Steve and Katie speak with Dr. Carina Popovici, CEO and Founder of Art Recognition, an art and technology startup that uses AI systems to evaluation the authenticity of artworks. They discuss the problems with authentication in the traditional art market and the promise and limitations of AI in solving these problems along with some real-world examples.
Steve Schindler: Hi, I’m Steve Schindler.
Katie Wilson-Milne: I’m Katie Wilson-Milne.
Steve Schindler: Welcome to the Art Law Podcast, a monthly podcast exploring the places where art intersects with and interferes with the law.
Katie Wilson-Milne: The Art Law Podcast is sponsored by the law firm of Schindler Cohen & Hochman LLP, a premier litigation and art law boutique in New York City. Welcome back, Steve.
Steve Schindler: Well, thanks, Katie. It’s great to see you again.
Katie Wilson-Milne: Yes, it’s great to see you.
Steve Schindler: So nice to be in the studio inside with no light. A perfect start for the season.
Katie Wilson-Milne: Yes, right. It’s both the end of summer and 98 degrees outside, so—
Steve Schindler: What could be better?
Katie Wilson-Milne: Yes.
Steve Schindler: But what can be better is that we’re starting our seventh season of the Art Law Podcast.
Katie Wilson-Milne: Yeah, unbelievable, and it’s been so fun and it will continue to be fun this season.
Steve Schindler: Yeah, fun. And we are sitting here with our great producer, Jackie Santos—
Katie Wilson-Milne: Yeah, who makes it all happen.
Steve Schindler: And will make it happen for another season. So, what’s up next?
Katie Wilson-Milne: So the episode our listeners are about to hear is going to be one of our first episodes on AI, specifically on how AI can help with authenticity questions, which is— as we’ve talked about many times on the podcast— a persistent problem in the art world. And I think for the rest our our episodes, our listeners are going to have to tune in and find out. But we plan to talk about copyright, we plan to talk about art theft and misappropriation, all the usual hot topics.
Steve Schindler: All the fun stuff. Alright, well let’s get to the episode.
Katie Wilson-Milne: So we’ve been wanting to talk more about AI meets the art world, and we’re going to do that today. And our guest helping us talk about how AI can help the art world understand works of art in particular is Carina Popovici. She is the co-founder and CEO of Art Recognition, which is an art tech startup based in Zurich. It uses AI to provide art authentication and forgery detection. In addition to being an art lover, Carina is also an experienced computer programmer with experienced developing algorithms with applications in physics and finance. She holds a PhD in theoretical physics and previously worked as a quantitative risk specialist at Credit Suisse. Combining her passion for art and her technical experience, Carina became an entrepreneur and founded Art Recognition in 2019. Today, the company is one of the few startups in the art market that operates at the forefront of AI technology, which is why we’re excited to talk to Carina today. So welcome, Carina.
Carina Popovici: Hi, Katie. Hi, Steve. Thanks for having me.
Katie Wilson-Milne: Carina, if you could just tell us a little bit about Art Recognition and what the company does.
Carina Popovici: Gladly. So, at Art Recognition we use an AI program which we have developed in-house. So we don’t have any dependencies on any external software. So we use this program to carry out art authentication analysis. Our AI is based upon something that’s called a deep convolutional neural network that can learn the main features of an artist from photographs of authentic works of art by that artist. And subsequently it can recognize those learned features on previously unseen artworks.
Katie Wilson-Milne: And so tell us a little bit about that learning process. I mean, our listeners are presumably going to have a range of understanding of this AI, the learning process, how it might work in the art world. So maybe just dumb it down a little bit and give us a— maybe we shouldn’t say that.
Steve Schindler: For us.
Katie Wilson-Milne: For us, yeah.
Steve Schindler: Not for our listeners, but for us.
Katie Wilson-Milne: Not for our listeners. Simplify things a little bit and explain the training process, yeah.
Carina Popovici: Of course. So what we do in practice is whenever we want to train the AI on a new artist, we feed the program with photographs of all known works of art by the chosen artist. And from all these images, the AI learns the main characteristics of that artist. The most important feature that’s being learned is certainly the brushstroke, but the AI also learns other features such as, for example, you know, edge location to distinguish objects from one another.
Also the chromatics, what are the used colors and some other high-level elements of compositions and maybe even additional features that are very typical for that particular artist. And this process, like you said, is called training, so all this learning. And we do that on one of the commercial cloud computing providers, so on AWS or Azure from Microsoft.
And this training can take between one and three days depending on the complexity of that artist. And once the training has been completed, the program, like I said, it compares the learned features with those identified on an image whose authenticity needs to be established. And based on this comparison, the program returns a probability for the authenticity of the new art piece. So basically it tells with which probability the new art piece is by the hand who produced all the images learned during the training.
Steve Schindler: And can I just ask a question about the photographs that you use? Do they have to be of any kind of particular quality, or does it not matter?
