I really like this essay, and I think your point about originality is good. But I’d have two things to add in support of AI art not really being art: first, as you say, getting that last percentage is the hard part. But it’s important to say _why_ it’s so hard: the real innovation in many works of art (broadly defined, including writing and music and so on) is in the subtle ways they break accepted rules and subvert expectations. Definitionally, that’s pretty hard to engineer into something that is essentially a giant interpolation engine that deviates from maximum likelihood based on local statistical fluctuations, particularly since most such violations just read as weird and dumb to the viewer. Second, intentionality matters at least to my mind because it’s a source of human connection across time and distance, e.g. David Foster Wallace died years ago but I can feel his pain and wonder and neuroses when I read infinite jest. Or even in this piece, I read it and wonder about the thought process behind each paragraph and clever turn of phrase. LLMs are lousy for writing long documents as they start to lose the thread after a while, but even once that’s eventually solved, I’d have no interest in reading an LLMs novel no matter how technically skillful. There’s an emptiness there that nothing short of machine consciousness can remedy, and we are nowhere near that (if it’s even possible).
The elephant is extrapolating its experiences. AI is extrapolating its training data. Same thing. You can try to complexify it with large words, but art can be done by many things, mostly animals, and now AI.
Just because the elephant does not know about Michaelangelo, does not mean that what it is producing is not art. Your explanation is ridiculous, and it is senseless.
You say:
"LLMs are lousy for writing long documents as they start to lose the thread after a while, but even once that’s eventually solved,"
That is right, an LLM that does 100 MILLION tokens using MAMBA 2 (SSSM6). That is right, longer context with 100% reliability up to tens of millions of tokens. Show me a human that can do that in 10 seconds. Your time is almost up.
By your definition an anthill is art, and likely a sunset is too. And I mean, great! People have argued about what art is and means basically since cave painting. But if you want to remove “something made by an intelligent being to convey meaning to intelligent beings” as a critical part of the definition of that word, then i don’t think there is much to debate, given an irreconcilable difference of definitions.
The elephant is a funny example because while it’s not clear either way if it knows it’s painting an elephant, they are among the smartest animals and likely have unique names, complicated language patterns, burial of their dead (which may imply the existence of elephant religion), and so on. I’d happily grant that an elephant painting can be art. But an elephant likely has intentionality the way all evidence so far suggests an LLM or diffusion model does not.
Likewise a context window of millions of tokens doesn’t mean that it can use all that coherently. ChatGPT already has tens of thousands of tokens in its context window, enough for at least short novels. Amazon is filled with them. They’re terrible and I don’t want to read them, and as I said, even once they become technically decent (which should happen eventually), I still have no interest.
In a discussion of art, saying you can create ten years of human speech in one output is not making your case stronger. Who wants to listen?
I do sympathize with your thoughts. There are lots of people out there that will need to change. They complain about it. The reality is that it will change just like your interpretation of art. People will not have jobs, but maybe, if they are smart, they can do their work for the billions of agents and stay in a museum where the AI can reminice about the old ways. I prefer not to live in a cave or tent, so I will go with the tech for more reasons than you know. I would be happy to discuss it if you want. The future I am helping make.
Art is subjective. That is final, what you define it as is your business. But since you did not create the word, you should not redefine it. Stick with the 2500 year old definition. techne (τέχνη), an ancient Greek term that encompassed both practical skills and craftsmanship, as well as intellectual and theoretical knowledge. Techne was associated with the mastery of skills, whether it was in the crafting of objects (like pottery or architecture) or in the practice of disciplines such as mathematics, astronomy, or philosophy. In that sense, I think AI has it and probably will be better at it in almost all senses. Just saying.
Accuracy retrieving hashes is quite impressive, but that’s a very different thing from being able to compose that stored information into a compelling next paragraph again and again in a way that maintains logical and thematic consistency for a hundred pages. One is retrieval, the other is retrieval and very complex processing of it. I have not seen any AI system that can even approach that yet. I bet we’ll get there eventually but it’s a hard problem.
And as I said, mastery of skills is not all there is. As Ted Chiang put it, it’s the choices that make things interesting. Imagine a version to Stanley Kubrick who only ever made HR training videos. Would they be better than average, for what they are? Sure. Would you want to watch one instead of the Shining?
Woah woah woah. I am a PhD, and writing hundreds of pages of anything at once has only happened once in my life. And it was tough. And it took years. Litterally. https://sakana.ai/ai-scientist/ This does it for 15 dollars.
Not sure why you have not seen this yet.
Ph.D.-level research with peer review quality is already here.
Ironically enough, your same points have been made before by Margaret Boden with the concept of the “superhuman human fallacy”—which posits that the inclination to dismiss AI's creativity because its outputs don't always reach the peak of human creativity overlooks the fact that most humans rarely reach those heights themselves. But to echo your sentiments, I don't think it'll be impossible to find a direct or indirect precursor to Boden's fallacy either.
I personally really resonate with Ben Davis's analysis of AI art in his book "Art in the After-Culture". He touches on a lot of historical currents and the political economy of artistic creation but the one undercurrent that continues to frame my view is that at the end of the day, it's not really about "what art is" but it's about "what we want artists to be".
The "artist" identify is itself constructed and retroactively applied to individuals from our past and through this process of social construction, we have and continue to redefine the identity. AI art threatens that definition more than anything else.
