Introduction
As an outstanding art historian and curator, Max J. Friedländer (1867–1958) also gained a reputation as connoisseur. In 1942, he summarized his knowledge and experience in a booklet that was later translated into other languages. „On Art and Connoisseurship“ („Von Kunst und Kennerschaft“) is still a timeless guide for all those working in this field. The most important method which is applied here is comparison. It is about comparing a work of fine art with other works and relating it to works whose attribution is as secure as possible. This is what connoisseurs did many hundreds of years ago when the term was beginning to become established. The process of comparison on which every connoisseurship is based is naturally prone to error, because any comparison can only be as good as the comparative material used, the selection of which is inevitably subjective. Attitudes towards connoisseurship tend therefore to be critical and the method defamed as “ahistorical, unscientific, even pre-scientific, subjective, unsuitable for theory, not self-critical enough, not satisfying the rules it has set itself and ultimately only serving the interests of dealers and collectors”.1
In 2021, a study using AI concluded that there was a 90% probability that Samson and Delilah (National Gallery, London) was not by Rubens.
Seeking greater objectivity, dealers and collectors in particular looked to the large number of art technological and technical procedures that provide objective results, such as dendrochronology, Raman spectroscopy for determining pigments, isotope analysis or the C-14 method for age determination. The manifold possibilities for making technical images by means of X-rays, infrared reflectography, thermoluminescence and magnetic resonance tomography provide a great deal of information, which, however, also has to be interpreted and, ideally, included in the expert analysis of the technical images and all data. In this way, the specifics of the material used and the images can help date a painting, but in most cases do little to answer the question of who held the brush, as in the case of the Flemish painter Peter Paul Rubens (1577–1640). Rubens produced paintings collaboratively with his workshop, designing or executing them himself and also having his collaborators copy his creations. Rubens himself made no secret of this practice and sold his own works to his clients at twice the price he asked for products from his workshop. Today, the price difference between a workshop copy and a work done by Rubens himself is a thousand to ten thousand times higher; the interest of owners and dealers in authenticating their works is correspondingly great. However, in situations where masters and employees use the same pre-primed panels bought by the dozen and dip their brushes into the same pot of paint, these familiar art-technological investigations do not lead anywhere.
A New Platform
Against this background, authentication by means of artificial intelligence appears to offer salvation. A company providing such a service promises “Art authentication without the uncertainty” – and only “by simply analyzing a photograph”.2 Upon uploading a photograph of a work of art, you will receive a detailed and encrypted AI Report and Certificate – no delays or extra transportation and insurance fees are needed as everything is managed digitally. The company claims to have developed its algorithm using “images of verified authentic works of art”.3 Nevertheless, it does not work with original works of art, but with digital images. So it is not original pictures that are analyzed, but digital ones. Interestingly, knowledge about digital images is relatively limited, especially in circles of art enthusiasts, and even with journalists who indulge in eulogies about the blessings of the new possibilities of image analysis by means of AI.4 The question of how artificial intelligence actually works is not asked.5
Using the same dots-per-inch resolution images for very large paintings, such as The Great Last Judgment (Munich), as well as for smaller works like Agrippina and Germanicus (Washington), results in inconsistent levels of detail to train an AI algorithm on.
A digital image reduces the relief of the three-dimensional surface of an oil painting to a plane, which is then divided into pixels using an algorithm. These pixels are then compared to the pixels of other digital images. Artificial intelligence compares pixels, not works of art. Anyone who enters the world of pixels, the internet, can easily see how much the color values of digital reproductions of famous works of art differ. Often the deviations are so great that the coloristic quality of the original painting cannot be recognized in a digital image. Such color differences become a potential source of error for the algorithm. The next problem is the resolution of the images and the relationship between the pixels of the digital image and the real painting. When using images with an average dpi of 300, for example, the specification “dots per inch” refers to the digital image and not to the painting itself.6 Thereby, the pixels of the digital image have no objective relation to the analogue original. The image of a 30 × 20 cm oil sketch has the same 300 dpi as the image of the Last Judgment from the Munich Pinakothek, which measures 608.5 × 463.5 cm. No question: the comparison is distorting, because while in the first case a pixel has a real relation to the original, the 30 pixels per square centimeter in the best available image of the Last Judgement mean an extremely low resolution. If you were to draw the digital image to approximately its original size, the pixels would appear in the format of bathroom tiles. It goes without question that such a comparison is not useful. If one wants to compare digital images of the same quality with a realistic reference to the object reproduced by them, one accepts deficient images. Or one must work with images that are anything but “high quality“. Alternatively, this problem can be ignored and we compare what should not be compared. Both end in a dilemma and lead to an unreliable result.