Carina Popovici: Well, I mean, they do need to be of a certain quality. So we have a protocol, so they need to be more than 1,500 pixels on the short side. If that number doesn’t tell you much, I can tell you that the modern, you know, so the last generation of iPhones, they can reach that quality.
So I mean, the training images, we buy them. I mean, there are data providers that have themselves, you know, contracts with museums, and we can get hundreds of images from them for a certain artist that we need to train. But from the client perspective, if you would like to submit an image to Art Recognition to have it authenticated, you don’t need a professional photographer.
I mean, you should not have, you know, a lamp behind your head or like a light that creates, you know, shadows. But if you keep your iPhone in like an upfront position, take a photograph, then that’s sufficient for this analysis.
Katie Wilson-Milne: One question I have, Carina, and then we want to hear more about why you started this company and sort of— you know, what Steve and I are really interested in is this problem of authenticity and the politics around it in the art world, which I know you’re also interested in, but what just struck me when you’re talking about feeding these authentic artworks into this system so that it can analyze the artwork in question, how do you know with certainty that everything you’re feeding into the system is authentic? I mean, I think it would be pretty easy to know which ones are fake because they’d be identified as fake. But given that part of the problem is that we suspect that many, if not most major museums around the world themselves have inauthentic artworks that haven’t been identified as such misattributed on their walls, is it possible that you’re feeding in incorrect data into the AI program?
Carina Popovici: I mean, you are right, of course, that this is one difficulty that we have in general. I mean, normally we use the catalogues raisonnés as the golden standard. So whenever we train an artist, we go to the catalogue raisonné first. It is of course possible that later in the future, you know, if say, I don’t know, 20% of the Monet catalogue raisonné turns out to be counterfeits, then this will of course impact the results. I mean, that’s clear. But we do our due diligence. You know, we are very careful in making sure that all the images that we use for training are documented in catalogues raisonnés or maybe in books. We also have an art historian in the team that literally, you know, inspects every image and makes sure that there is a provenance, you know, behind that image or it’s in a reputed museum also having a history behind.
If there is, for example, a painting that’s in a private collection but not in a catalogue raisonné and doesn’t have a provenance, we do not use it. So we are being as strict as we can, but of course there can always be an issue. Also, there are artists such as Modigliani as you know, where there is not just one catalogue raisonné, but several, three or four.
Katie Wilson-Milne: We do know. Yes. Very confusing.
Carina Popovici: Exactly. I mean, that’s very difficult to be able to find a subset of images where there is a consensus, you know, where all the experts agree that there is at least some common number of common images that are acknowledged as being real art, you know, by everyone. But you are right, of course, that there are difficult cases where we struggle a bit with authenticity of the training data.
Steve Schindler: So what you’ve just said is incredible. I mean, you’re basically in three to five days of taking images that are sort of widely available or readily available, you know, not high resolution kinds of images—
Katie Wilson-Milne: Not seeing them in person.
Steve Schindler: —not seeing the works in person that you can then, or your program can, come to a conclusion about the authenticity of a work to some reasonably high probability. That’s incredible. And talk a little bit, if you could, about the percentages of probability. How does that work? So we’ve looked at some of the literature on the company and some of the results. What is it that drives the level of the likelihood that something is authentic or not authentic?
Katie Wilson-Milne: Yeah. Like what would an 85% mean or any percentage. Nothing’s 100%, I assume.
Carina Popovici: No, I think in AI in general, because of the intrinsic nature of this model, there is nothing 100%. I mean, not just this but like any AI model that works on, you know, learning something from training data sets. So that probability is something that we call class probability. So basically it tells— we have two classes in our model. We have the authentic data set, but we also have what we call a contrast data set. So the data set of images that includes known forgeries when available, but also works by followers or imitators, artists who created during the same period with the analyzed artists. We also include digital forgeries actually. So we have developed in the house an AI that can generate art, and we also feed those kind of images into the contrast set. So basically we have what we call a tool balance data set, so containing about the same number of images, and this probability that the client receives at the end tells with which probability it belongs to the authentic dataset or the contrast dataset. So if it’s authentic or not authentic art. That’s what it’s behind it.
Katie Wilson-Milne: So who are your typical clients?
Carina Popovici: So our client base, it still comprises of private collectors seeking to sell artworks in general. So that’s the most typical case, but just having a personal interest, you know, in authenticating their art pieces. So that’s the majority of our clients, but we also collaborate with a much more, you know, diverse range.
I mean, we also have art galleries. We have some private art dealers, some auction houses here in Europe, also family offices, some art brokerage platforms. Since we started in 2019, we have analyzed more than 500 artworks from clients literally spanning the globe. So including Australia and New Zealand, which is quite amazing that they heard about us. And the spectrum of artists is very, very wide too. I mean, spanning from old masters to contemporary art, because the AI obviously doesn’t specialize like on a certain style or on a certain kind of artist. As soon as there is enough training data, then it can analyze any artist.
