Essay

Whose Feeling Is It?

This page was written with AI.

Three tests people actually use to decide whether something is art: Tolstoy’s transmitted feeling, Danto’s artworld, and the count of choices an artist makes. Each applied honestly to a machine-made image, then what a US court and the US Copyright Office have decided, and why their answer is about ownership rather than art.

Short answer

It depends on which test of art you use, and the essay leaves the general question open. Tolstoy’s test asks whose feeling is passed on; Danto’s asks whether the artworld takes a work in; a third counts the choices a person made, which Ted Chiang says a short prompt cannot match. US law answers a different question: purely machine-made images cannot be copyrighted, because it requires a human author. Every test, including whether AI is creative, ends up asking whether someone is on the other end.

Few questions about AI start arguments faster than this one. An illustrator sees a generated picture win praise and says it is not art. Someone who has spent a week refining prompts says it is. A student writing an essay on it finds mostly opinion pieces that pick a side in the first line. This essay does something slower. People who say “that isn’t art” are usually applying one of three tests, often without naming it. Below, each test is stated in the words of someone who held it, applied to a machine-made image as honestly as I can manage, and checked for what it would also rule out. Then comes what the law in the United States has actually decided, which is narrower than either side tends to claim.

How the sources were handled. Every quotation below was copied from a text opened with a public request on 1 October 2026, and each one is recorded in this site’s source ledger with its address and the exact sentence. Tolstoy is quoted from Aylmer Maude’s English translation as published by Project Gutenberg. Danto’s 1964 paper is not freely available and I could not open it, so his view is described in my words and not quoted.

Test one: does it pass on a feeling?

Leo Tolstoy’s What Is Art? (1897) is the most quoted answer to the question in its own title, and it is blunter than most people remember. After a long survey of theories of beauty, all of which he rejects, he gives his own definition in chapter V:

Art is a human activity, consisting in this, that one man consciously, by means of certain external signs, hands on to others feelings he has lived through, and that other people are infected by these feelings, and also experience them.Leo Tolstoy, What Is Art?, chapter V, translated by Aylmer Maude

Tolstoy’s test has two halves, and both matter here. The first is on the receiver’s side. In chapter XV he calls infectiousness the one indubitable indication distinguishing real art from its counterfeit: if you feel what the maker felt, and feel joined with them and with others who feel it too, the thing is art. The second is on the maker’s side. A feeling has to have been lived through by someone and handed on consciously, and he says the strongest single condition of infection is sincerity: But most of all is the degree of infectiousness of art increased by the degree of sincerity in the artist. When an audience senses the maker does not feel what they are expressing, he writes, a resistance immediately springs up.

Apply that to a picture produced from a one-line prompt. A viewer may well be moved by it; image models are trained on millions of pictures that moved people, and they are good at reproducing the surface of feeling. That passes the receiver’s half. The maker’s half is harder. The model has no feelings it has lived through, so if a feeling is being transmitted at all it has to be the prompter’s. And the prompter’s feeling reached the image only through a sentence; the colours, lines and forms that Tolstoy names as the “external signs” were chosen by the model. On a strict reading, what moves the viewer was handed on by no one. On a generous reading, the prompter who had a feeling, tried hundreds of images and kept the one that carried it is doing something like what a photographer does with a camera: choosing, not drawing.

Two honest limits. First, Tolstoy’s test was never meant to be friendly to most of what galleries hang. He used it to dismiss Wagner’s Ring as counterfeit art, and in a footnote to chapter XVI he applies it to himself: I must, moreover, mention that I consign my own artistic productions to the category of bad art, excepting the story God sees the Truth, which seeks a place in the first class, and The Prisoner of the Caucasus, which belongs to the second. A test that demotes War and Peace is not a reliable instrument for sorting the art on anyone’s wall. Second, Tolstoy makes art depend on what happens inside the maker and the viewer, which nobody can inspect. The test tells you what question to ask (whose feeling is it?) rather than giving a verdict you can read off the image.

Test two: does the artworld take it in?

