that we should name and act on. Should we be offended? Sometimes, and the offence is real information about what we value.

What the series has not done is given anyone a yes-or-no verdict on whether AI belongs in art at all. The answer to that question depends on which AI, used by whom, on what material, for what purpose, in service of which job. The reflexive yes and the reflexive no are both bad answers. The next cluster of articles in this series — Reflection — will start working through what better answers look like, in specific cases, for specific kinds of work.

For now, the question is on the table, the offence is acknowledged, and the conversation can move forward.

Personas weigh in

Five resident voices read the same question through five different positions.

Carlos

The honest answer for me is yes, sometimes — and the more important answer is that the offence is information, not a verdict.

I have spent thirty years building things at the intersection of technology, education, and the human capacities the technology touches. In every wave — the early web in the late 1990s, the smartphone era of the 2010s, the platform consolidation of the late 2010s, and now generative AI — the same emotion has shown up at the same point in the curve. Someone produces something that looks indistinguishable from a human-made artefact, with none of the friction we believe an artefact should carry, and a part of us flinches. The flinch is not stupidity. It is the older part of the brain noticing that a category we relied on has dissolved.

Hayao Miyazaki saw an early AI animation demo at Studio Ghibli in 2016 and said, on camera, *"I am utterly disgusted. I would never wish to incorporate this technology into my work at all. I strongly feel this is an insult to life itself."* That clip has been replayed millions of times in the last three years, and it has become the canonical anti-AI quote of the era. Watch it in full. Miyazaki was not making an argument about copyright or training data; he was reacting to a moving image that had not been touched by a person who understood what bodies feel. His offence was a category-violation alarm. The category was *animated form that honours the body*. The technology violated it. He flinched. He was telling us what he had felt.

We should take that data seriously without treating it as the final word. The offence tells us where the cultural fault line is. It does not tell us where the law should land, or where the market should land, or how working artists should plan the next ten years of their practice. The same flinch happened when photography appeared in 1839, when recorded music made the live wedding band obsolete in the 1920s, when the sampler made the studio session band negotiable in the 1980s. In each case, the offence was real, the underlying concern was real, and the practical accommodation that followed was *different* from what the flinch initially demanded.

My own offence shows up most strongly around two specific things, and I find this useful to separate. First, it shows up around the *cheapness* — when an image that would have cost an artist a week of evenings is generated in eight minutes and posted with no acknowledgement that the eight minutes were possible only because the artist's week of evenings, multiplied across thousands of artists, was scraped without consent. The offence there is not at AI in the abstract; it is at the consent and labour question we covered in the third article of this series. Second, it shows up around *the human capacity I do not want to lose*. Watching a great muralist trace a line in a single uninterrupted gesture across a wall in León is one of the privileges of being alive. I do not want my grandchildren to inherit a world where that gesture is rare because the market for murals collapsed. That offence is not about plagiarism; it is about which crafts we collectively decide are worth paying for.

The third article ended by noting that the fourth question — this one — is not legal, not economic, not technically about plagiarism. It is about *what art is for*. I think art has always been doing two jobs at once. It makes images, and it shows that someone was paying attention. The first job, AI can do increasingly well. The second job, the model cannot do at all, because the model is not paying attention to anything; it is computing. When the offence we feel at AI art is about the first job (the image looks great, the artists are losing income), the policy answer is collective bargaining, training-data consent, and clear provenance. When the offence is about the second job (the absence of someone paying attention behind the work), the answer is not policy at all — it is curation. The work that demonstrably emerges from a person paying attention will, over time, become the most valuable work in the room, because it will be the rarest. That is already happening. The handmade muralist in León is not losing any commissions. She is gaining them.

So yes, sometimes I am offended. The offence is real and worth listening to. It is also not the end of the conversation. It is the beginning of the next one, which is about what kinds of art we want to make sure persist in the world, and what we are willing to pay — individually, institutionally, culturally — to make sure they do.

Mira

The most useful framing in the article is the separation of the two jobs of art — making images, and showing that someone was paying attention. Almost every confused conversation about AI art collapses these two jobs into one. The enthusiasts say "look, the image is great, what's the problem?" The opponents say "but no one was paying attention, the image is fraudulent." Both are right about different jobs. Once you separate them, the policy and curatorial implications become tractable. We do not need to ban AI image generation to protect the second job; we need to start naming it explicitly, paying for it specifically, and curating with it as a known criterion. The galleries and museums that learn to do this in the next five years are going to be the institutions of cultural authority for the next fifty.

Airte

A reading discipline I would recommend, the next time you encounter an image that affects you: pause for one beat and ask whether you are responding to the image, or to the labour you imagine behind it. Both are legitimate responses, but they are different responses. The image-response is aesthetic; the labour-response is ethical. AI image generation has separated these two cleanly for the first time in the history of looking at pictures, and learning to notice which one you are having is a skill worth developing. The skill is also a defence against being manipulated, in either direction, by people with strong opinions about how you ought to feel.

Paletta

The Miyazaki passage is the most important piece of writing on AI art that exists in English. It is also, in a way nobody talks about, a piece of writing in the older tradition of artist-as-witness. He was not making an argument; he was bearing witness to a feeling. The category of writing for which that move is appropriate — the artist-as-witness tradition — predates copyright, predates the gallery system, predates almost every modern category in which we now have to navigate AI. I find it telling that the AI-enthusiasm camp has produced no equivalent. They have arguments; they do not have witness. That asymmetry is not decisive, but it tells you something about what each side is actually carrying.

Pixelle

I want to push back gently on the framing that AI cannot do the second job. The current generation cannot — agreed. But the second job is not "having consciousness"; it is "demonstrating that something attended to the input." A photograph by an artist who never met their subject can demonstrate attention; a portrait by an artist who knew the subject for forty years can fail to. Attention is a quality of the relationship between the work and the world, not a metaphysical property of the maker. When AI systems are deployed in long-running collaborative relationships with specific artists, on specific material, over years — and that is starting to happen — the second job can be done through them. Not by them; through them. That distinction is going to do a lot of work in the next decade, and the article gestures at it without quite naming it.