Proper Comparison of Digital Images
Apart from the quality of the digital images, another problem is the quantity of works used for comparison. One company states that “normally a dataset of around 200 images is enough to produce accurate results”. There are 1800 painted compositions by Rubens and about 800 oil sketches. The quantity of reference paintings is thus less than 10% of his oeuvre. As already indicated, Rubens’ paintings vary greatly in size, ranging from works as small as 20 × 20 cm to those many meters high. To judge a result produced by AI, it is necessary to know how the selection was made and which paintings were taken. More importantly, AI cannot verify who held the brush when the painting was finished. It can only be determined whether the general characteristics of a digitally reproduced composition, the distribution of color and brightness values, match the technical reference images. Futhermore, the fact that the structure of the layers of paint and the painterly technique of Rubens in the works executed on panel differ greatly from those on canvas is not taken into account. Instead of a streaky imprimitura, the canvas paintings have an opaque grey underpainting that appears completely different not only in reality but also in the digital image. Comparing reproductions of such works with the digital reproduction of a painting executed on a wooden panel exacerbates the problems described above. The promise of salvation propagated by the company offering such services is not fulfilled for the time being. The results generated in the way described are not suitable to justify an attribution or its downgrading.
The Possibilities of Digital Connoisseurship
However, the technology could possibly be applied in the future if one works with standardized images of the highest quality and, for example, incorporates the relief of the surface of a painting by changing the incidence of light. It would also be exciting to train an AI software with the largest possible number of repetitions of one and the same painting. There is no question that digital technologies have the potential to become a useful scientific aid. In the future, AI will be one of many tools with which art historians will work. However, for AI to be useful and impartial it must be programmed by scientists who are not driven by pecuniary interests. The data basis and the algorithm that make up the AI must be disclosed. Not only must the comparative images used be related to reality and to each other, but it must be disclosed what is actually being compared and how exactly the comparison works. In this context, art scholars and journalists must be required to acquire the necessary digital literacy to understand these practices. At the same time, it is clear that even the best AI can only be one of many methods used by connoisseurs today. It is also likely to remain the case for the time being that humans alone will draw the conclusions from the results generated with the most diverse technical possibilities. I am therefore not concerned about the extreme increase in artificial intelligence, as long as enough natural intelligence remains to critically question the coming developments.
Prof. Dr. Nils Büttner is Professor at the Staatliche Akademie der Bildenden Künste Stuttgart and Chairman of the Centrum Rubenianum in Antwerp. He has been an associate member of CODART since 2008.
This feature is the first in a series of articles exploring the rapidly evolving capabilities of artificial intelligence and its potential impact on art museums and the curatorial profession.
1 www.kunstgeschichte-ejournal.net/418/6/Mensger%20-%20Kennerschaft%20heute_20.03.2013.pdf (PDF file, 463 KB)
2 art-recognition.com/about-us/
3 art-recognition.com/our-technology/
4 news.artnet.com/news-pro/rubens-art-recognition-ai-authentication-2017274
5 A recent article explains how so-called vision transformers (ViT) work, i.e. how neural networks function that can be used for image classification and other computer vision tasks: doi.org/10.1007/s00521-023-08864-8
6 The company Art Recognition uses digital images from the picture agency “akg-images”, which makes all data available on its homepage, www.akg-images.de. All images offered have an average resolution of 300 dpi.