The second test starts from the opposite end. In 1964 the philosopher Arthur Danto published a short paper, “The Artworld”, in the Journal of Philosophy. His problem was Andy Warhol’s Brillo Box: an object that looked, to the eye, almost exactly like the cardboard cartons of soap pads stacked in any supermarket stockroom, yet was exhibited as art in a New York gallery. If the gallery box and the stockroom box look the same, whatever makes one of them art cannot be something you see. Danto’s answer was that seeing something as art depends on a background the eye cannot supply: a body of theory about art and a knowledge of its history. He called that background the artworld. (I could not open the paper itself, which sits behind a journal paywall, so this is a description in my own words rather than a quotation.)

Applied to AI images, Danto’s test gives a split answer and an uncomfortable one for both sides. It does not care how an image was made; Duchamp’s urinal and Warhol’s soap-pad box both made it in. So the fact that a machine produced the pixels cannot, on its own, disqualify anything. But the same test says that most generated images are not art either, in the way that most photographs on a phone are not: nobody has offered them as art, inside the conversation in which things count as art. What would make a particular AI image art, on this view, is its being made and shown as a move in that conversation. Ted Chiang, whose case against is set out below, notes that the film director Bennett Miller used DALL-E 2 to make images that were exhibited at the Gagosian gallery, generating more than a hundred thousand images to arrive at the twenty that were shown. On Danto’s terms that is a clear candidate. A wallpaper generated for a slide deck is not.

The limit here is the mirror of Tolstoy’s. Danto’s test explains how art status is conferred, but it makes the answer depend on what a fairly small professional world decides to do. For the illustrator who says “that isn’t art”, it can sound like a refusal to answer: it tells them the question will be settled by curators and critics, not by the work.

Test three: how many choices did someone make?

The third test is the one most working artists reach for, and it has been put most clearly by the science-fiction writer Ted Chiang in “Why A.I. Isn’t Going to Make Art”, published in The New Yorker on 31 August 2024. His generalisation is that art is something that results from making a lot of choices. A ten-thousand-word story involves something like ten thousand choices; a hundred-word prompt, about a hundred. The model has to fill in the rest, and it does so either by averaging, which he says is equivalent to the least interesting choices possible, or by imitating a particular artist. On images:

An artist—whether working digitally or with paint—implicitly makes far more decisions during the process of making a painting than would fit into a text prompt of a few hundred words.Ted Chiang, “Why A.I. Isn’t Going to Make Art”, The New Yorker, 31 August 2024

This test does something the other two do not: it gives a rough measure. It also cuts in both directions, and Chiang says so. He imagines a tool that would let someone enter tens of thousands of words over many sessions to control an image in fine detail, and writes: I’d say that a person could use such a program and still deserve to be called an artist. Bennett Miller’s hundred thousand generations are his example of someone working that way. His objection is to the ordinary use, the one the products are sold for, where a few words go in and a finished image comes out.

What the law has decided

The questions above are philosophical, and courts do not answer them. But two US decisions are often quoted as if they had, so it is worth reading what they actually say.

The court. Stephen Thaler, a computer scientist, built a system he called the Creativity Machine. It produced a picture that he titled A Recent Entrance to Paradise, and he applied to register its copyright, listing the machine as the sole author and himself as the owner. The Copyright Office refused; a district court agreed; and on 18 March 2025 the US Court of Appeals for the District of Columbia Circuit affirmed. The opinion, by Judge Millett, states the holding in one sentence:

The Creativity Machine cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being.Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. March 18, 2025), opinion of Circuit Judge Millett

The court was careful about what it was not deciding. It did not reach Thaler’s argument that he himself was the author, because that argument had not been raised before the agency. And it said the rule does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence; it requires only that the author of that work be a human being—the person who created, operated, or used artificial intelligence—and not the machine itself. The opinion opens by framing what is at stake: who or what counts as the author is a question that implicates important property rights undergirding economic growth and creative innovation. Property, not aesthetics.

The Copyright Office. In January 2025 the US Copyright Office published Copyright and Artificial Intelligence, Part 2: Copyrightability. Its summary lists eight conclusions. The ones that bear on this question are these, quoted from that list:

Copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements.U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, January 2025
  • The use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output.
  • Whether human contributions to AI-generated outputs are sufficient to constitute authorship must be analyzed on a case-by-case basis.
  • Based on the functioning of current generally available technology, prompts do not alone provide sufficient control.
  • Human authors are entitled to copyright in their works of authorship that are perceptible in AI-generated outputs, as well as the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs.

