Episode 80 - Art and AI with Raphaël Millière Transcript

David: 0:12

Hello and welcome to Overthink.

Ellie: 0:14

The podcast where two philosophy professors talk about how our discipline has lots of rich tools for helping us think about contemporary issues.

David: 0:23

I am your co-host,David Peña-Guzmán.

Ellie: 0:25

And I'm co-host Ellie Anderson.We have been getting so many listeners asking us if and when we are gonna do an episode on AI.And I will confess, I have found this subject quite daunting, which is partly why we're bringing in a wonderful guest a bit later today to help us think through this topic.But I actually went to a conference last weekend called Art and AI at Cal State Los Angeles, and I wanna share some thoughts from it with you, David, to get started.

David: 0:56

You I love conference gossip, so tell me all about it.

Ellie: 1:00

So much of this will be a report,but I actually do have a fun piece of gossip for you and our listeners, which is guess who was at this conference?

David: 1:09

Elon Musk I have no idea

Ellie: 1:12

No,That, would be extremely juicy,but less juicy in a personal way.Okay.If you remember from the emotional labor episode how I talked about this guy that I dated, who dumped me and then started dating an influencer who wrote a song about him,

David: 1:27

yes, I do remember

Ellie: 1:28

she was at this conference.

David: 1:30

Oh, is she an AI influencer or an art influencer?

Ellie: 1:33

No, and calling her an influencer isn't actually fair.I think she like has a real job,but she also did have a presence online after we stopped dating and then wrote the song about him.Anyway, she was there.

David: 1:42

Yeah, but why was she there?Oh, cuz she's a musician and so I guess art and AI, art and music.

Ellie: 1:47

don't know why she was there.I don't know her.I just saw her.

David: 1:50

I know, but now I wanna know what her deal is because she was there and you were there.So this guy clearly has a type,which is like women who go to AI and art conferences in LA.

Ellie: 2:02

That's a pretty badass type, if I'm being honest.Like I, I think I, and she does seem really cool, so yeah,

David: 2:08

Oh, Did you talk to

Ellie: 2:10

No.I'm saying like she just seems cool based on her social media,

David: 2:13

I know, but you wouldn't have approached her and hey,we have something in common.

Ellie: 2:18

Not randomly at a professional conference.Absolutely not.So moving on from conference gossip to what I actually learned at this conference, one of the speakers was Claire Evans, who is in the band YACHT and also is a really wonderful journalist and YACHT is this great band that I've actually seen before in Atlanta, maybe like 10 years ago.And they co-wrote a Grammy nominated album called Chain Tripping a few years ago with AI.

David: 2:48

Oh, nice.

Ellie: 2:50

Yeah, and one of the interesting things about what Evans said is that it was actually harder to create the album with AI, lest you think that co-writing an album with AI is being done for some sort of artistic shortcut.She said they had so many moments while they were doing this album where they were like, are we really doing this?This is so much harder than actually just writing an album ourselves, which they have done many other times in the past.

David: 3:16

That is really surprising to me because I would've assumed that the AI would have been a shortcut.So I'm curious about why it didn't make things easier for the creation of music in the same way that it seems to be making things easier in every other regard.

Ellie: 3:32

One thing she said about this was that the music that AI generated did not lend itself easily to performance.Like it was really awkward to perform the music that AI had written because it wasn't rooted in the rhythms of the body.So she speculated that when humans are writing music, it's emerging out of their own embodied rhythms.And so it feels, in a sense,natural to perform it, especially when you're the one writing it.And that was just not the case with AI.They had to make the rhythms work for them.

David: 4:03

Yeah, and insofar as AI doesn't have a body, it wouldn't have the kind of bodily rhythms that maybe guide our own approach to musical composition.So I can just imagine the kind of lyrics or the melodies even that AI produces as being very foreign to us.By virtue of, quite literally being produced by a disembodied intelligence that exists just somewhere out there in the computational ether of the world.

Ellie: 4:29

Exactly like they don't necessarily fit in with the rhythms that a human would sing.And the way that they created this album, cuz you're probably wondering like how did this actually take place was basically two main procedures.The first was a generation of the music itself and what they did to create the music was they trained the AI on their existing catalog of songs.I think they had 83 songs or something of that sort.And they used a thing called my,like scientific terminology here is going to be somewhat lacking,

David: 5:04

You're like, computers!

Ellie: 5:06

They used whole thing called latent space interpolation, where you take two different melodies in this case.And the AI creates a melody that is in between those two melodies by first starting with melody A, and then getting to Melody B by way of having all of these melodies in the middle.And so it's moving through all of these latent spaces in the middle, and then that's where the new melody comes from.So more or less, the new melodies that the AI was generating had their roots in the existing catalog of YACHT songs.

David: 5:43

Okay.Okay.

Ellie: 5:44

Lyrics were a little bit different.So for lyrics, what they did was train the AI on all of the music,not only that they've created, but that they grew up listening to.It was a giant catalog of songs and a giant catalog of human created songs, right?So the lyrics of Linda Ronstadt say, are probably gonna be better than an AI generated lyric, but even if an AI generated lyric is decent, it's in part thanks to Linda Ronstadt, that it is decent.So I just wanna note that at first,that there's this human creation that's behind both the music and the lyrics.And in the case of the lyrics, it's not just YACHT's human creation, but it's also that of a lot of other bands as well.Then they had AI produce lyrics from there.So they trained AI on this huge catalog of music.And then they used the lyrics that the AI generated as the lyrics for their album.And they told themselves we can't add anything that wasn't already in the AI generated lyrics.They could subtract and they could paste things together in different ways.Like for instance, the AI did not have a good sense of where the chorus would be versus the verses.And so they needed to decide what would be the chorus and what would be the verse.And so they're doing this arranging themselves and figuring out like what melody to put with what lyrics, and also what instruments to have, play particular parts of the music that they had come up with.But they're working really closely with AI and so it was that collaboration that caused so much trouble for them.It created this album that ended up being nominated for a Grammy.