Why do they come to us? Well, in general, they want to make an informed decision when buying or selling art. So the main reason is related, you know, to establishing the value of an artwork. So protecting the investments. We also work with some foundations, actually, who are willing to preserve an artist’s legacy also for cultural preservation. But what’s important to mention at this point in time is that, so the clients that we have, they don’t use, in general, our AI as a standalone method, but rather in conjunction with expert opinions. So there are actually two scenarios that we are looking at, and personally I find this quite interesting. Part of the private collectors, they use it as a filter. So they send us a photograph, you know, and they get a response like in less than a week. That’s, of course, very convenient, because they don’t have to pay for the transportation, shipping the artwork somewhere in Paris, somewhere else. So they get this result very quickly. And then based on this they decide whether it’s worth, you know, shipping the artwork to an expert in Paris or somewhere else in the world. So this is something that we see very often. The other situation is the reverse. So like when the client has seen several experts actually, and their opinions contradict each other. I mean, as you know, and your listeners know for sure, this is an incredibly frustrating experience, you know, to think that you have like on your hands a very valuable art piece, and you go from one to the other, keep hearing “yes,” hear a “no,” and no motivation whatsoever.
And so in that case they turn to us to just gain some clarity. I mean, we’ve seen many cases when we had to, you know, return a negative result, but they were happy, you know, because they appreciated the fact that we provided them a level of detail that they did not see before and also we did a transparent analysis for them. So all of these concrete findings, they are very much appreciated, I think, by our clients nowadays.
Katie Wilson-Milne: One question, you know, that we’ve talked about on the podcast before and that plagues this authenticity investigation, is that if a work is accepted as by an artist, nobody in the art ecosystem has any incentive or motivation to investigate it, right? In fact, they have every incentive to ignore red flags or concerns, because who will benefit?
The museum doesn’t want it to be fake, the owner doesn’t want it to be fake because that would gut its value. The person selling it doesn’t want it to be fake, because they stand to make a lot of money off of it. The person buying it certainly doesn’t want it to be fake, because they’re about to buy it, you know, and spend a lot of money on it. So one of the problems is like there’s just a lack of opportunity in a natural life of an art transaction or an artwork’s sort of relationship with the public and its owners where somebody is motivated to pay attention and find out. So I can see that your clients, if they were finding a work that had never been seen before, presumed inauthentic by an artist they suspected it was by that they would’ve nothing to lose except paying your fees, right? If a work is already sort of thought of as valuable, there’d be no reason to investigate it, right? So I mean, I’m curious your thoughts on that, and if that’s true with your clients, that they’re sort of always coming from a newly discovered work that has no value yet, but they’re hoping to establish value.
Carina Popovici: Yeah. I mean, like I said, so you are right in principle, but these are different situations of course. I mean, if your starting point is an authentic art piece, or if you’ve got a positive response from some expert, then of course, you don’t want to challenge that. But if you have seen one expert first and he said, “no,” and then you went to another expert and he said, “yes,” then the situation becomes a lot more complicated.
And unfortunately, I mean, this happens quite often as we can see it, and there we can help. But where I agree with you, certainly, is that museums, for example, they don’t have any incentive whatsoever, you know, to uncover their fakes if there are any in their collections. But the private owners, I think there are many different scenarios that are possible. Also, I mean, we did have cases of art pieces that have not been seen by anyone. I mean, they have been bought, for example, by grandparents who passed away and then they’ve been inherited. So there was the first effort to like clean up, so to speak, and find out, you know, what’s real art or not. So we also had these cases. So like I said, I think it’s different.
Steve Schindler: One of the things I’m curious about is at this point, how widely recognized is your technology within the art world? So for example, we talked about catalogues raisonnés and for most of the art world, right, if something is in a respected catalogue, it’s presumed authentic, and it can be sold.
And if it’s not in the catalogue, even though there may be a good story behind it, even though there may be some good provenance, if it’s not in the catalogue, Christie’s won’t auction it, Sotheby’s won’t auction it, and it’s really difficult to sell it and that’s because that opinion trumps all other opinions for that particular artist. And I’m just wondering whether your work can possibly erode some of that in some way.
Carina Popovici: Yes. I mean, there are a few interesting points that you’ve touched upon. First of all, there is the issue of the art experts. Currently, as we all know, they are the dominant power on the art authentication. So Christie’s and Sotheby’s certainly rely on their expertise. However, I think it’s very important to realize that this is a massive problem, because it concentrates an immense amount of power in the hands of a single person. I mean, this is incredible, you know, considering the value that these artworks can have. I think the situation, you know, that we see currently, it does raise legitimate questions about how justified are the decisions made by a single individual and of course, their consequences. We want to support the experts. I mean, we do not want to threaten them, you know, we don’t want to replace them. We do respect their vast knowledge, their expertise, but we believe that both approaches should work together, because only by collaborating we can make progress, we can establish the authenticity of artworks.
