< Wstecz

Artificial intelligence as authenticity tester or great art through the eye of the machine. AI will not replace humans, at least for the time being

fot. PAP/DPA/YURI KOCHETKOV
fot. PAP/DPA/YURI KOCHETKOV

If Philip K. Dick's androids could dream of electric sheep, why shouldn't artificial intelligence (AI) dream of rococo pastoral scenes? In 2020, scientists at the Los Alamos Laboratory, USA, proved that some machines can not only fall asleep, but should take a power nap. They need a break in order to absorb knowledge more efficiently. But the knowledge can vary. AI used to verify the authenticity of paintings has already had its first successes, but also failures. And where the machine fails, humans step in.

 


A post published on social media by the Royal Castle in Warsaw encouraged internet users to submit ideas for the use of AI in the castle museum’s mission. AI-generated images were added to the post. They were based on Rembrandt's painting ‘Scholar at his Writing Table’, which is owned by the Royal Castle. The scholar in several AI variations was surrounded by new characters and props. The portrait of the Jewish scholar (which is indicated, among other things, by elements of clothing characteristic of 17th century Netherlands) is usually mentioned together with another Rembrandt painting from the Polish collection: ‘The Girl in a Picture Frame’. To this day, it is unclear whether the portrait's heroine really existed or was a figment of the painter's imagination. Her wry smile, which has been compared to that of Gioconda, her suggestively painted hand "reaching out of the frame", as well as the buff sleeves and characteristic hat, make her stand out from the group of other castle ladies. One theory concerning the painting is that she was the scholar’s daughter.


In 1777, both works were part of the King Stanisław August Poniatowski collection and, after peripeties, were handed over to the Museum of the Royal Castle as part of the Lanckoroński’s family collection. The paintings were the subject of research by art historians, including Rembrandt experts of the Amsterdam-based Rembrandt Research Project (RRP), an initiative of the Netherlands Organisation for Scientific Research. It is a research team of chemists, conservators and art historians active since 1968, who set themselves the task of examining Rembrandt's paintings using modern techniques and verifying hitherto accepted attributions. The verification of the authenticity of the 'Warsaw Rembrandts' took place in three stages and ended in 2006 with the final recognition as works by the great Dutchman. But how is the authenticity of a given painting tested? 


"Establishing the authenticity of a work of art is a sophisticated process. First, there is visual identification, comparisons with other works by the particular master, ones that are unquestionably considered to be his works. Knowledge of the stages of a given artist's oeuvre is needed, and the no less important direct knowledge of as many of his original works as possible. In addition, literature on the subject and documents relating to the artist and his circle are researched to obtain any information that could confirm or deny the attribution of the work to the artist. Finally, chemical tests of the paint layers and the ground, as well as photographs taken in different light spectrums that help us to look "inside" the painting," Alicja Jakubowska, curator of the Royal Castle in Warsaw, explained.


The Rembrandt Research Project team, which certified the authenticity of Rembrandt's works from the Warsaw castle's collection, has a track record of exposing alleged 'Rembrandts'. The artist's works have been compiled and catalogued in three primary publications: Wilhelm von Bode and Cornelis Hofstede de Groot (1896-1905), Abraham Bredius (1935) and Horst Gerson (1968). In Bredius' catalogue, the number of authentic paintings was estimated at 611, in Gerson's there were 411. Today, around 350 paintings are considered authentic. The results of the RRP's works were published successively in six volumes titled ‘A Corpus of Rembrandt Paintings’. These publications caused a lot of excitement, as they denied the authenticity of many famous paintings previously considered as indisputable works by Rembrandt, such as the painting ‘The Man with the Golden Helmet’ from the collection of the Gemäldegalerie in Berlin. Meanwhile, this painting, which is the pride of the Berlin’s museum, is most likely the work of Johann Ulrich Mayr, an Augsburg portraitist and pupil of Rembrandt. Experts pointed out that the helmet depicted in the portrait must have been made by an Augsburg armourer, which was considered as one of the links leading to Mayr. The painting is undoubtedly of a high quality, but was not painted by the famous Dutchman.


In 2005, Stefan Koldehof wrote in the German daily newspaper ‘Die Welt’ that when Rembrandt experts go to a museum, directors panic because the authenticity of paintings owned by the museum may be questioned. Over the years, unfavourable news has reached not only the Berlin museum, but also institutions such as the Metropolitan Museum of Art in New York, the Bührle-Stiftung in Zurich and the National Gallery in Washington. On the one hand, one can look at such verdicts as a loss for the museum, but on the other, many museums, as a result of the surprising verdict, have strengthened departments of assessing the attribution. They also started to x-ray collections, which has contributed to broadening the knowledge of not only their creation but also possible modifications, e.g. when underneath one painting there is another, painted over for reasons known only to the artist. Sometimes, the earlier layer confirms the authenticity of a painting, as it was in the case of some Rembrandt portraits. However, regardless of the method used to examine the attribution of a given work, the success of the whole process depends on the individual and institution. It was said that the deceased (died in August 2021) art historian and head of the Rembrandt Research Project, Ernst van de Wetering, had to take a look at a particular painting to develop doubts about its attribution. And then he would start digging.