David: 7:18

Yeah, that's interesting.But so they clearly did some editing to the lyrics and they also did some manipulating some judgment calls actually about what is chorus and what is verse, and you mentioned that this happened a couple of years ago.Do you know what software they used for the generation of the lyrics?

Ellie: 7:35

It was an earlier version of GPT, I think GPT two.

David: 7:39

Okay.Like you, I am not an expert in AI or in computer science, but I wonder whether things would be pretty different today and maybe the difficulty that they experience with AI generating lyrics has less to do with there being something unique about lyrics that makes them recalcitrant to AI generation and more to do with the particular stage in the development of AI that we were in at the time that they were collaborating with this AI.So I suspect that nowadays AI might give you the lyrics fully formed with a clear and pretty good distinction between chorus and verse even.So the musical art that you would get nowadays in 2023 from working with AI would come to you in pretty finalized form.That doesn't need that further tweaking.I suspect that it might be passable from the get-go, which it doesn't seem like it was at the time.

Ellie: 8:41

There was this really interesting question at the conference that got at precisely this thing.The person who asked the question was wondering whether when YACHT created their album there was as the questioner put it, this magical moment of an uncanny valley where the AI's lyrics were still weird, like they didn't quite fit.And if you listen to the YACHT album, you can hear this quite a bit, which is really fun.And I think totally works with YACHTs aesthetic cuz they've always worked with technology in a fun and off-kilter sort of punk way.And this questioner was like, I don't know if you would really get that with GPT 3, for instance, you could type into GPT 3, come up with lyrics that sound like a YACHT song and it would probably be pretty accurate.

David: 9:28

Yeah.It would produce lyrics that would actually win a Grammy this time rather than just be nominated.

Ellie: 9:34

my God, so rude.

David: 9:36

But for people who have seen all the experiments that people have been posting about online, of asking the latest version of Chat GPT to create, whatever crazy concatenation of demands like produce a Nicki Minaj rap talking about Abraham Lincoln it's shockingly successful at mirroring style and being faithful to content.And so I do think we are now past that uncanny valley and entering closer to the territory of the Turing test,where humans cannot tell the difference between lyrics that are generated by a human being and lyrics that are generated by the latest version of GPT.

Ellie: 10:21

I will say though that I do think part of the beauty of the YACHT album is in trying to make those AI generated lyrics work according to human rhythms in a way that does inject human element,that lyrics created in the style of YACHT by Chat GPT probably would not have.Today.We're talking about AI and art.

David: 10:46

How is image generating AI upending the world of art?

Ellie: 10:50

Can we say that artificial intelligence is actually creating art or expressing itself through it?

David: 10:55

And who is really the artist in this case, the algorithm, the coders of the algorithm or the human users who feed instructions into it?Although artificial intelligence may not have co-written a major album before YACHT musicians and other artists have in a sense been outsourcing their creative process to technology for quite some time.David Bowie wrote many of his lyrics using what is known as a Verbasizer,which in short is a computer program that auto generates lyrics.So while many of us think about AI in art as a relatively recent phenomenon, it actually has a longer history than most people realize.In fact, we could say that there has been computer generated art almost for as long as there have been computers.

Ellie: 11:48

Hmm.

David: 11:49

In the late 1960s, a painter from London named Harold Cohen got really interested in the relationship between computers and art and the way in which computer programming could be used to generate new forms of aesthetic creations.And he started playing with something that he called algorithmic art.Art that is the result of algorithmic mediation and manipulation.And what he did in the late 1960s.This started in 1968 and continued well into the seventies, eighties, and nineties.He gave a computer a set of relatively simple rules for creating images.So he gave the computers on parameters about things like lines, geometry, figure versus background, and then had the computer generate images on its own that then a little robot that was attached to the computer, they called that robot a turtle, because it looked like a turtle would print these images on a surface.Kind of like a mobile printer,making the images real.

Ellie: 12:54

so the turtle's like a glorified printer.

David: 12:56

Yeah.It like, not even a glorified,it's like an animalized.

Ellie: 13:00

Okay.

David: 13:00

Yeah like the turtle is a printer that just called a turtle.

Ellie: 13:06

So this is really interesting.And yet it also sounds a bit different from what we have today because what we have today is generative AI and Dall-E,which open AI started same company as Chat GPT is the first widely accessible text to image generator where you can literally type in a text prompt.So I could type in, create an image in the style of Van Gogh of this turtle printer from the 1960s.It was 1960s, right?

David: 13:36

Yeah, late 1960s,

Ellie: 13:37

And then Dall-E would generate something from that, this computer program that Harold Cohen started could not do.Something like that.

David: 13:45

no.In the case of the program that Harold Cohen worked with, which was called Aaron, you would have to actually code for the rules rather than give it instructions in natural language,which is what we do with Dall-E, and that's also what we do with chat GPT.

Ellie: 14:01

And so Dall-E and now more and more programs like it have made widely available, something that only a few tech savvy artists who literally knew how to code were comfortable using or even had access to.