Ben Davis argues that "AI aesthetics" are inherently "prosumer". If I as a consumer want to hear a song that sounds like if Bob Dylan sang for the Beatles, the logic of AI art makes that possible and desirable for me. If I want a novel written based on my own life's events and written in the style of my favourite author, then so be it!
In those examples, the quality and originality of the art object is irrelevant. What is relevant is the power shift they represent. In that future, the "artist" identity could become "producer of art objects that cater to consumer preference" instead of what we currently valorize, which is something along the lines of "someone who reframes our world through their perspective".
Like Hito Steyerl alludes to in her essay "Mean Images", the verb "training" is occurring before and after the creation of the AI output: the training first applies to the AI and then it is applied by the AI after the object is produced. What is important to me is to what extent we allow the logic of AI aesthetics to "train" us into redefining what we want artists to do for us socially.
Art is the result of media, which comes from the word medi, or "in the middle".
The artist is literally a medium that has an input and output. Whether you like it or not, whether it "measures" up, or not, is subjective. You are playing on Dante and Locke, and even going back to Thales of Miletious from the Ionian school more than 2500 years ago at a time when art already existed. Art is the result of input and output through the interpretation language that is used to interpret it. It is the creation of something from another thing and impossible to have without that creation.
If god exists, then we could say that we are his art, and he is the first artist. So, what is art? It is subjective, as the word itself is an interpretation of the reality of humans that humans experience, as we create it. So, why not other species, and why not other entities or machines?
This was surreal and delightful to read because (1) I agree, (2) I wasn't expecting to get quoted, (3) I forgot I said that and still agree with myself, (4) I wrote an essay last week and organically came to the same exact conclusion, but this time instead of IDEA + CHOICES + EXECUTION I called it INSPIRATION + TASTE + CRAFT, which is the same thing in a slightly more upmarket fiction kind of way.
I frankly think that this discussion--"Is AI agentic? Can AI form opinions? Can AI make art?"--is a totally worthwhile discussion worth having. However, I think the 99% of people who are serving as spectators at this big intellectual tennis match eyeballs flicking from thinkpiece to thinkpiece would be better served asking, "Am I agentic? Can I form opinions? Can I make art?"
And, in what I find a kind of delicious turn of irony, the conclusions you trend toward here are the freeing, load-bearing ideas that can get people started toward answering yes to all three. Those conclusions: (1) most of being agentic is small decisions, and (2) most opinions are small and unimportant, and (3) most art is unoriginal.
I'd add one more factor to your formula: + iNSPIRATION. Knowing the history of the artist plays an important part in our response to that art. Does it come from exposure to trauma? Experiential insight? Unique observation? We judge art often on the influential inputs the artist can identify rather than just on the making of the art itself. I think that may be why a unique human experience trumps all artistic output ever as expressed by AI.
Writers and artists are being affected though - just not the ones higher up in the food chain.
The onramps to making a living from writing or art - freelance corporate copy writing and illustration - are already basically gone, outcompeted by 'good enough' and much cheaper AI.
It's the timeless progression of technology. Scribes also lost their jobs to the printing press, typists to computers, and now here we are writing with text online in seconds.
One more thing to consider. Art lives just as much in the experience of the viewer/listener/consumer as it does in the producer of the work. Does it make us think/see/feel things differently, in ways that we associate with the idea of what art is supposed to be/do (which can be different for different people)? Then it is art.
Can AI make things that do that? Certainly. Is it plugged in well enough to the deeper pulse of human experience to do it as well as or better than a human artist? Or can it arrive at something beyond the capacity of the human through its own path of production? Hmm, well that's worth thinking about.
Either way, though, as you speak to toward the end, any artist's work will touch some people more than others. And the relationship between the audience and the artist is something that goes beyond the momentary experience of a given piece of art. Will some focused AI personas capture the attention of a specific audience? Probably. Will that take away from the relationships and audience of human artists? Probably not.
I think this piece is great but I also think Chiang's essay does contain one slightly different point – that all of the details in a piece of art are not original necessarily so much as choiceful. And they can be that without being novel. When we engage with art we know that each choice that went into it, even when unoriginal, is made from a particular sensibility and for particular reasons, so there is a sense of communion with another creative intelligence that we will not get when the choices are made by machine.
I feel Chiang slightly undersells art, in reducing it to "choices". He rightly says that AI art system proponents think of art as all inspiration, no perspiration, and rightly condemns that attitude. But he replaces it with art being composed of many decisions at many scales. I think even if an AI art system presented all these decisions to the user one by one it would still not be equivalent to the art-making process. Art involves doing something, not just making choices. I feel I'm arguing against all of Information Theory here but so be it.
Because I had never heard of it before, this idea of ‘slim real creativity margins’ seemed very original to me. I have no idea to what extent it really is, but this made me think of a third reason why humans will remain far more relevant than AI in the world of ideas for a long time: they have the unique discernment to recognize which ideas—whether fully original or not—are worth pondering and spreading—as Eric exemplifies with this post.
Confirming your point and adding another that covers both this essay and Ted's is one I wrote Jan 2023 on whether AI could be creative where I came to the same conclusion that humans aren't as creative as we'd like to think.
What about the role of human unconsciousness in the creation of art? Chiang and others put so much attention on "conscious intent" and "communication" for defining art, but human artists have always talked about very original works of art "coming through" them without much conscious intent. And art that attempts to consciously communicate something (e.g., political propaganda "art") often feels lifeless or silly. How much does / could AI's fuzzy synthesis of prior material mimic this unconscious, flow state in humans, because I see some pretty bizarre output from AI art.
