Loss adjusting and art restoration: Will artificial intelligence disrupt or revolutionize the market?

April 13, 2023

A bust of a female head.
Share on LinkedIn Share on Facebook Share on X

Artificial intelligence (AI) technology has proven that it can dramatically change the analysis of artwork as it relates to restoration and loss adjusting. A recent authentication case concerns a painting attributed to the Italian Renaissance painter Raphael, entitled “The de Brécy Tondo (the Madonna of Brecy).” The British art collector George Lester Winward acquired the piece in 1981 and claimed it was painted by Raphael. To prove his theory, he compared the painting to Raphael’s other work, the “Sistine Madonna” that’s housed in a gallery in Dresden, Germany. For 40 years, Winward was unable to confirm his belief, even though in the eyes of many specialists, the similarities were obvious.

Recently, researchers at the British universities of Nottingham and Bradford used a facial recognition tool to compare the two paintings. The computer reviewed the artist’s work magnified by thousands of pixels. The comparison, using AI technology, affirmed that the two Madonnas depicted are 97% similar, and the children in the paintings are 86% similar. One press report on this event used the headline, “Painting’s author identified thanks to artificial intelligence.” But this so-called identification doesn’t automatically result in a certificate of authenticity or wider market recognition of the work.

Indeed, art market professionals, specialized experts, art historians, curators and gallery owners remain cautious of AI — for good reason.

Several elements to consider

When an internationally-renowned expert is known to be the premiere specialist of a particular artist, or artistic period, his or her expertise is valid. For example, in 2019 when the expert Eric Turquin attributed the painting “Judith Beheading Holofernes” to the painter Caravaggio, his attribution had to have been certain. If the expert had doubts, he wouldn’t attribute the work to an author in his certificate and expert report.

Turquin’s analysis went further, comparing it with the other 65 known paintings by the artist. The analyzed canvas and pigments showed it was Neapolitan piece originating from between 1600 and 1610. His expert knowledge and stylistic analysis — in collaboration with a college of art history specialists — allowed complete certainty in confirming the painting’s attribution to Caravaggio. It was demonstrated that the work could not be, for example, a copy by Louis Finson, as some had claimed.

This example demonstrates that:

  • Attributions of important works can’t be determined by a single person. Experts surround themselves with specialists to base and confirm their judgments. A simple analysis to compare two paintings — as mentioned in the earlier example regarding the work of Raphael — is insufficient.
  • Within the art market, it’s difficult to accept that a painting is 97% similar — or an even higher probability than that — to an artist’s other work. Analyses performed by laboratories using AI purport to accept that a 75% similarity is sufficient to validate the attribution of a work. But doubt over the remaining 25% is in no way acceptable for the art market and its insurers.

Authentication certificates qualify the appraised works in different ways. The pieces can be from the “artist’s studio” — meaning produced by the master’s apprentice — or from the “artist’s own hand.” Many artists have their students paint parts of their paintings such as landscapes, draperies, etc.

How will AI take these factors into account? Will it reject attribution of such paintings if compared to a portrait painted exclusively by the artist? Will it also consider restorations, repaints, etc. — and how much importance will that bear in its analysis?

These observations lead us to review the data that feeds AI.

In the “The Madonna of Brecy” example, the indication is that the painting was compared to Raphael’s other painting in Dresden. What other data was used? Were analyses of the materials carried out? Was the dating of the work in relation to Raphael’s stylistic development analyzed?

Another example is the validation of Renoir’s “Portrait of a Woman (Gabrielle Renard)” painting in 2022. Sotheby’s, the company that sold the painting, used AI technology from Swiss-based company Recognition to authenticate the painting. The machine analyzed the brushstrokes, colors and general style of the work to compare it to a database of more than 200 Renoir paintings. The technology established an 80.58% match. Although the auction house was able to attribute the painting in this way, specialists remain skeptical. The art market specialists’ doubt about attribution equates to 19.42%. 

Just as specialized experts call upon laboratories to analyze pigments with support from history experts, AI must be used in conjunction with knowledge of the artist and other techniques to authenticate — or not, as the case may be — a painting.

The responsibility of the expert or of AI?

Imagine a future where AI, with further improvements, is seen as reliable enough for the art market and insurers to rely solely on its verdict. Or, that in a first analysis with a minimum threshold of 75%, it will call into question the initial expertise of the expert who attributed or rejected the work of such or such artist.

It would therefore be necessary to demonstrate that the expert made a mistake. Throughout the history of art, the evolution of technology has revealed that even the most revered experts have made attribution errors. AI will perhaps reveal more mistakes. Still, AI will have to be accompanied by other techniques than those currently used for facial recognition.

An air of caution

AI must not be used to identify works of art in the future without control of the data that feeds it, or by intensive/exclusive use. One can imagine an application that proposes to authenticate a work of art using photographs of it, or even agree to insure it, from analysis carried out by AI.

It’s imperative to retain an air of caution. It’s likely that soon, forgers will use this technology to create imitated works with all the characteristics, qualities and brushstrokes of the original artist. 

Let’s take as example the painting “The Night Watch” (1642) by Rembrandt. The painting had been trimmed down on three sides in 1715 and the trimmed pieces were lost. Thanks to a copy of the painting from the seventeenth century and AI, the painting was recomposed. AI analyzed Rembrandt’s painting technique, his use of colors and his brushstrokes to print the missing parts on canvas. Some art experts are certain that forgers will use this technology to create works that will be difficult to prove are imitations.

Imagine a forger uses AI to imitate a work by painter Pablo Picasso, whose production was immense and whose pigments, canvases and mediums are easily found. The forger’s only difficulty would be to secure a pedigree. History is full of forgers who have succeeded in introducing false works. For now, the art market and its insurers are aware of this danger without being able to fully apprehend it.

AI and restoration of art works

“The Night Watch” example demonstrates that AI will greatly help the restorers of paintings in their approaches to restoration. It will permit a better understanding of missing or damaged sections, and enable confirmation of the colors, shapes, etc., that the artist would have used. Nevertheless, the technicalities and knowledge of restorers will be necessary to produce a perfect restoration.

What artificial intelligence lacks

Though AI can recreate missing parts of paintings and perhaps soon produce works in certain styles, experts still agree that technology won’t possess the level of sensitivity, taste and human intelligence the artist uses to create.