David: 14:14

Yeah, because even beginning from the 1960s all the way to the present moment, the majority of artists who were playing with AI were people trained in coding and knew how to code and felt comfortable around computers.And in fact, the generally agreed upon date when AI generated art really shook up the world of art internationally is quite recent, and that was in 2018.What happened in 2018 is that an art collective named Obvious used this open source algorithm to create a visual work of art.It was a painting called The Portrait of Edmond de Belamy, and this portrait, which was AI generated,became the first artwork by an AI that sold at a major auction house.It sold at Christie's in2018 for over $400,000.

Ellie: 15:08

Oh wow.

David: 15:09

When this happened in 2018,the controversy around this particular piece hinged on the fact that this art collective used an open source algorithm to create this image rather than actually writing their own code, using their own tech savviness.So it seems that in 2018, the assumption was that the creative element in AI art lies in the coding more than in the way in which an average person uses a particular software.So it's about creating the software rather than about just using it to create a great work of art.

Ellie: 15:51

It seems like what you're saying,David, is that the creation of the work of art was not being located in that generative ness, which the AI was doing,but more in the creativity on the human side, which was due to the coding.

David: 16:03

Yes.So people were saying, Obvious is not a real artist, and their painting should not have made that much money because they didn't even write their own code.And so if you wanted to really be an AI artist, you had to be a coder first.

Ellie: 16:16

Yeah.And now it seems like just within the space of a few years, the conversation has really changed because the focus has to do with whether the algorithm itself is artistic or whether the people using it, or artists, rather than on whether coding itself is the art.And with Dall-E, the question is,is Dall-E the one creating the art or is it the human putting the text prompt in, or is it not even art?And part of the reason that this all seems controversial goes back to what I mentioned Claire Evans saying, which is that for her, AI wasn't a shortcut at all with visual arts and the generative models that we have with Dall-E, what seems controversial is that it actually does seem like a shortcut, and this is especially the case if the artists aren't coding, they're just plugging in ideas and then generating the art from there.And this is indeed what I heard at an event a few months ago, back in December, 2022, there was an Atlantic event, a progress summit, that I for some reason got invited to.And I heard this guy who worked for Dreamworks talk about how AI was replacing entire teams of animators because all they needed now were text prompts.Obviously you still need real animators,but for instance, AI allows you to have major shortcuts in terms of thinking how a scene is gonna be composed,what a character might look like.Things that used to take weeks and months, he said would now take three to five seconds, which is really amazing if you're Dreamworks and you're driven by profits, but is obviously very scary if you're one of the animators.

David: 17:50

Yeah, and I think the democratizing effect that technologies like Dall-E and Chat GPT had on thinking about the aesthetics of AI generation, that democratization has meant that the focus has now shifted away from who is actually creating the algorithm to the way in which people are using the algorithm in order to create a particular final product.And there it's not so much about whether or not the use of the AI represents a shortcut, although there are also debates very much alive about that,but about what the appropriate use of language is in order to get the AI to produce a work that you intended as a human user interacting with the AI.Does that make sense?

Ellie: 18:40

Yeah, I feel like I'm not getting it, but that might just be because I'm feeling like I'm not firing on full cylinders today.

David: 18:47

No.So in short, new technologies like Dall-E and Chat GPT have had a democratizing effect where it shifts us away from caring about who created the code into thinking about how human users interact with the code in order to generate art independently of whether it counts or doesn't count as cheating, because it's a shortcut.

Ellie: 19:11

Yes.And that actually did come up at the conference as well in the opening remarks by the philosophy professor who organized the conference.This was Foad Dizadji Bahmani.And he had this really great way of putting it at the beginning of the conference where he said, the question is not, can AI create art?But rather, what does our tendency to shun or accept AI art, say about what we want the arts to do for us?

David: 19:36

Yeah, and I think when you frame the question in that way about what our willingness or unwillingness to accept AI in the context of artistic creativity, tell us about us, that really allows us to understand why there are all these concerns, for example,about using AI as a shortcut, right?Because we tend to think that art is something that requires effort,that requires time, and that should cost the artist something.If you can just snap your fingers and create a work that might win an art competition, many of us might, raise an eyebrow or two beyond this question of the shortcut in the sense of saving time, I think a lot of people also express a worry or a reservation about the use of AI because they see the turn to computers and machines and algorithms.Not just as a shortcut, but really as a complete departure from their conception of art.Because for some people,computers are machines.I mean for computers are machines.No,

Ellie: 20:40

"some people would say,"

David: 20:42

No.What I meant to say is, for some people, because computers are machines, they cannot be creative in the way that is required for art.So maybe they can generate,but they cannot invent.And I think one philosopher of art who I think would take this position if you were alive today, Is RG Collingwood, who is one of the main defenders of what is known as the expressive theory of art,which links art above all to emotion.Now, for Collingwood who lived in the first half of the 20th century,art is the expression of emotion.Usually what drives artists to create art is that they are aware that they are having some strong emotion, but they just don't know what that emotion actually is.It's unconscious and it is unintelligible to them, so they just feel as stirring within and they turn to art to clarify that emotion to themselves and to express it.And so according to him, it is through art that artists express their emotions and become conscious.Of them.Now, if we tried to bring this theory to bear on this contemporary controversy around AI, we would simply say that since AI obviously doesn't have emotions, then on Collingwood's Theory,it couldn't possibly be artistic.It might be productive, but there is no development of feeling and no expansion of consciousness that happens in the case of AI generated art, because Collingwood is very clear that art is never about the emotions of the audience.It's about the emotions of the artist.