Regarding art audiences, too often, humans simply attempt to make conscious meaning from art, especially abstract art, instead of focusing on our emotional experience. We intellectually try to figure art out -- "Hey, the lump looks like a face, and that might be a tree," instead of "Wow, I'm not sure why, but there's something interesting / appealing about the way I feel when I encounter this painting."
The irrational, the emotional, the unconscious -- aren't these essential elements in both the creation and experience of art?
Absolutely. When I write I often get into that flow state, where ideas and words seem to simply flow through. It’s a surreal experience. My 14-year-old, who’s a professional artist, does the same with her visual art. When she’s in the zone she can create a stunning work with virtually no correction. I’ve watched her do it—it sends chills down your spine bc she’s so young.
I think the essay trips on this acknowledgment: "Chiang executes his argument about AI’s inability to create art very well, since he’s a great writer."
You make good points regarding his ideas not being original. But I don't think that's important. It's true that it's very difficult to come up with a radically new idea, and so incredibly rare. But great writing, and art in general, don't depend as much on the production of radical new ideas as on the execution of culturally interesting ones.
You could have ChatGPT give arguments against its being an artist and it'd probably give you back a summary of the ideas Chiang and others have given, but it wouldn't be the piece Chiang wrote. And it's in the specifics of his essay, how he wrote it and what he illuminates, that make it interesting.
The myth of Oedipus existed before Sophocles wrote his play. What makes his play interesting and long lasting is how he did it. I think that's an essential feature of art and what Chiang is exploring.
First off, I'm surprised by your equivalence: art = non-fiction essay = all human cultural production. I'd argue that that the threshold for art is higher, and that while there are essentially never non-fiction (but non-scientific products) that are completely new, surprising, and shocking, the history of art is full of them.
But you're also making some statements that I find at least highly questionable:
"it’s likely that human artists and writers and thinkers will, for the foreseeable future, maintain control of the means of distribution of ideas and creative output. Humans will have sticky control over audiences and the cultural narrative in exactly the way that large publications have a sticky control compared to smaller outlets or individuals." - artists and thinkers already *aren't* in control of those means of distribution. Publishers and similar entities are, and more and more the big tech platforms. If those decide to foreground "AI"-generated artefacts over human ones, there's little that human artists can do.
Related to this: "Until they become better in a slim-margin business, then lacking the means of distribution and the stickiness of human creators, AI will be forced into the role of tool, not artist replacement." - possibly, but whose tool. If they are a tool for producing cultural products for a mass audience at scale, or for replacing concept artists and asset creators in the videogame industry (as is anecdotally already the case), they will very much replace current artists.
"AI must win consistently more points than the best humans, who already operate on razor-slim margins of originality, insight, or creativity to begin with (much like Federer’s mere 4% advantage). " - possibly but "AI" can also play much more than human creators. If human creators produce 1 artefact in 10 that speaks to audiences and "AI" produces 1 in 1000 then the latter draws equal if it produces 100 times as much as the human, which is easily doable. AlphaGo played billions of matches against itself, something that no human, no matter how smart, creative, talented, could ever achieve.
"Cars still passed on the highway. Here in New England, the leaves will change into orange and yellow and then die and come again. An ant makes its way across a playground structure and will trace the same path tomorrow. The magnetic poles of Earth remain unflipped for the last 780,000 years. " - of course, the material world, and even the consumerist society won't actually change if non-derivative art disappears...but that's not the impact that's claimed by defenders of this kind of art.
Well said Albrecht, I felt the same. Especially about the control of distribution point. Big Tech is solidifying universal control of human expression, be it art, communication, or influence, and that I find the most worrying.
Great, thought-provoking article. Same re comments. As I read, I keep asking a question of this article, a question that is simultaneously unoriginal (the other commenters make the point) and original (no one asked the question this way): where does a unique mind like Emily Dickinson fit into this analysis? She was part of the communal mind (and thus unoriginal) and entirely apart from it. And from there I can extrapolate to other great minds, and from them to ordinary minds occasionally having their own unique thoughts. There is mystery in creation, whether you look at the Big Bang or childbirth or Bob Dylan writing a song. Can AI experience mystery, can it be comfortable with uncertainty, can it imagine its way into that tiny percentage of new human thinking? I suspect (and hope) that AI will always be chasing the unattainable horizon of human imagination. (Of course, when it gets tired of the chase it will just wipe us out, but that’s another story.)
I suspect your reply to Chiang may be missing the point a bit. Strawson's essay, "Freedom and Resentment" really catalyzed for me (and later Wittgenstein helped) a particular way of thinking about how we treat other people. Strawson was talking about how we can take up one of two stances vis-a-vis another human being -- I may be getting the labels wrong here -- the 'objective' versus the 'personal' stance. The former is roughly the stance of science, which explains human behavior in terms of cause and effect. The latter is the stance crucial to relationships, and (Strawson's focus) taking and holding one another responsible for our conduct. So, for example, you might explain Sam's outburst this afternoon by pointing to their hunger and its influence on their mood and how this precipitated the outburst. Or you might explain Sam's outburst in terms of his taking himself to have been slighted. The former is the world of causes and effects; the latter, of intentionality and reasons. It isn't exactly that one is right and the other wrong, nor even that only one of these stances is the ethically appropriate stance (I think both a more 'objective' and 'personal' approach are, at times, required in caring for others). But what has proven intractably hard, so far anyway, is reducing one (inevitably the personal) to the other.