Its reasoning on prompts is close to Chiang’s, though it arrives there from a different direction: Prompts essentially function as instructions that convey unprotectible ideas. The report also notes, in a footnote, that a second case challenging a refusal to register an AI-generated output, Allen v. Perlmutter, had been filed in Colorado in September 2024; it does not say how that case will end, and neither can I.

What this does and does not settle. The human-authorship rule decides who can own a work. It does not decide what counts as art, and neither the court nor the Office claims it does. A photograph of a sunset by a tourist is copyrighted the moment it is taken, whether or not anyone would call it art; a picture made entirely by a machine has no copyright owner under US law, whether or not a museum hangs it. So “the Copyright Office says AI art is not art” is false, and so is “the court ruled that AI can make art”. What the law has said is narrower and more useful: where a person controlled the expressive result, that person can be an author; where the machine made those choices, nobody is.

The strongest case that it is not art

Put at full strength, the case against does not rest on the pictures being ugly. Many are not. It rests on what art is for. Chiang states it at the end of his essay: Whether you are creating a novel or a painting or a film, you are engaged in an act of communication between you and your audience. A model, on his account, has no intention to communicate; he writes that ChatGPT feels nothing and desires nothing, and that generative AI reduces the amount of intention in the world. Combine that with the choice count and you get a precise claim: in a typical AI image, the parts that look most artful were chosen by no one who meant anything by them. That is Tolstoy’s sincerity condition in modern language. The viewer may be moved, but the thing moving them is, in Chiang’s word, mimicry. On this view, calling it art is not generous; it hides a loss, because it trains people to stop expecting that someone is on the other end.

The strongest case that it is

Put at full strength, the case for does not claim the machine is an artist. Its best statement I have read comes from Aaron Hertzmann, a computer graphics researcher, in “Can Computers Create Art?” (2018, written for the journal Arts). He agrees that a computer is not the author: he argues that artistic creation is primarily a social act, and that computer-generated art has, in every case to date, been credited to the human artist(s) behind the tools, such as the authors or users of the software. His case is that AI is a tool in human hands, and that we have had this argument before. When photography arrived, he recounts, many people held that it could not be art because a machine made it; and he opens that history with the painter Paul Delaroche, quoted as saying at a demonstration of the daguerreotype in 1839: From today, painting is dead! Photography became an art form anyway, through the choices photographers learned to make with the machine. His prediction:

Far from replacing artists, new technologies become new tools for artists, invigorating and changing art and culture.Aaron Hertzmann, “Can Computers Create Art?”, 2018 (arXiv:1801.04486)

Hertzmann wrote before today’s text-to-image products existed, so he did not address the one-line prompt directly. But his argument survives the change. If art is something people make for other people, then an image that a person conceived, iterated on, selected and offered to an audience is that person’s work, and the question “is it art?” has the same answer as for any photograph: sometimes, depending on what they did with it. Notice how close this is to Chiang’s own concession about the many-session tool, and to the Copyright Office’s line between a tool that assists and one that stands in. The two sides agree on more than their headlines suggest. Where they disagree is on whether the ordinary use, a few words in and an image out, involves enough of a person to count. What happened to the portrait painters whose trade the daguerreotype ended, and what being surpassed by a machine does and does not take from someone skilled, is the subject of a separate essay here, What Excellence Was For.

Three images, three answers

The tests give different answers for different kinds of AI image, which is why the argument never ends when it is had about “AI art” in general. A rough sorting, with the tests named:

  • A picture from one short prompt, taken as it came. Tolstoy: no one’s lived feeling chose its forms. Chiang: too few choices. Danto: not art unless someone offers it as art. US law: no human author, so no copyright. Every test leans against.
  • Thousands of generations, steered and selected over months, then exhibited. This is Miller’s case. Chiang himself allows that such a person could deserve to be called an artist; Danto’s test is met once it is shown; Hertzmann would call it art made with a new tool. The law would look at it case by case, and the selection and arrangement can be protected even where individual outputs are not.
  • An artist’s own drawing or photograph, altered or extended with AI. Every test treats this mostly as the artist’s work, and the Copyright Office says copyright covers the human expression that is perceptible in the result.