Ellie: 22:29

Yeah, which I might disagree with Collingwood on.However, it seems like his criticism,if we're thinking about AI, would only even really apply if we're thinking about the algorithm as the artist.If we're thinking instead about the human prompt writer as the artist,then things look a little bit different because the human user does have emotions when they're typing in.Create a image in the style of Van Gogh that is this printer turtle.Arguably, they can express their emotions by telling the AI what to do.

David: 23:00

Yes, that's a very good point Ellie.But Collingwood also distinguishes between what he calls expression, which is central his to his theory and description.He says that description is what happens when you name an emotion using language that you've already recognized for what it is.So for example, when I say I am angry, I am simply describing an emotion that I've already identified.And in a sense, that's what we do when we use AI.We use words to tell it what to do, make a sad painting about this or that, make an angry nightmarish poem about this or that.The difference there is that expression is not, that expression is actually about the process of coming to discover.The emotion before you actually know what that emotion is and before you can ever name it.And so at the beginning of the process of doing art for Collingwood, we don't know what we will find because we don't know what emotion we are trying to manifest or externalize.And that's why he says in a passage from The Principles of Art, that when the artist begins to create a work of art, they actually cannot tell you if it's going to be, let's say, a comedy or if it's going to be a tragedy.And so in the case of the human prompt writer who is interacting with Chat GPT or Dall-E, the problem is that the human already knows what they're looking for.And in fact, they have to know it in order to give the instructions.And in that sense, Collingwood would say the human is not an artist either.If anything, I think the human would be a lot closer to what Collingwood calls a rhetorician.

Ellie: 24:47

Human artists everywhere now write grants, just saying, I can't tell you whether this is gonna be a comedy or a tragedy yet because I haven't had the requisite emotions yet.Please fund my play and let me see what comes out of it, because I do actually worry that the contemporary situation of the art world automatize even human artists in any case.But can you say briefly, I love this distinction that you're drawing, David.I think your calling woody and analysis is fascinating here.Can you say briefly what you mean by the rhetorician?

David: 25:17

So Collingwood draws a distinction between the artist and the rhetorician.The artist discovers something about their inner life in the course of doing art,so they discover themselves in a way,whereas the rhetorician is somebody who masters the rules of language in order to bring about certain effects in the external world, typically in the audience.

Ellie: 25:39

This is exactly what people are doing with the Dall-E prompts.

David: 25:43

Yes, except that I guess the main difference is that the Rhetorician tries to manipulate an audience in order to extract a certain effect, and the human prompt writer in connection to AI would not be dealing with a human audience.They would be dealing with the algorithm,and they would be manipulating language to bring about an effect in that algorithm.But both of them, the rhetorician and the human using AI are directed towards something external, whereas the artist has to always turn within and try to discover something about their own emotional life.

Ellie: 26:39

David and I are so excited to bring you the following interview in part because we actually recorded it in person with our guest in my office,and so you can watch the video version on YouTube in addition to listening to it here, if you would like, because we recorded it in person in my office, the audio quality might not be quite up to the standards you're used to with Overthink,but we think you'll enjoy it nonetheless.

David: 27:02

Dr.Raphaël Millière is the Robert A.Bird presidential scholar in Society and Neuroscience, in the Center for Science and Society,and a lecturer in the philosophy department at Columbia University.He's the author of a number of publications about phenomenology, the philosophy of self-consciousness, and the ways in which drugs such as psychedelics can alter our experience of our bodies.Welcome.

Raphaël: 27:26

Hi.Thanks for having me.

David: 27:28

Rafael, you recently wrote an article based on your research for Wired magazine entitled AI Art is Challenging the Boundaries of Curation,which is about the relationship between new AI protocols like, Dall-E, for example, and just our understanding of what it means to be an artist who is an artist, so on and so forth.And I will want us to talk about that.But can you tell us first, a little bit about how art works in the world of AI for the uninitiated like us?

Raphaël: 28:02

Of course.So there is first of all have said that AI has made great strides in the past 10 years.Because of this new development called deep learning.So this is this new class of algorithms,computer programs that can learn from a lot of data automatically.So from the bottom up.So you feed them.Data, you can collect a data set and then they can learn by themselves to do certain things, to complete certain tasks.As far as the algorithms that are used for artistic purposes are concerned,these are algorithms that are generative.So they are generating new pieces of text or images or audio or videos, for example.And to do that, they get trained on data and they get trained in such a way that they're trying to reproduce some of the features that are in, in the data sets including to create novel outputs.For example, in the case of image generation algorithms, these are algorithms that typically get trained on a data set composed of images and their captions.So you can scrape these images from the internet.So just.Take the public internet and just grab a bunch of images.And on internet, on the internet, often you will have images that have captions,for example, for accessibility purposes.So what you can do is ask your algorithm, can you try to guess which caption and which image go together?try to find relationships between the text of the captions and the visual features of the images.And if you do that enough times with data sets composed of hundreds of millions, or in some cases billions of images and captions, eventually your algorithm is gonna become very good at fitting together, at putting together the right images with the right captions.And then you can, once your algorithm is trained in that, when becomes very good at that, you can reverse the process.Once it's trained in the phase we call inference, which is when you generate the new things.And for that, you can feed a little bit of text to your algorithm that looks a little bit like a caption that describes a certain image.And then the algorithm will be able to produce a whole new image.That is that the algorithm's best guess as to which kind of image would fit that particular caption.And so that's the generative text to image process.You fit it a prompt, we call it.So a piece of text that describe the desire, output, and then you'll go with the mold, generate a corresponding image.And, this is one of the main categories of algorithms people use for artistic purposes.Because a lot of the people who operate in the realm of AI art are visual artists and they use these new algorithms to generate images and create visual artworks.And thatraises all sorts of fascinating questions, I'm sure.Totally.