Anyway, I suspect that christening a work as 'good art', 'original', 'creative', etc. attributes it to a *person*, as their (non-moral but still ethically loaded) *responsibility*. And we just haven't incorporated AI into these interpersonal practices yet -- they remain *things* to be explained in terms of cause and effect -- and so an AI can't make art in the same way that the color blue can't be (literally) 100 degrees Celsius. It's a category mistake. Might we one day? Well, maybe. I think in a vaguely Wittgensteinian way that we don't have cut-and-dried rules for inclusion here -- witness our changing and contested circles of concern over human history!
But so you may well be able to point to the products of AI and the products of people and say, look, we cannot identify any significant causal differences -- are people 'really' not copying like AI does? -- but that does not necessarily mean that you should be agnostic as to whether one is an original/creative/good work of art and the other not, or that you should presume they both are. Those issues turn crucially on whether one is the work of a person, and the other the output of a thing.
If I interpret you correctly, I think we actually probably agree on this. I still stand by everything I said in the original AI Art is an Art essay, that the philosophical definition of art for me is a communication across consciousnesses. There's no subjective experience behind whatever the AI is outputting, so I don't consider it to be art. But there are also other definitions of art. Some of those focus more on the effects of the viewer, and clearly you can still be emotionally affected by some piece of AI-generated art. So I think technical ability must come into play, and on that I think the opposite of Chiang: most stuff is deeply unoriginal, most choices are deeply unoriginal, and really good artists/writers/thinkers are just *slightly* more original.
I think we're sort of on the same page. I don't like your first definition even if it sounds like I do, because I'm again more Wittgensteinian about 'subjective experience' and 'consciousness' than most people nowadays are. (Relatedly, regarding 'communication across consciousnesses': what would the criteria for success even look like?) My starting point is that there are communities with conceptual toolkits for engaging in and developing the practices of those communities, and this is the context in which things, including works of art, 'light up with meaning', and in which they relate to one another as people. We haven't sorted out how AI will fit into our practices, but they definitely do not yet enjoy the status of persons/fellow practitioners.
Great article, Erik, thanks! It's easy to argue, as Ted Chiang did, that the key feature of great art is originality. But this ignores two aspects of creativity that I believe are more important: surprise and craft.
Great art can open our eyes to a new perspective on existing art or the real world. So originality might actually be more about surprise than uniqueness. This is why "original ideas" never seem particularly original once you start digging a bit. Surprise brings delight and activates fundamental emotional reactions. Originality might bring delight, but you could fill a screen with a number to the ten-thousandth decimal that has never been seen before and that would be original without producing the emotions of surprise and delight.
There's a reason that the word "arts" is frequently accompanied by "crafts". The perception of originality in art can also be provoked by a uniquely personal method of execution on the part of the artist. This craft-driven level of artistry is something that AI de-emphasizes. Sure, you can produce a draft of a short story by inputting an idea into an LLM, but if you want it to be "artistic" you're going to have to take that lumpen draft and rewrite it into a shimmering gem using your own ability as a writer. No amount of re-prompting will achieve that outcome.
So my diagnosis of AI's ability to produce great, or even good, art, is that it's not necessarily originality that matters, but the human input that produces surprise and exhibits craft.
You cited "Pierre Menrad, Autor del Quijote", instant upvote. But the art of Pierre Menrad lies in writing the Quijote as himself, and just by Pierre Menrad having existed, the existance of the Quijote's changes. First, because you'll never know wether you hold an original or a Menrads copy. His endeavour is far more profund than a mere copy, could you imagine if you yourself had written 'La Odisea'? The writing changes, it's not longer is a mythological recount of a possible event, but pure fantasy. As Menrad did, GPT in it's willingess to summon the Hyperspace that it represents into an answer that a human Made, too shaped the perception of text written by other humans. The same effect Menrad had on all other copies of El Quijote, GPT has over all text written post birth.
I hope that I can create such endearing art as Menrad, unfortunately not everyone is a Menrad. Is quite hard to be Pierre. The transformation that he made is not a mere blog, or a novel or a book: Is a change of perception to everyone who willl ever read the art he transformed. On a superficial level the shortstory is about transforming something through your own point of view. But Borges never is superficial.
I don’t see why this mythical “true originalism” should be north star of discourse. Many of the best ideas are actually knowing *when* and *how* to combine existing ideas, especially ideas from far flung places or disciplines. When an engineer takes inspiration from the paws of a little known animal in the Amazon rainforest, or when a linguist takes an idea he heard about in an obscure physics course and adapts it to his own craft - none of these are original thoughts, but they’re incredibly valuable.
Forget this “originalism”, I for one welcome the chance to see how AI can recombine existing ideas in a myriad of incredible ways we mere mortals would never dream of.
I really like this essay, and I think your point about originality is good. But I’d have two things to add in support of AI art not really being art: first, as you say, getting that last percentage is the hard part. But it’s important to say _why_ it’s so hard: the real innovation in many works of art (broadly defined, including writing and music and so on) is in the subtle ways they break accepted rules and subvert expectations. Definitionally, that’s pretty hard to engineer into something that is essentially a giant interpolation engine that deviates from maximum likelihood based on local statistical fluctuations, particularly since most such violations just read as weird and dumb to the viewer. Second, intentionality matters at least to my mind because it’s a source of human connection across time and distance, e.g. David Foster Wallace died years ago but I can feel his pain and wonder and neuroses when I read infinite jest. Or even in this piece, I read it and wonder about the thought process behind each paragraph and clever turn of phrase. LLMs are lousy for writing long documents as they start to lose the thread after a while, but even once that’s eventually solved, I’d have no interest in reading an LLMs novel no matter how technically skillful. There’s an emptiness there that nothing short of machine consciousness can remedy, and we are nowhere near that (if it’s even possible).
