Welcome to episode 273 of The Digital Life, a show about our insights into the future of design and technology. I’m your host Jon Follett, and with me is founder and cohost Dirk Knemeyer.
For our topic this week, we’re going to chat about the next step in art and artificial intelligence. What do we mean by the next step? We’ve talked previously on the show about artificial intelligence creating artwork. We can debate the word creation around artwork for AI, but nonetheless the instantiation of that has begun. We talked about The Next Rembrandt, which created a Rembrandt-like artwork from an algorithm that learned based on the master’s portfolio of work. That was one of the first times we discussed AI in artwork and the creation of that artwork via algorithms.
There’s a piece in the news, Dirk, that I think we both find pretty interesting, which there’s an announcement that the famous art action house, Christie’s, which for listeners that may not be familiar with it, it’s a British company that was founded in 1766. Their New York branch is going to be selling an AI produced work of art for the very first time. That’s happening in late October of this year. The print that they’re selling is on canvas, and it was created by an algorithm that was developed by a French company or collective, an art collective called Obvious. That was, like I said, created through the use of AI and algorithms to generate this artwork.
Dirk, does this feel like a revolutionary moment, when serious art collectors take Christie’s seriously, right? That’s where the high end art scene was established, or at least has been very influential. Does this feel like a moment where AI artwork is perhaps gaining some acceptance?
I think the answer is, “We’ll see.” It could be a milestone. It could be a revolution. It’s probably something more modest. It’s certainly interesting. First of all, we’ll see how the world of money reacts to this. Is it a huge purchase? Is it a modest purchase? What’s the result? That will tell us a lot.
Then what will also tell us a lot is beyond this. This is the first one. Does it, over time, as more come through, does whatever happen in this moment end up being just a mirage, something that was too low or too high, the start of something, a curiosity?
We’re not going to know until those data points are recorded, but it’s certainly interesting.
Yeah, you know I think these all start as a curiosity, right? How could it not be? From that perspective, and you also mentioned this is the world of big money. This is art as, not just collection, but also investment, right? When you think of art movements, you think about types of art that weren’t really considered art when they were produced. There’s aspects of that whether Andy Warhol’s work, which the printmaking, which is famous now, or even going back further, looking at the use of technology, you know, photography initially sort of seemed to be as not necessarily artwork, but sort of a technical and technological feat, but now it’s certainly, fine art photography is considered important and notable.
I like that word curiosity. It’s this initial that’s at the track on this horse. Is it really something? Is it really going to turn into something? If it is, these initial offerings from the French company Obvious, those are going to be big ticket items someday, because they’re groundbreaking right now.
Let’s talk a little bit about how art gets created by algorithms and dig into that a little bit, because I think Obvious has an interesting approach to the creation of their algorithmic artwork. They use a model called generative adversarial network or GAN. They basically feed this AI a data set of thousands of portraits. In this case, they were portraits that were created between the 14th and 20th centuries. That’s the training data set that this algorithm digests and learns from. That enables it to create these pieces that I’m sure you took a look at some of the pieces on the Obvious site there. They’re interesting. They definitely call into that uncanny valley, for me anyway, where it doesn’t seem quite human. We’ll put some links up on the site so listeners can judge for themselves.
But clearly this training set, what I find so interesting about this act of creation is there’s a lot of curation on the human side that enables this to happen. This isn’t an AI that just is generating these portraits without context. The context is being set by human beings who are curating thousands of portraits from a specific time in history to push out this new type of algorithmic portrait, so I’m not sure …
In some sense I get it. Okay, the AI is creating something new, but at the same time, it’s really based on a historical data set that has been established by people and then curated by people. So it’s really more of a hybrid. I don’t know if it’s AI art necessarily, but really sort of a human, AI enhanced human art. What do you think of that take, Dirk?
It certainly is a hybrid, right? That gets into definitions of what is this stuff, and what constitutes art created by AI. Going back at least to the Rembrandt project, which we talked about before on the show, the examples so far are all different levels of that kind of collaboration. The designation of this really is AI or it isn’t I don’t think is particularly useful unless you are somebody wanting to spend a lot of money at Christie’s for the privilege of making some claim about your art, although even the piece in the Christie’s auction, no one is claiming is the first piece of AI generated art. Yes, so the specific designators, I think, aren’t as important, but these are, in fact, human machine collaborations at a level far beyond the human machine collaborations of the human and the brush, for example.