Ellie: 30:39

And it also, I think that helps me understand.Or put into perspective something that I guess should have been obvious to me.But that wasn't really, which is why so many of these AI images look like composites.Like I feel like especially with faces, sometimes you'll see these AI generated images that look like sort of bizarre composites of different elements of the face.I certainly think probably what immediately comes up for a lot of people, including myself, is questions about what this means for human artists, like in the labor market,like whose images are getting scraped.Maybe we can come back to that a little bit later, but I just wanted to flag that, that I think is of interest to me.But I'm curious where you see creation fitting in here.Because I think when I hear about the debates about AI art, a lot of them have to do with whether the AI is creating the art or not.And you have, I know thoughts on not only this, but also what the role of the person who's putting in the prompts.Has as is this person an artist in the sense we usually think about it, right?

Raphaël: 31:44

Yeah.One thing that I think is important to, to have in mind is that there is a really deep sense in which these algorithms are not autonomous, right?And that's important when we discuss questions such as, are these algorithms artists themselves, right?So essentially you need a human in the loop in order to produce any of these images, cuz you need a human to input the prompt to the algorithm at the minimum and to generate an image.But generally, the process that leads to the creation of the images that, are, offered as artistic works in, in, in the realm of AI.Art is more complicated.So it's not just putting a text prompt and then clicking a button,but even more generally, even in the most minimal case where you're just giving a description of an image, say portrait, of a cat sitting on a mat.So you're giving a text description is generally generating a picture,and on the basis of that description,And you might think that's very, it's a very lazy process.You're just clicking a button and you don't have any, you don't need any particular artistic skill.But it's interesting because that's the, a very similar debate was going on when photography was invented, right?Yeah.Yeah.And there was a lot of pushback with the artists, publishing the equivalent of op-eds at the time, essay in magazines and saying, this will never be art.at best it's an attempt at archiving reality.But it will never be anything like art, because it's just pressing button.And I think we've moved past that because we, all of us in our pockets these days, we have a miniature camera.I don't think that automatically makes all of us artists, yeah.And I think we can all agree on that.So yeah, the mere ability to press the button is not making us artists.Okay.And in order to be an AI artist or an artist, as I prefer to say, using AI as a tool in their, artistic toolkit and in their process, You'd presumably need more than just access to these algorithms that are increasingly easy to use and you need, in particular,perhaps a, vision, inspiration, talent,and in new kinds of skills that come with crafting specific prompts that can accomplish specific results.

David: 33:46

So let me ask you a question about this, because when I've talked to people about Dall-E and then there was Dall-E too, I actually dunno if there's still open access at the

Raphaël: 33:54

moment or not.Yes.Okay.

David: 33:57

So one of the questions that I have is when somebody puts in a prompt, do you know if.If I put in the same prompt back to back, will it generate the same image?

Raphaël: 34:08

No.So it's stochastic.So there is some degree of indeterminacy in the generation of the images.Okay.And so I'm gonna

David: 34:14

try to push back and put on my hat of the 19th century critic of photography,transposed onto the 20th, 21st century.Because I think one argument that could be made that would redeem photography,but then poses a challenge for AI art is that at least in the case of photography, we can articulate a case for the dependence of the final product on the action of the individual.Yeah.In such a way that it could not have happened without the individual.And the same set of actions by the individual at the same moment would've led to the same result.So that there's a kind of codependence of product and producer.And this stochastic nature of the algorithm throws a little bit of a wrench into maybe that articulation of the relationship between the two.Because it would mean that maybe the creative power doesn't lie, or at least doesn't lie exclusively in the side of the artist.If anything, it might be the algorithm that adds that creative stochastic element.Or if we wanted to push it back a little bit in a different direction, we could say it's not the algorithm, but the coders of the algorithm that are the real artists.And so I'm curious about that.What are your thoughts here?Yeah, so I think essentially

Raphaël: 35:29

that's a question about how much control the artist has over the output.And I think you can raise similar concerns or points about photography, right?So many photographers say, those who are not working with the studio and with carefully a carefully composed scene will perhaps go out in the streets and take snapshots of things they see.They don't have full control over what's happening.And they might be that someone, like maybe a cyclist suddenly, passes through the shot and gets captured, gets, as part of the scene that's maybe a happy accident that leads to a particular photograph that we deem, very interesting.There are many famous examples in the history of photography of these kind of haphazard events that lead to a particular interesting shot.And it's not like this was under the control of the photographer in any robust sense.So I think this element of stochasticityity, you can find it also in photography, even though it's not in the, in on the side of the parameters of the camera, but it's on the side of what's happening in the scene.So there are some respects in which perhaps the AI artist using an image generation algorithm, Might have less control than the photographer because of this stochastic element.There are other aspects in which they might have more control in the sense of this is not coming from the external world in this way.That said, there are also techniques to reduce this element.For example, the come from the fact that when you generate an image,the model of the algorithm is using what's known as a random seed to generate any particular image.You can actually make the image reproducible if you use a determinate seed.So a specific number that is seeding the initial noise added to produce the image.Without getting into the technical details, it is possible to make it deterministic.So keeping the pharmac fixed such that the very same prompt with the very same model.We'll generate some image.Yeah.I see.Yeah.