Katie Wilson-Milne: Yeah. You’re touching on, I mean, it’s what Steve was raising was just that there’s a monopoly on sort of the market, deciding the market value of certain works based on their attribution in the hands of a very few group of people who are absolutely experts, yet they are human beings who have biases, who make mistakes. And you know, while the world doesn’t fall apart, if they refuse to say something is authentic, you know, that value of that artwork certainly will. So it’s an odd, perplexing situation and I think, as with anyone in a position of power like that, you guard it. And so there’s some politics to it, too, and it can feel impenetrable, you know, to people who aren’t in that inner circle. So this idea of having an unbiased, like there’s no human opinion involved, I guess other than the selection of the data set, which I guess if it’s small enough or the life work of an artist is small enough could be an issue. But the idea of taking human bias and review out of it from an outsider seems like a no-brainer. It’s just that again, like who in this system is truth seeking?
Carina Popovici: Right.
Steve Schindler: Right. And I think just to add onto what you just said, one of the frustrating aspects of dealing with certainly some of the catalogues and some of the authors of the catalogue raisonnés is that the process by which they determine whether something is authentic or not is the opposite of transparent.
Katie Wilson-Milne: We don’t know what it is. Yes.
Steve Schindler: It’s 100% opaque. You give the work in and then it is either accepted or not accepted and you have no idea why. And sometimes it’s just the result of an examination by a single person expressing an opinion and you don’t know how that opinion was reached. But as far as the market is concerned, at least with some artists, that’s it. And so having a process that, you know, that is open and transparent, it certainly seems like a good development.
Katie Wilson-Milne: What I’m quite curious about is how they would criticize your product. What would be the basis to disregard a report that your company produces about an artwork? I’m guessing that they aren’t kind of like running to you to work with you to review the scientific, or sorry, the technical analysis, but why not?
Carina Popovici: I mean, our experience with experts, because we talk to them a lot as you can imagine, so our experience is very mixed in fact, and it does vary quite a bit. We know some experts who are there, you know, since decades belonging to families with longstanding traditions.
And surprisingly some of them did embrace our technology in fact. We work with an expert here in Zurich, and that works quite good. And he sometimes looks at what we analyze and would check whether it corresponds to his own opinions. I mean, we had this case several times, but most of the experts are not that receptive.
What I find quite sad actually about the situation is that some of them dismiss this method without testing it. They could just send a photograph, it would be very easy to do a test run, but they don’t do it. They also don’t make an effort to kind of understand the inner working. So it’s a very straight disapproval. But again, without the motivation, without a conversation, without giving us the chance to explain how it works, at least I think that—
Katie Wilson-Milne: What do you think the reason for that is? I mean, or if they have given any reason for not finding it useful to use your platform or thinking that it will not be accurate. I don’t know.
Carina Popovici: Yeah. I mean, I would guess that most of them they simply feel threatened by this technology. They feel that we might replace them, which is not the case, because AI is not perfect for sure. Just as the experts themself, they are not perfect. Data situation is a big issue, because in order for the AI to learn properly, it needs a minimal amount of images to train on. That would be around 100.
And as you know there are artists such as Vermeer, for example, a very famous example, who produced around 30 paintings. So that’s certainly not sufficient to train the AI. And there are others also. So that’s a case where we cannot help. Also, what’s very important is that the art pieces, so either drawings or paintings, they have to be manifestly made by the hand. So if there is dripping, for example, or other modern techniques, the AI cannot help there either at this point in time. Other issues also related to the data situation, like the old masters, as you know, they have often worked together with their assistants together with a workshop.
So you might have a figure by an old master and the rest completed by his apprentices. That’s very difficult actually to analyze. And in some cases we would just say it would better to consult an expert first. That’s on the AI side, of course. On the expert side, we know, as you mentioned, there is the human bias. I mean, there is just the human error. I think it’s really important, you know, to acknowledge that none of these two methods are perfect and to, you know, simply work together and sit at a table. So that would be my approach.
Katie Wilson-Milne: So just to— because I thought you raised another point we wanted to touch on, which is just the flaws or limitations of the AI training, so there’s certain artists that you just cannot issue or report on, because there’s not enough of a data set for works we know are authentic.
Carina Popovici: Indeed. Indeed. So I mean, it also depends a bit like on the complexity of that artist. So like Cezanne, for example, is a very complex artist. He has several different periods. And in that case the data set needs to be bigger to make sure, you know, that it captures the specificities of each period. But there are artists that are more straightforward, that they have preserved their style across their careers, and then the dataset doesn’t need to be as big. But from experience, we know that it’s necessary to have at least around 100 images just from empirical evidence or trial and error.