What you describe is human emotion. Not art. Art is simple. You try to humanize it. Check the elephant painting (it is real) https://www.youtube.com/watch?v=foahTqz7On4
The elephant is extrapolating its experiences. AI is extrapolating its training data. Same thing. You can try to complexify it with large words, but art can be done by many things, mostly animals, and now AI.
Just because the elephant does not know about Michaelangelo, does not mean that what it is producing is not art. Your explanation is ridiculous, and it is senseless.
You say:
"LLMs are lousy for writing long documents as they start to lose the thread after a while, but even once that’s eventually solved,"
I say pay attention: https://magic.dev/blog/100m-token-context-windows
That is right, an LLM that does 100 MILLION tokens using MAMBA 2 (SSSM6). That is right, longer context with 100% reliability up to tens of millions of tokens. Show me a human that can do that in 10 seconds. Your time is almost up.
As a reference, that is ten years worth of human speech in one output. https://cloud.google.com/blog/products/ai-machine-learning/magic-ai-100m-tokens-cloud-supercomputer
By your definition an anthill is art, and likely a sunset is too. And I mean, great! People have argued about what art is and means basically since cave painting. But if you want to remove “something made by an intelligent being to convey meaning to intelligent beings” as a critical part of the definition of that word, then i don’t think there is much to debate, given an irreconcilable difference of definitions.
The elephant is a funny example because while it’s not clear either way if it knows it’s painting an elephant, they are among the smartest animals and likely have unique names, complicated language patterns, burial of their dead (which may imply the existence of elephant religion), and so on. I’d happily grant that an elephant painting can be art. But an elephant likely has intentionality the way all evidence so far suggests an LLM or diffusion model does not.
Likewise a context window of millions of tokens doesn’t mean that it can use all that coherently. ChatGPT already has tens of thousands of tokens in its context window, enough for at least short novels. Amazon is filled with them. They’re terrible and I don’t want to read them, and as I said, even once they become technically decent (which should happen eventually), I still have no interest.
In a discussion of art, saying you can create ten years of human speech in one output is not making your case stronger. Who wants to listen?
I do sympathize with your thoughts. There are lots of people out there that will need to change. They complain about it. The reality is that it will change just like your interpretation of art. People will not have jobs, but maybe, if they are smart, they can do their work for the billions of agents and stay in a museum where the AI can reminice about the old ways. I prefer not to live in a cave or tent, so I will go with the tech for more reasons than you know. I would be happy to discuss it if you want. The future I am helping make.
First, read, then speak. https://magic.dev/blog/100m-token-context-windows 100% accuracy across the board until 32 million tokens. 3 years of human speech.
Art is subjective. That is final, what you define it as is your business. But since you did not create the word, you should not redefine it. Stick with the 2500 year old definition. techne (τέχνη), an ancient Greek term that encompassed both practical skills and craftsmanship, as well as intellectual and theoretical knowledge. Techne was associated with the mastery of skills, whether it was in the crafting of objects (like pottery or architecture) or in the practice of disciplines such as mathematics, astronomy, or philosophy. In that sense, I think AI has it and probably will be better at it in almost all senses. Just saying.
Accuracy retrieving hashes is quite impressive, but that’s a very different thing from being able to compose that stored information into a compelling next paragraph again and again in a way that maintains logical and thematic consistency for a hundred pages. One is retrieval, the other is retrieval and very complex processing of it. I have not seen any AI system that can even approach that yet. I bet we’ll get there eventually but it’s a hard problem.
And as I said, mastery of skills is not all there is. As Ted Chiang put it, it’s the choices that make things interesting. Imagine a version to Stanley Kubrick who only ever made HR training videos. Would they be better than average, for what they are? Sure. Would you want to watch one instead of the Shining?
Woah woah woah. I am a PhD, and writing hundreds of pages of anything at once has only happened once in my life. And it was tough. And it took years. Litterally. https://sakana.ai/ai-scientist/ This does it for 15 dollars.
Not sure why you have not seen this yet.
Ph.D.-level research with peer review quality is already here.
Ironically enough, your same points have been made before by Margaret Boden with the concept of the “superhuman human fallacy”—which posits that the inclination to dismiss AI's creativity because its outputs don't always reach the peak of human creativity overlooks the fact that most humans rarely reach those heights themselves. But to echo your sentiments, I don't think it'll be impossible to find a direct or indirect precursor to Boden's fallacy either.
I personally really resonate with Ben Davis's analysis of AI art in his book "Art in the After-Culture". He touches on a lot of historical currents and the political economy of artistic creation but the one undercurrent that continues to frame my view is that at the end of the day, it's not really about "what art is" but it's about "what we want artists to be".
The "artist" identify is itself constructed and retroactively applied to individuals from our past and through this process of social construction, we have and continue to redefine the identity. AI art threatens that definition more than anything else.
Ben Davis argues that "AI aesthetics" are inherently "prosumer". If I as a consumer want to hear a song that sounds like if Bob Dylan sang for the Beatles, the logic of AI art makes that possible and desirable for me. If I want a novel written based on my own life's events and written in the style of my favourite author, then so be it!