Ellie: 37:18

Yeah.So it seems fair to say that on your view, you don't think that the algorithm is an artist.And it also seems fair to say that you think the prompter may or may not be an artist in the same way that I may or may not be an artist when I'm taking a, a photo with my camera.Do you think that in the future that might change and we might get to a point at which artificial intelligence is becoming an artist?

Raphaël: 37:45

I think that would take a very drastic change in the way these algorithms work.Okay.So one thing I would like to flag perhaps,is that, sometimes people ask me, do you think these algorithms are creative?And I think there are two different ways in which we could pass that question, right?So there is a more deflationary sense of creativity or or originality in which you might say, yes, these algorithms are creative in the sense that, and that goes back to your other comments about the composite nature of these images.That could be a little bit misleading because sometimes people think that all these algorithms are doing is just taking some of the samples from the training data.So for example, artworks have been created by humans and doing a kind of collage, just taking ahead from there and maybe like a body from there and putting them together and plagiarizing different artworks from the training data and putting them together.What they're actually doing is more sophisticated.So in the limits you will have some edge cases in which indeed they might produce images that look very similar to images from the training data, but most of the time they're capable of creating genuinely novel outputs that,don't really look like any of the specific outputs they've been trained on.So they're capable of generalizing beyond, the specific distribution of the training sample.And the same way text generation,algorithm language models or chatbots like Chat GPT can create totally novel sentences that you've never seen before.Now, when we talk about creativity in humans, we generally assume that it comes with something more, which is,intentions to realize a particular art vision, for example, for an artist, right?So we have intentions, goals, intrinsic goals that, lead an artist to approach an artwork in a certain way, to draw things in a particular way and so on.These algorithms, in my opinion,don't have anything like intrinsic goals in that sense.They don't have any desire,they don't have any vision,artistic vision or otherwise.They don't have any intention in that sense.And so that's why I would be very reluctant to say that any of these algorithms are artists.Maybe in the future you might have more autonomous algorithms that could genuinely form their own intentions to create something generally have their own artistic vision.It seems like it would take a much richer set of psychological capacities than current algorithms have.And essentially it would take something like a more general and human-like form of intelligence that we are still marketing quite far away from.

David: 40:05

So when I think about one of the main challenges that people pose or that people see in the development of AI is precisely that lack of intention that is tied to the myth of the genius and the artist, the one who produces from the depths of their being and generate something new.And the argument typically, especially when it comes to AI, is that because this is all machinic and computational, you just cannot have that.And now I'm going to go in the other direction in this question because.There is a whole tradition of automatism in art.It, it's there in literature of people who have tried to write texts that are freed from any kind of subjective conscious control where you type as fast as you can so that something automatic kicks in.And we recognize that as art that's been replicated in drawing,that's been replicated in painting.And so I'm wondering whether this is a new version of that.And so maybe the real challenge, and this is where I'm heading, maybe the real challenge is not so much that it's automated or machinic or computational,maybe the real challenge in the case of AI art is that it distributes the category of the artist in a way that we haven't really had before.Because with previously automated art,you still can identify the person.Doing that, like the, pull doc,trying to like spread paint on the canvas is still an individual, but here you have multiple individuals.You have the person putting in the prompt, you have all the people who coded the algorithm, and so it's unclear where the locus of production is.

Ellie: 41:47

Can I just add a couple quick points to that?So one is, I think with respect to the myth of the genius already the story is complicated because in some of the main theories of Genius, for instance, Kant's theory of Genius,you already have this idea that the genius is overtaken by something beyond them that is outside of them.So I might even say that they're, it's a bit more complicated, but I would also say, I think this idea of the distribution of art is new with respect to painting and this image generation, but it's not new with respect to other arts.For instance, theater and film are already distributed styles of art.Anyway, that was just like, to jump in, but back to David's question.

Raphaël: 42:27

Yeah, no, that,that's a great question.And there's a lot there.So you're absolutely right, and this is usually what I bring up right after I say that, these machines like intentions that it's true.Many artists, you talk to them and they will tell you actually, I don't, go to my canvas having a particular vision in mind.Yeah.Or, forming in this very procedural way,a certain intention and then realizing it.But it just happens.And a lot of artists working with AI algorithms will themselves be very happy to say things like, I collaborate with the algorithm, or things like that, that perhaps we feel first are a bit more reluctant to say.So it's interesting that, it seems artists have perhaps a less,romanticized view of the artistic process and the and the creative intentions that go into it that sometimes we do.But I would still say, though, to push back on this, that first of all, there's a long tradition of generative art that,even predates the invention of computers where you would have things like an artist putting a canvas in the forest and letting the elements oh, do their work.And then it's a process that is much more impersonal in some sense.And less intention driven than the typical process of artistic creation.And nonetheless, I think it is still art that is made by humans, for humans in a very substantial sense.So there is still, even though there is an intention to not project any particular human intention as it were, it's like a meta intention to put the canvas in the forest and just have the elements work on that canvas and then exhibit it in the museum or, you can think of the ready-mades from Duchamp and so on, right?So even then, I think the human intention is still important even though the specific process of creation with the canvas, for example, or the object that's taken from a shop to put in museum is something that perhaps was not imbued with an artistic intention.And the object of the artwork is also produced for human consumption.I don't really like that term, but like for the enjoyment of other humans.And I think that's still the case with AI algorithms where humans use these algorithms as tools to produce artworks that are distinct for other humans.With respect to your point about the distribution of artistry or artistic authorship as it were.I think it's tricky because you could, I think this point might over generalize to a lot of other things.So you could say why stop there?Why not think that the, in the Fuji film engineers that created your camera should be credited as well with a particular photograph that, that you've taken or,the people who made the brush that you're using the paintbrush create a painting.These are tools.They are new tools, but it seems to me that we ought to be careful not to over generalize descriptions of artistic authorship just because,there are people, engineers involved in creating these algorithms.But again, the algorithm is learning from the data set and these engineers are not themselves by hand making artistic decisions about specific features that the algorithms should represent more than others.