But 30, I mean, is definitely not enough to train on AI. And there are several cases, most of the valuable artists did produce enough material to train on, but there are some exceptions, like the one I had mentioned that didn’t. So this is one case where the AI cannot give an answer. Also, there is the issue with multiple catalogues raisonnés.
Steve Schindler: Right.
Carina Popovici: I mean, Modigliani comes in mind, but there are also others. So there are flaws. I mean, there are limitations and we acknowledge them, and we do strive to enhance the capabilities of our AI technology and to support the authentication process and the experts themselves.
Steve Schindler: And I think I understood you to say that one of the limitations was that this technology is really suited for works that were painted by somebody’s hand with brushstrokes, so to speak. I guess just to use an example, say Jackson Pollok. Would that kind of work be susceptible to your product, or is that not a kind of work that you can analyze?
Katie Wilson-Milne: Right. There are a wealth of imitators, you know?
Steve Schindler: Right. No, of course. And I’m just curious if that’s a limitation as well.
Carina Popovici: Yes. It is actually. I mean, Jackson Pollok did paint something by hand. I mean, not that much, but there are some paintings that he produced really by his own hand, so to speak, that we can do. And we did have artworks by Pollok, but if it’s pure dripping, then we don’t do it.
Steve Schindler: I see. And I have just also another question which you touched on earlier and it’s a little fascinating to me. It’s a bit of a circle, but we know that AI is capable of learning and learning what, say, a Renoir or a Cezanne looks like. And then theoretically it can take that same data set and create a Renoir and a Cezanne that might look very convincing, because it has mastered the art of the artist. And how does your program then detect that kind of work if it’s actually generated by the same kind of technology?
Katie Wilson-Milne: Or by itself? Yeah.
Carina Popovici: Yeah. So actually, I mean, we recognized this problem already last year, you know, when the generative models started making headlines. What we do actually quite simply is to teach the AI how a digital forgery, or a synthetic image as we call it, looks like. So you remember I mentioned that we don’t feed just authentic, so images of authentic artworks into the system, but also images of fake forgeries, imitators, followers, and digital forgeries. So during the learning process, the AI gets to see some Renoirs that have been produced by a difference or by a generative AI. And that’s how it’s able to distinguish them. We started doing that last year. So this is like a new model. It’s something that we had introduced in our technology.
We didn’t have it in the beginning. I mean, in 2018 the generative models were quite simple. But meanwhile, I mean, we know it is a trend and unfortunately, we know that also the hardware has advanced a lot. So we are hearing that there are factories in, well, somewhere in the world that use digitally produced forgeries combined with very good 3D printers, for example, and create art pieces that you can buy at auction that looked very, I mean, obviously not at Christie’s, but unfortunately, you know, there are enough people that think that they can get the Renoir, you know, for $15,000 and this happens and we got to analyze some of those imitations. So that’s unfortunately a trend.
Katie Wilson-Milne: That’s actually fascinating that you’re training the AI to detect AI. Seems like it might be a race to something. So we touched on this too when we’re talking about like concerns with authentication, but one of the other concerns is that the act of authenticating, or even when you don’t use that word of commenting in any way on the authorship of a work, can subject someone to liability. I think that’s probably particularly true in the United States where we have a very litigious system and a, you know, business— interference with business torts that can be brought when someone destroys the market value of a work by issuing an opinion. And so there is a fear, a legal kind of form of intimidation or just liability protection that would lead someone not to issue an opinion.
And you know, we would advise them to be very careful not to do that too, or to do it in a certain carefully crafted way. So do you worry about that liability? I don’t know if you know you’re based in Zurich, if that’s a factor, but I know you are not giving an opinion, you yourself, Carina, but your company is issuing a report, presumably that puts a percentage likelihood that affects potentially the market’s perception, the value of a work. So I don’t know. I mean, that’s been one of the major roadblocks in the United States.
Carina Popovici: Yes. I mean, I’m aware of the fact that that’s an issue of course, but what you just mentioned, I think it’s quite important. I mean, I’m clearly not a legal specialist, but the fact that what we return to the clients is a, well a number, you know, coming out of the computer program, it’s not quite the same as issuing an opinion as an expert.So I think actually that it could even be the other way, like the AI as it is, it does have the potential to address this liability concerns by providing all the objectivity, you know, evidence-based to reduce the impact of the human biases and errors, because there is no human behind it. I mean, except the creators of the program and the selection of the training sets. So I think what we need first of all is a robust legal framework. I mean, I know there are discussions, there are initiatives to kind of include, you know, artificial intelligence like in the, well, legal infrastructure, but I don’t know honestly if there is that much progress on that route. So I think before discussing the details, so to speak, of the liability issues, we need the legal framework, and that should of course consider all the issues, the transparency, but also the accountability and allocating the responsibility. So if something goes wrong, who is actually responsible?
So obviously not the AI, but then who? There is a team of developers maybe. So I think there are many, many open questions actually. And also with the trend. I mean, we know what happened to Stable Diffusion, because they got sued and also there are all the other generative models on the internet, but the issues were different there, right?
