<iframe src="//www.googletagmanager.com/ns.html?id=GTM-PWMWG4" height="0" width="0" style="display:none;visibility:hidden">

Can artificial intelligence really identify artistic influence?

19 September 2014

Influence has long been a key idea in art history – hunting for similarities, sources and affinities between artworks is fundamental to making connections between artists and creating narratives. So should art historians be horrified to hear the news that artificial intelligence can now identify incidents of artistic influence, including one never previously noted?

The paper ‘Toward Automated Discovery of Artistic Influence’ by four computer scientists at Rutgers, The State University of New Jersey, reports on algorithms which automate the comparison of images and discovery of influences. The data set contained 1710 high-resolution images of paintings by 66 artists spanning the time period of 1412–1996. As a control it included a ‘ground-truth data set’ of well-established examples of artistic influence: Velázquez and Bacon, Van Gogh and Mirò, which the computer correctly identified. It also discovered previously unidentified similarities between Frédéric Bazille’s Studio 9 Rue de la Condamine (1870) and Norman Rockwell’s Shuffleton’s Barber Shop (1950).

I do not think art historians have too much to worry about as yet, firstly because, despite appearances, the automated process cannot actually tell whether one artist influenced another. It can identify visual affinity, and then using data provided by art historians about the artist and date of the paintings, work out who may have influenced whom. Of course this technology could aid art historians in comparing artworks in order to establish authorship and date, but I very much doubt that connoisseurial art historians who revere the power of the human eye looking at the actual object, as well as other forms of attribution research, will be keen to submit all authority to an algorithm with a high-resolution image.

Art historians are not only required at the start of the process, they are also needed to contextualise the results. This is where the notion of ‘artistic influence’ becomes particularly sticky. Nearly 30 years ago Michael Baxandall argued that an influential artist does not affect their successors, instead the successor appropriates their predecessor, and they are the active party, that is Bacon appropriates Velázquez, rather than being passively influenced by him. Visual affinity only becomes really interesting when you identify motive and opportunity in context.

Moreover, the kind of close visual affinity this process can identify is not always how artistic influence, or rather, appropriation, is expressed. It would be interesting to see whether the algorithm could identify a link between John Deakin’s photograph of Lucien Freud and either Francis Bacon’s Study for a Self-Portrait (1964) or many of the works in Jasper Johns’s Regrets series, currently on display at the Courtauld Gallery.

Griselda Pollock has criticised the computer science paper for its out-dated understanding of art history. Given that this development in artificial intelligence does not appear very valuable, neither for those undertaking attributions nor those exploring larger narratives about art, cultures, societies and histories, this is a timely reminder of the importance of improving awareness of what art historians actually do – not necessarily to discourage attempts to produce digital tools for art history, but in order to make them as useful as possible.

Related Articles

Editor’s Letter: Connoisseurship now (Thomas Marks)

‘The Educated Eye? Connoisseurship Now’ at the Paul Mellon Centre (Katy Barrett)

Curators, connoisseurship and the art of looking (Danielle Thom)

The Art of Digital: Your Paintings, Art Detective and the PCF (Andrew Ellis)