Ellie: 45:33

Yeah.So I think this is a perfect place then to ask the labor question.I think we probably think this should probably be our last question although I'm sure we could talk to you all day, but I think.There's obviously been a lot of concern among artists in terms of the fact that these images are getting scraped off of the internet by the algorithm and ultimately by the engineers.So there's a potential exploitation argument to be made.And so that I think relates to the idea of artists' existing work.So it's a sort of past, or it's a question about how we in the present and future relate to the past,but I also think there's a labor question towards the future, which is how are artists in the future going to interact with these AI and.I was at an event a couple of months ago where somebody was talking.It was like somebody who works for, what's the company that does Dall-E and chat GPT?Open AI.Open AI, yeah.Yeah.One of the engineers for Open AI talking about how Dall-E was already changing animation studios and the way that animators are iterating on different possibilities for scenes.And that seems like it could really open up human creativity in exciting new ways by allowing us like certain shortcuts.So I'm curious like what you think about that sort of exploitation question with respect to scraping existing images from the internet and also whether you're optimistic about the future for artists.

Raphaël: 46:58

Yeah.Thank you.That's that's also very.Thorny question.I think there are legal ramification Yeah.Moral ramifications.But the copyright question from a legal perspective is, hasn't been settled yet.Okay.So there are ongoing lawsuits, and the next few years will be very interesting.But because traditionally scraping data from the internet was considered fair use.Yeah.But also because traditionally the outputs of algorithms was deemed not copyrightable.So there's the question about whether there's a question about the protection of the copyright of artists on which the data is scrapped on.There's also a question of the protection of the copyrights of artists who are make, who are using AI to create artworks, right?Yes.Is very complicated.I think what's interesting to me is that there is this really blurry boundary in the art world between plagiarism, pastiche, and homage, right?So there's a, copyright is a modern western concept and there is a long history of artists riffing off each other, imitating each other, paying homage to each other.A lot of artists famously,Picasso has this whole series of imitating the great masters.We, when we look at that, we think of it as a form of homage and, not something that's morally reprehensible, but there is a fine line and all the artists that work within a certain movement,artistic movement, for example, the Impressionists or the Cubists you could say, there is perhaps one artist that is the true pioneer that really like is the first to create that artistic style, and then all of the artists are just copying that style.But of course we don't really see it that way, or we don't think of this as theft of intellectual property and human artists, they also learn from exposure to other artworks, in a bit like these algorithms learn.We get exposed to a kind of dataset of artworks just by studying in art school and so on and going to museums and we use these inspirations in, in, in creation.And So it's very tricky to exactly pinpoint where, to draw the line.Where does it become plagiarism?Where does it, actually infringes on, on, on copyright, and where is,does it become immoral, for example?That said, my personal view is that I think there should be,a blanket option for artists to at least opt out of scraping.So there is already a mechanism in place.For example, for webpages, if you don't want your webpage to be scraped, there is, you can add a file on your website.That is traditionally named robots.txt.So it's just textile that says to any automatic scraper that scrapes webpages do not scrape.

Ellie: 49:19

We need to figure out how to do this for our website.

David: 49:21

No, it's okay.People can scrape us.That's fine.You can plagiarize us.You can copy us.

Ellie: 49:25

Oh my God.I do not sanction, I do not sanction that.Anyway yeah, so I think we,

Raphaël: 49:30

we probably need some kind of standard format that can be embedded in an image.In the metadata that says do not script or do not use to train an AI model.And then, beyond that with respect to the other part of your question, I do think it's really lowering the barrier to entry to artistic creation in, in, in a way that is reminiscent of how, photography or photo cameras have done so as well.Many people who were not particularly talented with a paintbrush.Were able to express themselves in different way using a camera.And, but it's not just, I would say it's complicated because it's not just lowering the barrier to entry, it's also creating a new kind of skill or skillset set.So being a photographer, it's, again, it's not just being lazy and pressing a button.There's actually a whole new skill set about framing, Composition, exposure,balance and so on that comes into play.And these algorithms similarly, you can use them in a very kind of, just take off the shelf algorithm or use Dall-E press a button.But most of the time that's not gonna lead to something very interesting.It's this generic output that we've seen a million times that's really not very appealing.The AI artists that I found most interesting are people who actually tinker a lot.They have a very complicated process.They sometimes train their own algorithms or they have a, they have a really tremendous skillset set that is highly nontrivial.I think it will transform art a little bit like.Photography,transformed art, but not only transform art by creating new media for artistic creation, but also in return,maybe changing traditional media.So for example, photography has had a really tremendous indirect influence on painting because realistic paintings were not as appealing anymore.You could just use the camera.It's if you're going to do that, what's the point?Yeah.So you might, as well,create a more abstract painting and explore that more abstract realm.And I don't know yet what kind of similar influence AI art will have, but I would venture it will have similar effects on other traditional media.

Ellie: 51:28

Fascinating.Raphael, thank you so much for joining us.It's been an absolute delight hearing you talk about your expertise, and I'm sure that our overthinkers feel the same way.