In 2016, teams from the Dutch museums Mauritshuis and Rembrandthuis, with support from Microsoft, ING and Delft University of Technology, created a painting in Rembrandt's signature style. The work bore the classic hallmarks of the Dutch master's brush, but was designed by computer and printed in a 3D printer. Technical data of Rembrandt's works—346 digitised paintings—were used. The result of this experiment was a portrait created by an algorithm. The image was supposed to depict "a middle-aged white man with stubble, looking to the right, wearing black clothes with a white collar and a hat". From the data collected, it was possible to create ears, eyes, pupils, mouths, etc. in a style of painting typical for Rembrandt.

One of the experts involved in the project stated that the data collected from all of Rembrandt's paintings were to those involved in the experiment what paints and brushes were to Rembrandt—a material as well as an instrument. And indeed, the portrait created based on these clues was deceptively reminiscent of the Dutch master's original works. It was not just the subject matter, colours, particular features of body parts and composition, but also the texture as the computer programme harnessed to the project even reproduced Rembrandt's brushstrokes. The image generated was printed in a 3D printer using a technique that was capable of reproducing an oil painting. The whole thing impressed viewers much more than 'ordinary images' generated by the algorithms.A separate issue is the assessment of the authenticity of works of art using AI, which has been done for several years now and is gaining popularity. In theory, the antlike work of experts, such as the Rembrandt Research Project, can be reduced to analysis using AI. In theory. 


Will AI replace experts in the art market? "This raises questions, not answers," art dealer, Robert Swaczynski says. "We do not know to what extent AI is developed. Is it able to study the composition of a painting, the time when the canvas was created, its fibres, scan a work of art, use a UV method that uses the electromagnetic and X-ray spectrum? Will it know the origin of pigments and their composition in paintings by Jan Vermeer or Hieronymus Bosch? However, the most important question is– does AI have intuition? Humans seem irreplaceable in the study of the authenticity of works of art. But knowledge and education may not be needed any more as AI certainly learns faster," he added and suggests a test. "Let's put two identical sculptures side by side. For example, by Igor Mitoraj [Polish sculptor-FH]. One original with a certificate signed by the artist and its cast. I'm sure that an expert would immediately pick up the differences. AI may have a problem, as the dimensions, design and material of the sculptures will be identical, but the details of the work, such as patina, may cause difficulties for AI. But let's remember that AI learns quickly. And that depends entirely on humans," Robert Swaczynski concluded.


It is hard to disagree with his opinion if we take into account a recent case with conflicting verdicts from two competing AI algorithms issued in relation to a single work of art. It concerned two Madonnas, the puzzling painting 'de Brécy Tondo' and the debate whether its author was the Renaissance master, Raphael Santi. The painting was purchased in 1981 by the Cheshire businessman George Lester Winward, who thought it was a 19th century copy of Santi’s 'Sistine Madonna'.


In early 2023, an AI model used by the University of Bradford and the University of Nottingham in the UK confirmed a 97 per cent similarity between the ‘Tondo Madonna’ and the ‘Sistine Madonna’. Then, the painting was also examined by experts from the Swiss company Art Recognition, which specialises in the authentication of works of art by using its own AI model. The researchers ruled out the likelihood of the painting being a work by Raphael Santi by 85 per cent. The conflicting verdicts electrified the art world and questioned the reliability of assessing artworks by using AI. On the Art Recognition website, experts addressed the discrepancy between two AI models, the one used by the British researchers and their own. They pointed out that the British universities’ AI focused on facial recognition, as it was trained on facial data and could indicate similarities between paintings depicting the same person, regardless of variants, which way the face in the painting is facing, lighting or the quality of the artwork.


According to the experts at Art Recognition, the programme itself performs well in recognising similar faces, but it was not designed to recognise the authenticity of artworks. Unlike the British model, the one used by Art Recognition has the ability to recognise various artistic elements, i.e. brushstrokes, chromaticism, object placement. The Swiss company emphasised that its team is made up of both art historians and AI specialists, and that the training material on which the machines learn includes authentic and fake works, to help the AI models better distinguish one from the other. Thus, the cause of the conflicting assessments of the Madonna paintings resulted from different assumptions made by the two AI models. Art Recognition emphasised the ongoing need for transparency and accountability in the application of AI in the art world.


However, the dispute over who invented the better model for recognising artworks is not over. On the Bradford University website there is a detailed article about the method and conclusions of the study of Madonna and Child paintings. It stated, among other things, that "the similarity between the Madonnas is 97 per cent and the Infant is 86 per cent". It is worth reading the arguments of both research teams as they may be interesting not only to art connoisseurs. After all, AI models trained to recognise faces are applied not only in the study of art, but also in medicine, where robots analysing a patient's facial expression can determine what condition they are in.


Prepared by: Olga Doleśniak-Harczuk

03.01.24