Right. I think right now we’re answering the question, pretty significantly, the idea whether or not AI can be a tool for artists, right? If you look at this as human enhancement, as an evolving tool set, certainly AI can be applied to any number of industries, but in the area of art, in the creative industries like fine art, we’re really going through this process of answering the question, “Is this AI tool set a valid, a useful way of generating new artwork that’s interesting and compelling, and that people can relate to, and that might have value in the longterm?” I think the answer to that is yes.
I will say there’s also this interesting layer over top, which is just all about the human decision making that makes the AI artwork possible. As we reveal more of that, I think it actually makes the artwork seem more human. If you know that, as I mentioned earlier than human beings are curating this, and then even more so, that they’ve created also this AI code, right?
It really is a very active and astonishing and interesting accomplishment that human beings have done that we’ve created this or not we, but that the folks at Obvious have created this digital machine that can generate artworks. That’s an accomplishment in and of itself that’s sort of separate from the artwork, right? Now there’s this mathematical model that generates things that humans can relate to. I find that, maybe that’s a piece of artwork in and of itself.
Yeah, so much of this, and this gets to a lot of things that have to do with AI, is that we don’t know what the human directions programmed into the AI were, and the degree to which or how that influences what the AI is doing. For example, if you look at on the Obvious website, the photos on the homepage, like start there, but even going into the gallery, this software has an art style. These aren’t different images with completely different styles. They’re compositionally classical, but then they are almost Impressionistic to Surrealistic in terms of the style on top of it. That’s consistent through all of the works that they have up there. I think maybe one of the subjects looks more 19th century, as opposed to earlier, but there’s a style here.
Is that something that was the guiding hand of the human, either in terms of the specificity of the style or the fact that the machine has settled on, apparently settled on a style for putting these things together? The unknowns behind what is the machine figuring out for itself versus what are the humans responsible for is another ambiguity. To me, putting all of these things together, it just does come back to being a curiosity more than something important. It may turn out to be important in the longer now, but where we are now, to me, there’s a lot of novelty here, to take nothing away from it. It is impressive, and it is cool, but it’s not entirely clear what it is yet.
Right. I think the creation of the algorithms, that might be useful for us to dig into that a little now. These generative adversarial networks were first introduced in a paper by Ian Goodfellow and some other researchers at the University of Montreal. What I find interesting about the underlying thesis around these generative adversarial networks is that they can be applied to a variety of creative outputs. Obvious has used this for their fine art painting, but these could be used, I imagine, for photos or other types of images, as well as potentially other types of artworks.
The way it’s done, and I’m certainly not an AI researcher, but in layman’s terms, the way it’s handled is, there are two algorithms that are essentially competing with each other. There’s a generative algorithm, which creates these new images based on the training data set. Then there’s an algorithm that tries to categorize that image. It’s discriminatory, so it’s trying to categorize that as not a valid image. So you have the one generator that’s creating the image, and another one that’s saying, “Okay, that’s not really quite right yet. It doesn’t match up with the parameters from our training data.”
When the generator can fool the discriminator, then that output is created, right? So when the generator creates something that’s so close to the training data set that the discriminator can no longer determine whether it was generated by human or by machine, then that is the triumphal moment. That’s when you have this artwork completed.
I thought that competitive … It really is two algorithms, two mathematical models, that are battling it out to make the artwork come to life. That approach, apparently, is quite powerful and has the ability to create stuff that human beings are interested in as well.
Dirk, I don’t know if you’ve heard of this type of modeling before, or are familiar with it, but just from a layman’s perspective, your thoughts.
The specifics of this model aren’t one that I’ve researched, but it is reflective of practices in AI to get to these sort of outcomes.
Especially in the creative field, I feel like this approach has been generating some really interesting outputs. I’m sure we’re going to encounter more and talk about those more on the show as time progresses.
Listeners, remember that while you’re listening to the show, you can follow along with the things that we’re mentioning here in realtime. Just head over to thedigitalife.com. That’s just one L in thedigitalife. Go to the page for this episode. We’ve included links to pretty much everything-