Katie Wilson-Milne: They’re different.
Steve Schindler: Right.
Carina Popovici: So they got sued, because of the copyright.
Katie Wilson-Milne: Right.
Steve Schindler: Right.
Carina Popovici: So we don’t have that problem actually unfortunately. But I think it’s difficult, yeah.
Steve Schindler: Yeah. I mean, here, as Katie said, we’ve had more lawsuits, you know, where an opinion is issued and that has had some negative effect on a work, whether it’s a catalogue author saying that they’re not going to accept a work or whether it’s just an expert. And it mostly comes up when a work is deemed to be inauthentic, and then whoever it is that owns that work is looking at a potentially big loss and then goes after whoever it is that that issued the opinion, which is why a lot of experts here have stopped authenticating works because of the liability that’s attached. And a lot of the, you know, catalogues now are very careful, at least here, about the contracts that they make people sign before they look at a work
Katie Wilson-Milne: And they’d only work for the owner, for example.
Steve Schindler: Right. And so there is some movement on that, but it is something that I think people here are concerned about. And I would suspect if you were looking at a work, say at an auction for somebody that was valued at a very high number and then your program determined that that work was inauthentic and then that result got out, that would have some pretty significant impacts, I would think, and there might be some consequences for that.
Katie Wilson-Milne: Might be some liability. We very much want there to be more of this activity. And it’s a tragedy in the United States that so many artists, foundations and other organizations that were the experts in, and the only people authenticating a body of work can’t do it anymore. So it’s left a huge vacuum, which to some extent for a certain type of art, your company can provide some of that information, but it’s just like, will you have the same issues? Hopefully not. So tell us how you got into this area to begin with. I mean, it’s not what you studied. I mean, we know you have a lifelong interest in art, but what was the compelling circumstances that led you to try to develop this type of AI and solve, I guess the problem we’ve been talking about?
Carina Popovici: So the inspiration for creating this program arose from discussion with an art historian back in 2018 who made me aware of the issue of authentication in the art market with at least a half of the artworks being counterfeits, maybe even more. And as an art enthusiast and art lover, I saw this as a tremendous opportunity to put my technical skills, which I had acquired during my past career, at the service of art. Because at that time there were no computer programs capable of addressing this challenge. And I remember I started writing the program myself during my spare time while I was still employed at Credit Suisse. And I was quite amazed actually, to realize that it works. And then I eventually decided to leave my job at the bank and to establish the company. I started submitting this idea to some competitions, some innovation prizes, and it got the first prize here and there. I also started talking to some local experts, and eventually I realized that this little idea has a huge, an enormous potential on the art market. And I decided to do it. So that’s how it started.
Katie Wilson-Milne: That’s great. And I recall when we were speaking before, part of what enabled you to start was that you didn’t immediately need to go into a round of private fundraising for the company, that you got some kind of government support. I don’t know if you want to mention that as well. Because as I thought about that, it seemed quite relevant, actually, to create this free bubble of innovation for you to get going.
Carina Popovici: Indeed, indeed. So we partnered up quite early with a research group at the Tilburg University in the Netherlands. We drafted a project together, and we submitted to a funding agency of the European Union and the project got accepted. So it was about developing this technology essentially.
And with the help of that funding, like you’re saying, Katie, we had the freedom to work on the codes on this product without being pressured, like either to get clients, you know, or to get some kind of funding. So we didn’t go look for investors immediately. We were in a comfortable situation in the first year and a half. And that was great, because we could develop it freely and bring it up to a level where we could bring the product on the market.
Katie Wilson-Milne: And now how are you structured? I mean, do you have investors? I mean, you’re presumably—
Carina Popovici: Yes. So we have several investors actually. So we have a VC based in Bern. We also have several private investors, in fact. The company has grown. So we started as two persons. Meanwhile, we are seven people. I don’t work on the program anymore. We have a very strong tech team maintaining the software and staying up to date, improving it of course. We’re a bit more structured as a company. I mean, I would honestly describe this as a success story.
Katie Wilson-Milne: Yeah. That sounds quite successful for a startup venture. Carina, so one thing, you know, that interested us when we were hearing about your company for the first time was reading about a case you worked on— which listeners, we’ll link to the Wall Street Journal article in the show notes about this— a work that may or may not be by Raphael that was discovered by an American, bought it on— this is like every story we tell about works like this, bought it on vacation somewhere and—
Steve Schindler: It’s the Salvator Mundi sort of revisited.
Katie Wilson-Milne: Yeah. And it’s like, we’ve said this before. It always seems to be the case. I need to buy more old art on vacation. But the story is incredibly interesting. And I wonder if you could tell us a little bit about your work with this particular Raphael.