In those examples, the quality and originality of the art object is irrelevant. What is relevant is the power shift they represent. In that future, the "artist" identity could become "producer of art objects that cater to consumer preference" instead of what we currently valorize, which is something along the lines of "someone who reframes our world through their perspective".
Like Hito Steyerl alludes to in her essay "Mean Images", the verb "training" is occurring before and after the creation of the AI output: the training first applies to the AI and then it is applied by the AI after the object is produced. What is important to me is to what extent we allow the logic of AI aesthetics to "train" us into redefining what we want artists to do for us socially.
https://www.youtube.com/watch?v=foahTqz7On4 I guess this elephant did not listen to art critics.
Art is the result of media, which comes from the word medi, or "in the middle".
The artist is literally a medium that has an input and output. Whether you like it or not, whether it "measures" up, or not, is subjective. You are playing on Dante and Locke, and even going back to Thales of Miletious from the Ionian school more than 2500 years ago at a time when art already existed. Art is the result of input and output through the interpretation language that is used to interpret it. It is the creation of something from another thing and impossible to have without that creation.
If god exists, then we could say that we are his art, and he is the first artist. So, what is art? It is subjective, as the word itself is an interpretation of the reality of humans that humans experience, as we create it. So, why not other species, and why not other entities or machines?
Excellent comment
This was surreal and delightful to read because (1) I agree, (2) I wasn't expecting to get quoted, (3) I forgot I said that and still agree with myself, (4) I wrote an essay last week and organically came to the same exact conclusion, but this time instead of IDEA + CHOICES + EXECUTION I called it INSPIRATION + TASTE + CRAFT, which is the same thing in a slightly more upmarket fiction kind of way.
I frankly think that this discussion--"Is AI agentic? Can AI form opinions? Can AI make art?"--is a totally worthwhile discussion worth having. However, I think the 99% of people who are serving as spectators at this big intellectual tennis match eyeballs flicking from thinkpiece to thinkpiece would be better served asking, "Am I agentic? Can I form opinions? Can I make art?"
And, in what I find a kind of delicious turn of irony, the conclusions you trend toward here are the freeing, load-bearing ideas that can get people started toward answering yes to all three. Those conclusions: (1) most of being agentic is small decisions, and (2) most opinions are small and unimportant, and (3) most art is unoriginal.
PS- here's the essay I wrote: https://charliebecker.substack.com/p/if-you-let-it-ai-will-prompt-you
I'd add one more factor to your formula: + iNSPIRATION. Knowing the history of the artist plays an important part in our response to that art. Does it come from exposure to trauma? Experiential insight? Unique observation? We judge art often on the influential inputs the artist can identify rather than just on the making of the art itself. I think that may be why a unique human experience trumps all artistic output ever as expressed by AI.
I agree, actually. The formula is inspiration + taste + craft.
Writers and artists are being affected though - just not the ones higher up in the food chain.
The onramps to making a living from writing or art - freelance corporate copy writing and illustration - are already basically gone, outcompeted by 'good enough' and much cheaper AI.
It's the timeless progression of technology. Scribes also lost their jobs to the printing press, typists to computers, and now here we are writing with text online in seconds.
https://cloud.google.com/blog/products/ai-machine-learning/magic-ai-100m-tokens-cloud-supercomputer
;) 10 years of human speech in one output. Yeah, I agree with you except... AI will be the higher-up in less than a year from now
One more thing to consider. Art lives just as much in the experience of the viewer/listener/consumer as it does in the producer of the work. Does it make us think/see/feel things differently, in ways that we associate with the idea of what art is supposed to be/do (which can be different for different people)? Then it is art.
Can AI make things that do that? Certainly. Is it plugged in well enough to the deeper pulse of human experience to do it as well as or better than a human artist? Or can it arrive at something beyond the capacity of the human through its own path of production? Hmm, well that's worth thinking about.
Either way, though, as you speak to toward the end, any artist's work will touch some people more than others. And the relationship between the audience and the artist is something that goes beyond the momentary experience of a given piece of art. Will some focused AI personas capture the attention of a specific audience? Probably. Will that take away from the relationships and audience of human artists? Probably not.
I think this piece is great but I also think Chiang's essay does contain one slightly different point – that all of the details in a piece of art are not original necessarily so much as choiceful. And they can be that without being novel. When we engage with art we know that each choice that went into it, even when unoriginal, is made from a particular sensibility and for particular reasons, so there is a sense of communion with another creative intelligence that we will not get when the choices are made by machine.
I feel Chiang slightly undersells art, in reducing it to "choices". He rightly says that AI art system proponents think of art as all inspiration, no perspiration, and rightly condemns that attitude. But he replaces it with art being composed of many decisions at many scales. I think even if an AI art system presented all these decisions to the user one by one it would still not be equivalent to the art-making process. Art involves doing something, not just making choices. I feel I'm arguing against all of Information Theory here but so be it.
Insightful essay.
Because I had never heard of it before, this idea of ‘slim real creativity margins’ seemed very original to me. I have no idea to what extent it really is, but this made me think of a third reason why humans will remain far more relevant than AI in the world of ideas for a long time: they have the unique discernment to recognize which ideas—whether fully original or not—are worth pondering and spreading—as Eric exemplifies with this post.
