Scientists use AI to indicate authorship of works of art – 02/23/2024 – Tech

Exhibited at the Prado Museum in Madrid, the “Madonna of the Rose” by the Renaissance painter Rafael has generated suspicion for years due to two elements: the face of the biblical character Joseph, husband of Mary, and the flower at the base of the painting. Historians and dilettantes dispute the authorship of the painting.

In 2023, an artificial intelligence (AI) model pointed out that the face of Jesus’ adoptive father was not painted by Rafael. The flower, in turn, would be a work by the artist, according to the AI ​​developed in England.

This prediction has a 98% chance of being correct, according to the model’s creators, a team led by computer scientist Hassan Ugail. “We did cross-validation tests to ensure the results were robust,” says Ugail, who is a professor at Northwestern University.

The art historian Jürg Meyer zur Capellen, author of a critical catalog of Raphael’s work, was the first to dispute the authorship of parts of the painting Madona da Rosa, which was produced between 1518 and 1520, according to carbon dating. He attributed the authorship of different excerpts to his disciple Giulio Romano.

Raphael was so popular in the early 16th century that he was invited to paint several murals in the Vatican. The first of them, the “School of Athens”, made between 1509 and 1510, brought the painter to fame. The last works, from 1513 to 1517, as well as other paintings, were completed by Renaissance apprentices.

Initially, the technology developed under Ugail’s leadership will focus on Rafael’s work, about which there is controversy.

The researcher led a multidisciplinary team with scientists based in the US and England to train the model. In addition to the computer graphics specialist, a physicist, a chemist, a historian and a painter and gallerist participated in the effort.

In the end, this group arrived at an artificial intelligence model that, according to Ugail, considers patterns, brushstrokes, color palette and composition of elements characteristic of Raphael’s art.

In January 2023, the technology, still in its initial phase of development, had already attributed to Rafael the authorship of the painting “de Brécy Tondo”, until then without a recognized origin.

The verdict of artificial intelligence may, however, be just another argument in the detailed investigation necessary to pinpoint the authorship of a work.

In the absence of complete proof of the origin of a given canvas, it is necessary to convince the entire visual arts community to reach a consensus.

The painting “Salvator Mundi”, sold in 2017 for US$450 million after attribution of authorship to Leonardo Da Vinci, still has its origin disputed by experts. “Da Vinci would never make this monotonous hand,” said former Louvre consultant Ben Lewis, in an interview with Sheet.

Gallerist Thiago Gomide, from Gomide&CO, states that the most objective way to authenticate a painting is to survey all the places the painting passed through until it reached the artist’s studio. This becomes more difficult the older the work.

Visual artist and professor at the Faculty of Architecture and Urbanism at USP Giselle Beiguelman remembers that productions prior to the establishment of copyright, during the 18th century in England, often did not take the same care in attributing authorship as recent productions.

For Beiguelman, the use of artificial intelligence serves as a useful, but limited, resource.

“The AI ​​model can capture elements that cannot be detected by the human eye,” says the professor. “On the other hand, a work by the same author can gain different characteristics over the years depending on how it is preserved”, she adds. Paints, for example, can suffer different degrees of oxidation.

Therefore, a very robust database would be needed to account for these nuances. This is often not the reality in the body of works from the Renaissance or Baroque period, when the production of a single canvas could take years.

Rafael’s catalog used to train the English AI model, for example, had one hundred frames. In the field of deep learning with artificial intelligence, it is common to use databases with millions or billions of data.

Researchers led by Ugail tried to overcome this limitation with a technique called data augmentation, in which each piece of information is sliced ​​and then shuffled to generate new data.

Still, considering the margin of error of the artificial intelligence model’s conclusion, there is a possibility that the machine’s verdict could be inaccurate or incorrect. This is where human participation comes in.

There are physical and chemical characteristics that AI may not detect, such as the pigment present in the frame. “It may come from paint not used in a certain historical period”, says gallerist Gomide.

“The authenticator observes the wood of the screen’s chassis, the fabric, whether it was linen or cotton with a thick or fine weave”, explains the specialist.

In addition, there is historical knowledge of when an assistant arrived and left the studio. “A disciple’s work in one setting can influence the entire work.”

Another common technique is to evaluate the composition of the elements and anatomical characteristics of the characters. “An apple painted by Cézanne has its peculiarities”, says the gallerist.

In this process, human work resembles that of the machine.

Beiguelman, from FAU, remembers the Italian art historian Carlo Ginzburg who evaluated the anatomy of ears depicted on canvas, inspired by the fictional detective Sherlock Holmes.

“Many have begun to understand the value of combining scientific analysis using AI tools with historical and stylistic expertise,” says Northwestern University’s Ugail.

For the English professor, the legitimacy of AI in this field will grow as its accuracy and usefulness are proven in successful case studies.

Gomide claims that 99% of fakes are gross. “Technology can make a difference in the remaining 1%.”