Carina Popovici: Yes. I mean, I can also give you a few background information, because I know the story of course. So it’s been bought in England, so to settle that, I think it was in the late nineties, maybe 1995, I’m not totally sure. And the painting depicts biblical scene. So with Mary, Jesus, Elizabeth, and John the Baptist, I believe. So the person, of course, first believed it was masterpiece painting by Raphael. Then I think he did invest some considerable amounts of time and money into researching it.
Katie Wilson-Milne: I just want to interject and say this person, his name was Anthony—
Steve Schindler: Ayers.
Katie Wilson-Milne: Yeah. Anthony Ayers, who was an American furniture maker, cabinet maker, an amateur artist on vacation with his wife.
Carina Popovici: Right. But so I think what’s important, you know, from the background story is that some expert, including the art historian, Larry Silver, believed the painting to be by Raphael or someone from his circle, whereas others suggested that it could be the work of a less known artist.
And so actually to settle this debate, the owners, because meanwhile the painting is owned by a group of investors, so they have commissioned us— so Art Recognition— to run an AI brushstroke analysis on this painting. And what we did, and this is something that we don’t commonly do on every painting, is to analyze every figure separately. And our AI concluded with a 97% probability, which is really very high, that the faces of Jesus and Mary in the painting were painted by Raphael.
Steve Schindler: Wow.
Carina Popovici: And the rest, presumably by his workshop. I mean, we cannot tell who created the rest of the painting, but considering that it was usual in those times, probably he had the workshop finishing it. And this is actually pointing something of the strength of our technology, the fact that we can indicate very accurately, especially on the old masters, what parts of the painting are by the hand of the presumed artist. I don’t think any expert can do that. And in this case, it turned out to be a major discovery. I mean, we were really absolutely thrilled by the result, because as you know, finding a genuine Raphael— it’s a sensation. I mean, it’s a fantastic discovery, because they’re so rare and of course so valuable, his works.
However, I believe that the overall situation is not that clear. I mean, the acceptance of the painting as a genuine Raphael, it’s still uncertain at this point in time. I mean, we also read the article in the Wall Street Journal, and we know that some leading experts such as Nicholas Penny for example, are willing to use the AI findings to provide their own assessments, so they are positive towards this technology.
Other experts, well I think they expressed some reservations, but emphasizing that AI’s findings should be combined with other factors. I think they talked about the painting’s conditions. So the future will show how this develops, but it’s a very exciting story and for us a major work that we’ve been running an AI analysis on.
Steve Schindler: Is the condition of this work pretty much intact or has the work been worked on, you know, that we saw with the Salvator Mundi and we had actually the person who discovered that work on the podcast, that there was a lot of conservation work that needed to be done on it before it could sort of reveal itself. And is the Raphael in that category or is it just the way it was found at the sale?
Carina Popovici: I mean, I’m not totally sure actually, but there has been perhaps a bit of restoration done on it, but certainly not at the level of the Salvator Mundi. Not at all. Because we’ve been asked, actually, very often if we analyzed that painting, but fact of the matter is that there is about 20% left of the original painting.
Steve Schindler: Right.
Carina Popovici: I mean, it’s been so massively restored that it’s practically a painting by the restorer and we could not analyze that, so we did not do it. But this was certainly not the case here.
Katie Wilson-Milne: I mean, the story so reminds us of the Salvator Mundi story. And from what I remember from the article, Mr. Ayers goes to England on vacation, he finds this work, he buys it for like $3,000 with maybe some other people, through some provenance research realizes it came from a convent in Kentucky that had been gifted this work by European figures in the church that were trying to find artwork to donate to the new frontier of US churches. And that’s how they traced it, sort of, you know, to the US back to Europe, and then did some materials analysis on the work, I believe before you were involved to say, okay, the dating on the wood and the paint does date back to the 1500s in Italy— Florence, I believe.
And so kind of just set up the space for more questions and that okay, it’s possible to now see if it’s by one of these great artists and looked like from initial expertise that if it was by anyone, it would be Raphael, because of the way the faces were painted. So the process over many, many years, maybe two decades, yeah, two decades, was slow and gradual and you came in only a couple of years ago after a lot of that was discovered.
Carina Popovici: Yes, that’s true. That’s true. I think what we did really made a difference. I’m not sure if anyone has thought about like, analyzing in that kind of detail the painting, so to be able to distinguish between here is Raphael, here is not Raphael. Because that wasn’t possible. I mean, if you look at the wood, you would not get that kind of information. Also, the pigments, the colors, I mean they would all be from the same epoch. So I think all the techniques that were available before we came were not able to make that distinction.
And that’s where we really make a difference. I mean, that’s where we were able to help, you know, to say, okay, this is not entirely by Raphael, but these figures.
Katie Wilson-Milne: And how big was your data set? I mean, did you have a robust data set to examine it with the question of whether parts of it were by the hand of Raphael? I assume you did, just give us a sense of sort of size.
Carina Popovici: We did. I mean, unfortunately I don’t remember that. I mean, there was certainly more than 100, but I would need to check the number actually.