Confirming your point and adding another that covers both this essay and Ted's is one I wrote Jan 2023 on whether AI could be creative where I came to the same conclusion that humans aren't as creative as we'd like to think.
https://www.polymathicbeing.com/p/can-ai-be-creative
Perfect!
What about the role of human unconsciousness in the creation of art? Chiang and others put so much attention on "conscious intent" and "communication" for defining art, but human artists have always talked about very original works of art "coming through" them without much conscious intent. And art that attempts to consciously communicate something (e.g., political propaganda "art") often feels lifeless or silly. How much does / could AI's fuzzy synthesis of prior material mimic this unconscious, flow state in humans, because I see some pretty bizarre output from AI art.
Regarding art audiences, too often, humans simply attempt to make conscious meaning from art, especially abstract art, instead of focusing on our emotional experience. We intellectually try to figure art out -- "Hey, the lump looks like a face, and that might be a tree," instead of "Wow, I'm not sure why, but there's something interesting / appealing about the way I feel when I encounter this painting."
The irrational, the emotional, the unconscious -- aren't these essential elements in both the creation and experience of art?
Flow state is the product of actual, living human experience arising out of a unique human mind.
Absolutely. When I write I often get into that flow state, where ideas and words seem to simply flow through. It’s a surreal experience. My 14-year-old, who’s a professional artist, does the same with her visual art. When she’s in the zone she can create a stunning work with virtually no correction. I’ve watched her do it—it sends chills down your spine bc she’s so young.
I think the essay trips on this acknowledgment: "Chiang executes his argument about AI’s inability to create art very well, since he’s a great writer."
You make good points regarding his ideas not being original. But I don't think that's important. It's true that it's very difficult to come up with a radically new idea, and so incredibly rare. But great writing, and art in general, don't depend as much on the production of radical new ideas as on the execution of culturally interesting ones.
You could have ChatGPT give arguments against its being an artist and it'd probably give you back a summary of the ideas Chiang and others have given, but it wouldn't be the piece Chiang wrote. And it's in the specifics of his essay, how he wrote it and what he illuminates, that make it interesting.
The myth of Oedipus existed before Sophocles wrote his play. What makes his play interesting and long lasting is how he did it. I think that's an essential feature of art and what Chiang is exploring.
That was one of your stranger ones! :)
First off, I'm surprised by your equivalence: art = non-fiction essay = all human cultural production. I'd argue that that the threshold for art is higher, and that while there are essentially never non-fiction (but non-scientific products) that are completely new, surprising, and shocking, the history of art is full of them.
But you're also making some statements that I find at least highly questionable:
"it’s likely that human artists and writers and thinkers will, for the foreseeable future, maintain control of the means of distribution of ideas and creative output. Humans will have sticky control over audiences and the cultural narrative in exactly the way that large publications have a sticky control compared to smaller outlets or individuals." - artists and thinkers already *aren't* in control of those means of distribution. Publishers and similar entities are, and more and more the big tech platforms. If those decide to foreground "AI"-generated artefacts over human ones, there's little that human artists can do.
Related to this: "Until they become better in a slim-margin business, then lacking the means of distribution and the stickiness of human creators, AI will be forced into the role of tool, not artist replacement." - possibly, but whose tool. If they are a tool for producing cultural products for a mass audience at scale, or for replacing concept artists and asset creators in the videogame industry (as is anecdotally already the case), they will very much replace current artists.
"AI must win consistently more points than the best humans, who already operate on razor-slim margins of originality, insight, or creativity to begin with (much like Federer’s mere 4% advantage). " - possibly but "AI" can also play much more than human creators. If human creators produce 1 artefact in 10 that speaks to audiences and "AI" produces 1 in 1000 then the latter draws equal if it produces 100 times as much as the human, which is easily doable. AlphaGo played billions of matches against itself, something that no human, no matter how smart, creative, talented, could ever achieve.
"Cars still passed on the highway. Here in New England, the leaves will change into orange and yellow and then die and come again. An ant makes its way across a playground structure and will trace the same path tomorrow. The magnetic poles of Earth remain unflipped for the last 780,000 years. " - of course, the material world, and even the consumerist society won't actually change if non-derivative art disappears...but that's not the impact that's claimed by defenders of this kind of art.
Well said Albrecht, I felt the same. Especially about the control of distribution point. Big Tech is solidifying universal control of human expression, be it art, communication, or influence, and that I find the most worrying.
Great, thought-provoking article. Same re comments. As I read, I keep asking a question of this article, a question that is simultaneously unoriginal (the other commenters make the point) and original (no one asked the question this way): where does a unique mind like Emily Dickinson fit into this analysis? She was part of the communal mind (and thus unoriginal) and entirely apart from it. And from there I can extrapolate to other great minds, and from them to ordinary minds occasionally having their own unique thoughts. There is mystery in creation, whether you look at the Big Bang or childbirth or Bob Dylan writing a song. Can AI experience mystery, can it be comfortable with uncertainty, can it imagine its way into that tiny percentage of new human thinking? I suspect (and hope) that AI will always be chasing the unattainable horizon of human imagination. (Of course, when it gets tired of the chase it will just wipe us out, but that’s another story.)
