CODART, Dutch and Flemish art in museums worldwide

The role of Artificial Intelligence in art history is gaining increasing attention, particularly in relation to attributions. Will AI eventually replace the expert? And how can AI be used constructively in our field? Charlotte Rulkens, who is currently developing an attribution consensus method at the Vrije Universiteit Amsterdam, discussed this with Anna Tummers, a former CODART member and newly appointed Professor of Early Modern Art History at Ghent University. Tummers received a European Research Council grant for the ARTDETECT project, which aims to develop tools to help experts more effectively identify forgeries.

First of all, congratulations, Anna, on the grant as well as your appointment to full professor! Can you share your plans for ARTDETECT, and the role that AI will play?

We aim to develop smart, effective ways to detect forgeries by establishing ‘red flags’ or warning signals. Surprisingly, there has been little comprehensive research into forgeries. Therefore, the first phase of the project focuses on an in-depth study of a number of suspected forgeries in both early modern and more modern styles. Additionally, a group of reference paintings is studied, with the primary aim to establish beyond doubt that the case studies are indeed forgeries and to identify their distinguishing characteristics. We will also look at early modern forgeries that were painted while the original artists were still alive.

In the second phase, we will use the initial findings to devise quicker detection methods for forgeries. The analysis of craquelure (the fine crack patterns on paintings) is one of the over-arching themes in the project – and one that will involve the use of AI. This kind of research is already carried out by my new colleagues in Ghent, Professor Maximiliaan Martens and Dr. Aleksandra Pizurica. Craquelure often escapes the expert’s eye, as connoisseurs tend to focus on aspects like invention, brushstrokes, the artist’s expected development, studio practice, and so on. But charting anomalies in the craquelure with the help of AI, may reveal artificial ageing – a clear red flag. This process can help experts swiftly identify possible forgeries.

AI is already being used in the art world. In his article on CODARTfeatures, for instance, Professor Nils Büttner criticized a company that promises rapid, reliable authenticity checks with the help of AI and digital images. What do you think of these kinds of methods and promises?

I agree with his criticism and his concerns about the overly high expectations of AI, especially when companies claim that their AI tool can effectively replace the expert. This can be confusing for buyers who are unfamiliar with the problems and complexities of attributions. AI is certainly advanced software, but advanced software is not necessarily right. It depends on how it is used. There are many relevant underlying questions that such a method might bypass. In other words: you cannot really claim that you have solved something if you did not understand the question in the first place.

If you compare it to healthcare no one would say: We do not need doctors anymore – just use AI! And the self-driving car is not quite there yet either. Who says AI can make the highly complex decisions? That is often the problem in practice. Computers are excellent for processing large quantities of data, but making complex decisions is a different story. I think that we, as art historians, have a duty to temper those high expectations.

In what ways do you think the field can benefit from AI?

We are living in an interesting era, in which both AI and technical research methods advance rapidly. However, if you claim you can make art experts redundant in one fell swoop, you are saying that you can deploy AI immediately for the most complex kinds of interpretation. And that is something that remains challenging and difficult in all fields, not just in art history. If you look at the well-known knowledge pyramid, which starts with data at the base, and progresses through information and knowledge to wisdom at the top, I believe the areas where we have most to gain from AI are at the base of this pyramid.

And what aspects of connoisseurship in these lower levels can benefit from AI?

I think AI can be very useful specifically in areas where numbers and statistical data are relevant, such as canvas and craquelure analysis. These analyses involve large amounts of data and can be said to constitute networks. That is quite different from brushstrokes, which, I should add, have been analyzed using machine learning since the 2000s.¹ However, the eye of an art expert has a well-developed ability to assess brushstrokes, and to do so with far greater nuance. So it is reasonable to question whether AI has much added value in that area.

Do you already have an idea of the type of AI modes you will apply in ARTDETECT?

Not yet. There are many different types of AI that one can deploy, and I have tried out some models – for instance when I was working with Professor Robert Erdmann.² However, we will not start this phase before 2026, and this field is developing so fast that we will have to assess that when the time comes.

One potential concern is that researching forgeries might unintentionally help forgers. How transparent can you be about the applied AI models and the results of the project?

These are relevant questions that we are considering in the run-up to the project. At this stage, we do not have concrete answers yet. We have taken the first step towards developing a data management plan. It is imperative for us to handle our data meticulously, since the kind of information we store can be highly sensitive for the owner, buyer, or seller of an artwork. And if you cooperate with the police, you cannot share anything whilst their investigation is ongoing. Even if you have discovered something or developed an AI tool, it is questionable whether you need to share all the details about how the tool works to be able to apply it in practice. And it is possible that we will gain so much in-depth knowledge that forgers will be virtually incapable of imitating the originals in all their intricacies. An expert committee will be involved in our project to help us deliberate on these issues. We want to advance scientific knowledge, but at the same time, we do not wish to play into the hands of forgers.

You have plenty of data at your disposal to train an AI. But for the vast majority of paintings there is not just too little technical data, but no technical data at all. Won’t that limit the capabilities of the use of AI in your research into forgeries?

Methods of technical analysis have expanded so astronomically in recent years that it is certainly true that far from all paintings have been examined with the latest techniques – not even those by the most famous masters. That means that if you study a suspected forgery alongside two artworks chosen as reference material, you are very likely to confirm your suspicions. For forgers mainly look at the top paint layer and cannot imitate things that were unknown, for instance about the layers beneath.

There is a huge difference between what is possible in theory and what works in practice. The forgeries that have been unmasked in the past ten years have really given us a jolt and forced us to confront the degree and scale of the fraud that is taking place. The big problem is that in practice, is it not feasible to start examining all the paintings that are in circulation using the latest techniques. With ARTDETECT, we are trying to develop shortcuts that can help us make an initial selection.

Can you envisage a future in which attributions are carried out entirely by AI, or do you think experts and the human eye will always have a role to play?

I foresee human qualities and art experts always playing an important role. Initially, people in numerous fields were extremely enthusiastic about AI. You see that changing a little now, in discussions such as those about ChatGPT or driverless cars. I now see a development that is more geared towards finding the best ways to combine AI and human qualities.

A potential downside or even danger of using AI in connoisseurship could be that people no longer go through the intrinsically valuable process of gathering knowledge to reach conclusions. If we start using AI more and more in this field, what kind of connoisseurs will there be in fifty years’ time?

A hundred years ago, before the advent of photography, experts had a huge visual memory. Today, much of that ‘data storage’ is external. However, the human capacity for critical reflection is still essential, even more so perhaps than before. As connoisseurs rely more on tools for data storage and to facilitate comparisons, it frees us up to specialize even more in the top part of the knowledge pyramid. So the experts of the future may well be different from those of today, but not necessarily worse. I think it is true that if the software fails, tomorrow’s experts will suddenly find themselves more incapacitated than their counterparts a century earlier. On the other hand, with the effective use of new technical research methods and digital tools, they will rapidly be able to take far greater strides.

Charlotte Rulkens is a Research Associate at Vrije Universiteit in Amsterdam. She has been an (associate) member of CODART since 2016. Anna Tummers is Full Professor in Early Modern Art History at Ghent University. She was a member of CODART from 2008 to 2020.

Notes

¹ See e.g. Jafarpour, S., Polatkan, G., Brevdo, E., Hughes, S., Brasoveanu, A., & Daubechies, I. (2009). “Stylistic analysis of paintings using wavelets and machine learning.” European Signal Processing Conference, 1220–1224.

² Tummers, A. & Erdmann, R. (2024). Frans Hals or not Frans Hals: Connoisseurship, Technical Analyses and Digital Tools. Cham: Springer.