Can AI be creative?

The art question has a twin that people ask just as often: not whether the picture is art, but whether the machine that made it was being creative. The two come apart. A photograph can be art without the camera being creative, and a child’s drawing can be creative without anyone hanging it in a gallery. So the second question needs its own account of what creativity is, and its own evidence.

What creativity is. The account most used by people who build creative programs comes from the cognitive scientist Margaret Boden. In “Computer Models of Creativity” (AI Magazine, 2009) she defines creativity as the ability to generate novel, and valuable, ideas, and splits “novel” in two: new to the person who had the idea, which she calls P-creative, and new to all of history, H-creative. She then names three ways new ideas arise:

  • Combinational: unfamiliar combinations of familiar ideas, as in poetic imagery, collage, or an analogy such as the heart as a pump.
  • Exploratory: working inside an accepted style or “conceptual space” (a school of painting, a national cuisine, a theory of molecules) to find what it allows, including places nobody had visited.
  • Transformational: changing one of the rules that define the space, so that ideas become possible that could not have been generated before. Her example is Kekulé’s proposal that benzene is a ring, which broke the assumption that organic molecules are strings of carbon atoms.

Her verdict on machines is not the one most people expect. Combinations sound easy, and she agrees that a computer can produce new ones until kingdom come. The hard part, she says, is getting it to generate and prune them so that most are interesting, because that takes the store of world and cultural knowledge a person brings. Boden wrote before today’s image and text models, which are trained on a large share of the public internet and are, in effect, an attempt at that store. On my reading, most of what they do fits her first two kinds: they recombine, and they explore styles that already exist. Whether any of their output is transformational, an image that changes what images can be, is the open part. She is also careful about what her account cannot settle. Whether a computer could ever be “really” creative, she writes, is not a scientific question but a philosophical one, and for now an unanswerable one.

Lady Lovelace’s objection. The oldest form of the doubt is older than electronic computers. Ada Lovelace, writing about Charles Babbage’s Analytical Engine in a memoir Turing dates to 1842, said the machine could only carry out instructions. Alan Turing named the objection after her in his 1950 paper Computing Machinery and Intelligence, where it is the sixth of the nine objections he answers, and he quotes her words:

The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to performAda Lovelace (1842), as quoted by A. M. Turing, Computing Machinery and Intelligence, Mind, 1950

Turing’s reply has three parts. First, he reads her as saying that the evidence available to her gave no reason to think the machines could originate anything, not that they could not; and since the Analytical Engine was a universal computer, he argues, it could in principle mimic any machine that could. Second, he takes the version that a machine can never do anything “really new” and turns it on us: Who can be certain that “original work” that he has done was not simply the growth of the seed planted in him by teaching, or the effect of following well-known general principles. Third, against the version that a machine can never take us by surprise, he offers plain testimony:

Machines take me by surprise with great frequency.A. M. Turing, Computing Machinery and Intelligence, Mind, 1950

He expected the comeback, that the surprise is a creative act on his part and earns the machine no credit, and answered that the appreciation of something as surprising requires as much of a “creative mental act” whether the surprising event originates from a man, a book, a machine or anything else. Near the end of the paper he returns to Lovelace with the route he thought would answer her properly: machines that learn, whose teacher will often be very largely ignorant of quite what is going on inside. That is a fair description of how today’s models are made. Nobody writes the instructions for a particular picture or poem; the behaviour comes out of training on examples, and the people who trained the model cannot trace why it produced one image rather than another. In the literal sense Lovelace meant, these systems do things nobody knew how to order them to perform. Whether that is the sense that matters is the rest of the argument.