I suspect your reply to Chiang may be missing the point a bit. Strawson's essay, "Freedom and Resentment" really catalyzed for me (and later Wittgenstein helped) a particular way of thinking about how we treat other people. Strawson was talking about how we can take up one of two stances vis-a-vis another human being -- I may be getting the labels wrong here -- the 'objective' versus the 'personal' stance. The former is roughly the stance of science, which explains human behavior in terms of cause and effect. The latter is the stance crucial to relationships, and (Strawson's focus) taking and holding one another responsible for our conduct. So, for example, you might explain Sam's outburst this afternoon by pointing to their hunger and its influence on their mood and how this precipitated the outburst. Or you might explain Sam's outburst in terms of his taking himself to have been slighted. The former is the world of causes and effects; the latter, of intentionality and reasons. It isn't exactly that one is right and the other wrong, nor even that only one of these stances is the ethically appropriate stance (I think both a more 'objective' and 'personal' approach are, at times, required in caring for others). But what has proven intractably hard, so far anyway, is reducing one (inevitably the personal) to the other.
Anyway, I suspect that christening a work as 'good art', 'original', 'creative', etc. attributes it to a *person*, as their (non-moral but still ethically loaded) *responsibility*. And we just haven't incorporated AI into these interpersonal practices yet -- they remain *things* to be explained in terms of cause and effect -- and so an AI can't make art in the same way that the color blue can't be (literally) 100 degrees Celsius. It's a category mistake. Might we one day? Well, maybe. I think in a vaguely Wittgensteinian way that we don't have cut-and-dried rules for inclusion here -- witness our changing and contested circles of concern over human history!
But so you may well be able to point to the products of AI and the products of people and say, look, we cannot identify any significant causal differences -- are people 'really' not copying like AI does? -- but that does not necessarily mean that you should be agnostic as to whether one is an original/creative/good work of art and the other not, or that you should presume they both are. Those issues turn crucially on whether one is the work of a person, and the other the output of a thing.
Does this make any sense? Is this helping? :D
If I interpret you correctly, I think we actually probably agree on this. I still stand by everything I said in the original AI Art is an Art essay, that the philosophical definition of art for me is a communication across consciousnesses. There's no subjective experience behind whatever the AI is outputting, so I don't consider it to be art. But there are also other definitions of art. Some of those focus more on the effects of the viewer, and clearly you can still be emotionally affected by some piece of AI-generated art. So I think technical ability must come into play, and on that I think the opposite of Chiang: most stuff is deeply unoriginal, most choices are deeply unoriginal, and really good artists/writers/thinkers are just *slightly* more original.
I think we're sort of on the same page. I don't like your first definition even if it sounds like I do, because I'm again more Wittgensteinian about 'subjective experience' and 'consciousness' than most people nowadays are. (Relatedly, regarding 'communication across consciousnesses': what would the criteria for success even look like?) My starting point is that there are communities with conceptual toolkits for engaging in and developing the practices of those communities, and this is the context in which things, including works of art, 'light up with meaning', and in which they relate to one another as people. We haven't sorted out how AI will fit into our practices, but they definitely do not yet enjoy the status of persons/fellow practitioners.
Great article, Erik, thanks! It's easy to argue, as Ted Chiang did, that the key feature of great art is originality. But this ignores two aspects of creativity that I believe are more important: surprise and craft.
Great art can open our eyes to a new perspective on existing art or the real world. So originality might actually be more about surprise than uniqueness. This is why "original ideas" never seem particularly original once you start digging a bit. Surprise brings delight and activates fundamental emotional reactions. Originality might bring delight, but you could fill a screen with a number to the ten-thousandth decimal that has never been seen before and that would be original without producing the emotions of surprise and delight.
There's a reason that the word "arts" is frequently accompanied by "crafts". The perception of originality in art can also be provoked by a uniquely personal method of execution on the part of the artist. This craft-driven level of artistry is something that AI de-emphasizes. Sure, you can produce a draft of a short story by inputting an idea into an LLM, but if you want it to be "artistic" you're going to have to take that lumpen draft and rewrite it into a shimmering gem using your own ability as a writer. No amount of re-prompting will achieve that outcome.
So my diagnosis of AI's ability to produce great, or even good, art, is that it's not necessarily originality that matters, but the human input that produces surprise and exhibits craft.
You cited "Pierre Menrad, Autor del Quijote", instant upvote. But the art of Pierre Menrad lies in writing the Quijote as himself, and just by Pierre Menrad having existed, the existance of the Quijote's changes. First, because you'll never know wether you hold an original or a Menrads copy. His endeavour is far more profund than a mere copy, could you imagine if you yourself had written 'La Odisea'? The writing changes, it's not longer is a mythological recount of a possible event, but pure fantasy. As Menrad did, GPT in it's willingess to summon the Hyperspace that it represents into an answer that a human Made, too shaped the perception of text written by other humans. The same effect Menrad had on all other copies of El Quijote, GPT has over all text written post birth.
I hope that I can create such endearing art as Menrad, unfortunately not everyone is a Menrad. Is quite hard to be Pierre. The transformation that he made is not a mere blog, or a novel or a book: Is a change of perception to everyone who willl ever read the art he transformed. On a superficial level the shortstory is about transforming something through your own point of view. But Borges never is superficial.
I don’t see why this mythical “true originalism” should be north star of discourse. Many of the best ideas are actually knowing *when* and *how* to combine existing ideas, especially ideas from far flung places or disciplines. When an engineer takes inspiration from the paws of a little known animal in the Amazon rainforest, or when a linguist takes an idea he heard about in an obscure physics course and adapts it to his own craft - none of these are original thoughts, but they’re incredibly valuable.
Forget this “originalism”, I for one welcome the chance to see how AI can recombine existing ideas in a myriad of incredible ways we mere mortals would never dream of.