Raphaël: 51:38

Thank you for having me, thank you.

Ellie: 51:39

Thank you.If you're enjoying, overthink,please consider supporting the podcast by joining our Patreon.We are an independent, self-supporting podcast, a subscriber, you can help us cover our key production costs,gain access to our exclusive digital library of bonus content more.David, this interview was So fun to do.It was great to be talking with Raphael in person, first ever in-person interview,but also definitely gave so much food for thought and helped me, I don't know,think about things in a different light.What were your thoughts on the interview?

David: 52:22

Yeah.The more I think about this connection between AI and art, the more I become convinced that the real issue is the question of whether or not algorithms can be creative.So I'm trying to go back to that older debate.

Ellie: 52:37

Yeah, and the question I thought at the beginning you said that wasn't the question you were interested.Have you changed your mind?Maybe because the

David: 52:45

reason that I think we sometimes end up in these really weird positions around AI and art.Oh no, the art is really the human creating a prompt, or it's really the coder,or it's really the algorithm itself is because of the intervention of computers.And they upset our assumptions about the inherently creative nature of human nature that presumably computers don't have.So I think underneath,really, we are having a debate about the possibly creative potential of man-made tools.And so if we think about it that way the question is whether this can be art in the first place.And can I

Ellie: 53:29

just add to that, that I think underlying that question is a question of whether creativity requires consciousness.Not necessarily like you could have a view of saying that computers are or aren't creative without having a stance on whether or not they're conscious.But I think usually when we talk about this, that is underlying it.It reminds me a little bit of what we talked about in the productivity episode, which was that our capitalist focus on productivity by making us into sort of robots undermines our own potential for creativity.And I think for me, a big part of that is because I think that creativity requires not only consciousness,but a certain relation to your consciousness that is active and free.

David: 54:12

Yes, I think you are right that, whether it's emotion, whether it's imagination or creativity, we think that there is this element that only we have that computer slack that is a necessary condition for artistic

Ellie: 54:24

Which I do, by the way,know from our episode that we did with Hi-Fi Nation,you do not think that AI has, you do not think that AI has or ever will have consciousness.

David: 54:33

Yeah, that's correct.I do not think AI is the right kind of thing.So when thinking about art, my intuition is to close the door on the robots that are coming down the street to the museums and to the auction houses,and to the Biennales To present

Ellie: 54:51

their canvases.

David: 54:52

yet to present their canvases or their

Ellie: 54:54

Please, sir.Show my work.Or their turtle printed drawings.It's the turtle bringing the turtle printed.Drawing.

David: 55:02

Yes, but we also have to keep in mind that definition of art as creativity,emotion appreciation, is actually a relatively recent historical way of understanding art.And so here I'm gonna make a point that I thought about while reading Collingwood's on the principles of art, which is about the history of the concept or of the very term art Collingwood says, if you want to develop a theory of art, you first have to get clear about what you mean by the word art, what it actually refers to.And of course, since the 18th and 19th centuries, we have come to think about art as this object of appreciation that we attend to because of its beauty, and that we can evaluate precisely because it incites in us certain feelings or activates our cognitive capacities in a particular way.But Collingwood says,let's be honest, that's not how people historically understood art.In Greek antiquity art was considered a form of techne.It was just another version, basically, of carpentry where it is about learning a technique that involves the use of tools.So the Greeks didn't recognize a distinction between painting and carpentry, and so for the Greeks, art is just a kind of techne.It's connected to the use of tools.If you move forward historically,and then you arrive to the medieval period, the medievals use the Latin term ars to talk about art.But for them, ars meant mastering certain tools in order to bring about certain effects.For example, think about ars grammatica, where you learn the rules of grammar in order to speak well.Or think about even ars magica like the art of magic, where you learn certain rules to conjure up magically certain effects in the world.Yeah.There, there's

Ellie: 57:01

no conception of the fine arts before the 17 hundreds.

David: 57:04

Yes.And this is Collingwood's point that before art had a connection to first and foremost tools with the Greeks and later to rule manipulation with the medievals.You do get in

Ellie: 57:15

the16 hundreds there's a conception of the beaux arts that emerges,I believe that's in the 16 hundreds, but it really doesn't get codified in academia until the 17 hundreds.

David: 57:24

And so if we think about art in this modern sense of beaux arts, the fine arts, the beautiful arts, then yes,AI doesn't seem to fit very comfortably into the world of art.So what then I thought about in connection to this Collingwood reference to the Greeks and the Medievals, is that maybe contemporary AI generated art is actually sending us back to these pre-modern conceptions of art.It's sending us back to the Greek conception of art as techne because it is about the use of tools,namely machines, namely computers, and arguably it is also sending us back to the medieval conception of art as the incorporation or the internalization of rules to bring about certain effects,because that's precisely what we do when we use Dall-E and Chat GPT.We try to use language in order to manipulate the rules of the algorithm to bring about.Certain effects.And I like how Raphael used the word conjuring to talk about this relationship that human users have to contemporary AI, where you have to get the prompt just right in order to bring about the right effect.So it, it's like magic.And magic in the medieval period was precisely considered an ars for that reason, because you have to learn to manipulate rules in order to change the world.

Ellie: 58:53

And if we acknowledge that a romantic conception of art where it is this like free creation of the genius is not the way that art has always been produced or necessarily should be produced, maybe we should stop worrying about this.I'm not sure where I land on that yet,but I'm leaving with food for thought.

David: 59:12

yeah.And for me, the food for thought is that maybe contemporary AI art is just non-romantic.It's just another kind of art.