Steve Schindler: Right. This actually seems like a perfect scenario for you, right? Because you have a picture with aspects of it, whether it’s the provenance or the materials that are not inconsistent with it being a Raphael or from the school of Raphael, but not enough to sort of put it over the top, so to speak. And so having your work then on top of that gives it a whole new level of seriousness, I would think.
Katie Wilson-Milne: Right. It’s more important than if there’d been a consensus somewhere else. But I guess what we’re seeing here which is what you said, Carina, at the beginning of our discussion today, is that still the experts are wary of these results even though they don’t seem to be specific criticisms about the AI software itself or the learning mechanism.
But someone might hear about this report and say, “well, that settles that.” Like go right to Christie’s or Sotheby’s and try to sell it. And you know presumably the entire point of having all these investors buy this painting and go to you and work hard over these couple of decades is they want to get an opinion that it’s a Raphael and they want to sell it. And no one will sell it for them at a price related to what Raphael will get at market, unless there’s some consensus among experts and to be seen, I guess, whether your report moves that consensus along or not.
Carina Popovici: Yes, I mean, this is something that the time will decide. I mean, this story I think puts everything in the broader context. Like when we started in 2018, I mean, people couldn’t even pronounce AI, you know, in the art world. Things have changed a lot during the past three years.
And you know, with advent of ChatGPT, with having AI all over the place, they are changing at an even faster pace than before. So personally, I believe that there will be a consensus, but it takes time. I mean, we are the forerunner, we were the very first company on the market to offer an AI art authentication service commercially. And of course, it takes time to gain the trust, to become established. But my personal overall impression is that the art owners, I mean the collectors, the investors such as the owners of the Raphael, they welcome the new technology, because it can make the process of course, safer, more transparent, more objective. And if that critical mass of collectors will reach a certain level, the auction houses, the Christie’s and Sotheby’s of this world, they will have to embrace that technology eventually, because they will have no choice. But this process takes time.
But again, I mean, I can see that we’re getting there. You know, we are advancing. Many things have happened as compared to where we started. It might take another year or two but the trend is there, certainly, and it’s irreversible. I mean, AI is here to stay for sure, also in the art market.
Katie Wilson-Milne: I think that’s right. I do think that what may happen is that these traditional experts, whether they’re curators at the world’s major museums, or they work for catalogues raisonnés or artist estates/foundations, whoever they are, like, they may not want to weigh in.
And I think we see this with people at museums. Sometimes they’re just not going to say one way or the other, and it doesn’t really matter what evidence comes across their desks, they’re just not going to opine on it. And that can be a problem if a major auction house doesn’t want to sell it unless they do opine and they just won’t, because they don’t want to get into it. And so what is another thing those market makers can look to, and will they over time start to put enough value on this sort of technical analysis that they’ll feel okay not getting a positive analysis from one of these few experts?
Steve Schindler: Right.
Katie Wilson-Milne: It could be quite significant.
Steve Schindler: I think so. I mean, it also seems to me that to be a little analogous to what we saw with Jamie Martin’s company, right, it’s— an individual who’s been on the podcast and who we know who had a forensic materials analysis company, which was eventually acquired by Sotheby’s. And it eventually came to be that a major auction house saw the value in having—
Katie Wilson-Milne: The irrefutable evidence that he produced.
Steve Schindler: —that technology in-house. And so now they offer that when people come to consign works of, you know, some importance and age, Jamie Martin will look at it. And so that seems to me that what you’re doing is kind of right in line with that, too. And probably inevitably will have to be accepted by auction houses and dealers and the marketplace in general
Katie Wilson-Milne: And maybe come par for the course as you’re saying.
Steve Schindler: Yeah. Alright.
Katie Wilson-Milne: Alright, well, thank you, Carina. This was excellent talking to you, and we’ll watch how the story develops and watch how your company moves forward.
Steve Schindler: Yeah, we’ll keep an eye out for it. Thank you.
Carina Popovici: Thank you very much.
Steve Schindler: And that’s it for today’s podcast. Please subscribe to us wherever you get your podcasts and send us feedback at firstname.lastname@example.org. And if you like what you hear, give us a five-star rating. We are also featuring the original music of Chris Thompson. And finally, we want to thank our fabulous producer, Jackie Santos, for making us sound so good.
Katie Wilson-Milne: Until next time, I’m Katie Wilson-Milne.
Steve Schindler: And I’m Steve Schindler, bringing you the Art Law Podcast, a podcast exploring the places where art intersects with and interferes with the law.
Katie Wilson-Milne: The information provided in this podcast is not intended to be a source of legal advice. You should not consider the information provided to be an invitation for an attorney-client relationship, should not rely on the information as legal advice for any purpose, and should always seek the legal advice of competent counsel in the relevant jurisdiction.
Music by Chris Thompson. Produced by Jackie Santos.