Authentication of works of art: »I simply love paintings«

Ms. Popovici, which artists are particularly popular for counterfeiting?

Generally speaking: the most expensive. Simply because they are the way to make the most money on the market. At the forefront are the great impressionists and post-impressionists such as van Gogh, but the old masters also enjoy great popularity.

Why the painters of past centuries?

Old masters such as Rubens have worked with various apprentices throughout their careers. That alone makes the process of creating their works more opaque. This makes it more difficult to distinguish whether a specific picture was painted by the artist himself or one of his apprentices. This fact is also known in the world of art forgery.

What makes a good fake?

Of course, a good fake is one that is not recognized as such by an expert or by technology.

How do art counterfeits even get into circulation?

Admittedly, I am not an expert in this area. But I know there are different strategies. In the past, there have been repeated cases of corruption in which counterfeiters worked directly with experts. However, the private art market, in which works are offered via online platforms, is far more vulnerable. The authentication processes there are much less strict.

Is art forgery a widespread phenomenon?

According to current statistics, there are not so many fakes in circulation on the auction market in Europe and the United States. In other countries, such as Russia or China, things look different. Over time, a wealthy oligarchy has emerged there, whose actors want to adorn themselves with status symbols similar to those of their European friends. There just aren’t that many van Gogh originals in the world, which is why new ones have to be made.

You have a doctorate in particle physics, worked as a financial risk analyst and co-founded the start-up Art Recognition five years ago, which uses artificial intelligence (TO) Works of art are examined for their authenticity. How do these biographical stages fit together?

At first glance they don’t fit together at all. The transition from theoretical physics to the world of finance is not that unusual. In the financial industry there is the area of ​​risk management. To put it simply, a lot of physicists and mathematicians are busy solving problems: by abstracting them and translating them into programming language, i.e. code. From statistics to AI, the tools used in both areas are more or less the same.

And how does art fit into this?

I just love paintings, art has been my big hobby for a long time. At some point I came up with the idea for Art Recognition through conversations with a former neighbor who is an art historian. At the time, I was thinking that working at the bank was great, but it wasn’t that intellectually inspiring either. So I wanted to do something new that combined my time-tested programming skills with my love of art.

How does Art Recognition work?

Before we can check a work, we must have trained our AI with the corresponding artist. Customers upload an image of their work of art to our platform, it is checked using our AI, the probability of authenticity is determined and at the end the client receives a certificate confirming the authenticity of their work of art.

And if the algorithm hasn’t yet been trained with the artist?

Then the very first thing we need is data. In our case, this means that all photos of authentic works by the respective artist that are documented in the catalog raisonné are used. It is important that there is always a contrast set with which the algorithm is trained with negative examples, such as known forgeries, photos of paintings that come from apprentices or AI-generated images in the artist’s style. In this way, the program learns to recognize the image composition or the individual brush strokes of the respective artist and to distinguish originals from fakes or works by apprentices. Depending on the scope, such training usually takes a few days.

What is the advantage over traditional methods for determining the authenticity of an image?

When you have an expert examine an image, they usually want to have the image in front of them. This means that in addition to the expertise, you also have to pay transport, insurance and sometimes even customs fees. This is not only expensive, but also takes a long time. When you take the exam using AI, all these fees are eliminated and you have the result on the table within a week. In addition, the subjectivity factor plays a major role. It is not uncommon for customers to receive two reports from two experts for the same work of art with completely different results. On the other hand, we offer the objectivity of our algorithm and the greatest possible transparency in how it arrives at its results.

Many people worry that when using AI, the human factor will be displaced and decisions will only be made based on technical rationality. They argue exactly the other way around. Why?

AI is now used more as a filter. For example, if the program’s answer is very clear, you don’t necessarily need to consult an expert. However, if the algorithm confirms the authenticity of a work of art by 60 to 70 percent, the involvement of an expert for the assessment is essential. But we neither see ourselves as competition to the experts nor do we claim to do the work better than them. Ideally, human work complements that of AI and ultimately leads to better results for everyone.

What do art historians and experts say about your work?

The reactions are mixed. Five years ago, no one in the industry believed in us. Things are different now. Of course, there are still experts who are afraid of being replaced by AI. But that is also changing dramatically. We have spent a lot of energy in the past explaining how we work to the industry and reassuring that we don’t want to take anyone’s job away. We now also work closely with several experts and art historians.

If you have succeeded in training an algorithm that assigns distinctive brush strokes to their authors, wouldn’t it then be equally possible to train a corresponding AI to produce works of art in the same style over and over again and thus produce almost perfect forgeries?

Clear. With programs like Midjourney or Stable Diffusion, anyone can already have pictures designed in the style of historical artists. This trend is definitely on the rise. In order to be prepared for this and to be able to recognize counterfeits created using generative AI, we have long since trained our algorithm with works of art that were created using such programs.

The interview was created as part of a research trip by the journalists.network association.