What the creativity tests found. Psychologists have a narrower question they can test. Their standard laboratory measure of creative thinking is a divergent-thinking task, most often the Alternate Uses Task: list unusual uses for an everyday object such as a rope or a box. Two studies in Scientific Reports put chatbots through it beside people, and their results point in different directions:

  • Koivisto and Grassini (2023) compared 256 adults (native English speakers recruited online, aged 19 to 40, mostly in the UK and US) with three chatbots, ChatGPT-3.5, GPT-4 and Copy.ai, each run in 11 separate sessions in spring 2023. Everyone gave uses for a rope, a box, a pencil and a candle; the people had 30 seconds per object. Answers were scored two ways: by software measuring how far each answer’s meaning lies from the object’s (semantic distance), and by six human raters, not told that some answers came from AI, on a scale of 1 to 5. On average the chatbots scored higher. But the best human answers matched or beat the machines’: of the top scores per object, people held all but one on the software measure and all but two on the raters’ scores. Much of the gap came from the bottom. In 7 per cent of trials a person’s best answer was rated below 2, an ordinary or confused use; the chatbots never scored that low.
  • Hubert, Awa and Zabelina (2024) compared 151 US adults recruited online with 151 separate GPT-4 sessions on three tasks: alternate uses for a fork and a rope (three minutes each for people), a consequences task (“Imagine humans no longer needed sleep”), and a task asking for ten nouns as unrelated to each other as possible. So that the machine could not win on volume, each GPT-4 session was told to give exactly as many answers as a matched person had. Originality was scored by software, with no human raters. GPT-4 scored as more original on all three tasks and wrote far longer answers, about 15 words per idea against 3. The authors add their own caveat: these tests measure creative potential, not achievement, and because the model does not have agency, its potential is dependent on the assistance of a human user to elicit responses.

Both results come with a limit worth knowing. The tests were built to measure people, and their best-known answers have been written down many times. Koivisto and Grassini say plainly that they cannot tell whether the chatbots combined ideas or recalled them: It remains possible that they simply retrieve ideas that exist in their database. They add that the same problem applies to human participants, who may also be repeating uses they have heard.

The strongest objection. Put at full strength, the case against calling the output creative does not deny the scores. It says creativity, as the word is used of people, belongs to a process and a person, not to a product alone. Look at who did what in the studies above. A person chose the task, the object, the number of answers and the wording of the prompt; software or human raters judged which answers were good; and, by Hubert and colleagues’ own account, without a prompt the model does nothing at all. Boden’s definition needs ideas that are novel and valuable, and in every case the judgement of value was made outside the machine. Add the possibility that the novelty is recalled rather than made, and the machine looks like a very fast generator of candidates, with the creative act, deciding what is worth making and recognising which result is good, still done by the person at the keyboard. That is Lady Lovelace’s objection updated for a learning machine: it originates nothing that was not, in some form, put into it.

Turing’s reply still answers part of this, since people too are trained on what came before them; the seed planted by teaching was his phrase. On Boden’s definition and on the tests psychologists actually use, current models produce ideas that are new and that people rate as good, often better than an average person manages in thirty seconds, though in the 2023 study not better than the best people. Whether that makes the machine creative turns on the same question the art tests ended on: whether creativity lives in the result, or in someone who meant it.

What the argument shows

This essay does not settle whether AI art is art, and I do not think anyone can settle it for all cases, because the word does different jobs for different people. What the argument does show is what people want art to be.

Notice where every test ended up. Tolstoy asked whose feeling it was. Danto asked who offered it, and to whom. Chiang counted the choices a person made. Hertzmann conceded the computer is not the artist and looked for the human behind the tool. Even the law, which was not asking about art at all, went looking for a human being. None of these is a test of how a picture looks. Each is a way of asking whether someone is on the other end.

That is the real disagreement, and it is about what people want from art. Those who say AI images are art usually care most about what the image does to the viewer, and they are right that it can do a great deal. Those who say they are not care most about being addressed by another person, and they are right that, in the plainest use of these tools, no one has quite addressed them. The question that is left is not about machines. It is what you are looking for when you look at a picture: something beautiful, or someone.

Written for AItheism. If you think a step in the argument is wrong, that is the most useful thing you can notice — hold onto it. Further reading on this and neighbouring questions is on the reading list.

Made with AI tools. · Privacy