In the past few years, there has been a rise in the application of computer vision, machine learning, and artificial intelligence to problems in the history and interpretation of fine art paintings and drawings. This new field, which some call computer-assisted connoisseurship, is rather different from familiar digital humanities. While some of these computational efforts have been rather ham-fisted and based on unsound methodology or incomplete art historical context, other efforts have provided valuable cultural insights or have resolved art-historical debates for which traditional (non-computer) methods have proven inadequate. For instance, computer-assisted connoisseurship proved central in rebutting David Hockney’s claim that some Western artists, starting at the time of Jan van Eyck, secretly traced optical projections as part of their praxis.

It is natural that these computer efforts, most of which are built upon methods for analyzing photographic images, medical images, and such, address the formal visual aspects of artworks—form, color, style, brush strokes, marks, figure identities, and so on. It is thus “computational formalism” that art historian Amanda Wasielewski explains and critiques. In doing so, she starts a thoughtful and likely productive conversation between the “two cultures” at hand, one that will help ensure that future work in computational methods provides genuine value to the study of art.

One of her central concerns is that many machine learning projects are biased … that is, because the images being processed are themselves selected in a biased or unrepresentative manner, any conclusions are therefore biased, flawed, or misleading. This is a valuable and often just critique. However, she seems unconcerned that non-computer methods are provably more biased; even formalists art scholars are highly biased—in their exposure to and selection of works, the choice of formal aspects they consider, the ideological biases they implicitly or explicitly employ, and so on. For example, Wasielewski is particularly concerned that the scourge of fake and misattributed works in museum collections and electronic databases will lead to biases in computer studies but fails to acknowledge that such flaws in image collections have necessarily plagued traditional art scholarship for decades and indeed centuries, with scant attempts to compensate. In fact, it is the computer scientists who continue to be far more concerned about fakes and misattributions than are humanist art scholars and are developing new methods for ending the scourge of fakes.

More generally, computer methods are superior to traditional methods in regard to bias because the digital data sets and the processing algorithms in question are frequently publicly available for analysis, replication, and especially refinement by others. For this reason, we can expect any biases inherent in current computer methods to be reduced over time, whereas there are little or no assurances for equivalent improvement through traditional approaches.

For the most part Wasielewski is a reliable interpreter of the computer science she considers, though she makes a few small errors and oversights. For instance, she is concerned that because basic computational clustering takes art data and splits it into distinct clusters, each painting will be placed into a single category or cluster, such as Impressionism—an unacceptably coarse simplification given that many paintings span two or more such categories. However, there are many machine learning techniques that can yield graded or partial category memberships, so that for instance a single painting might be assigned memberships into multiple artistic movements. Moreover, so-called hierarchical or taxonomic computational clustering, familiar from biology (kingdom, genus, species, …), can and has been applied to artworks. Thus, while at one level of abstraction a typical de Kooning figurative painting would be distinct from a (non-figurative) dripped Pollock, at a higher level of abstraction both would be classified as Abstract Expressionist.

Wasielewski makes a few other errors. She incorrectly reports that Thomas Hoving stated that 40% of the works in public museums, such as the Metropolitan Museum of Art, were fake or misattributed whereas Hoving in fact stated that 40% of the works examined for consideration for admission to that collection were fake. (Regardless, the number of fake works in museum collections is surely far higher than the general public knows or even imagines.)

Wasielewski touches upon but does not adequately address what is perhaps the most important challenge or opportunity for art scholarship arising from the adoption of computer-based connoisseurship. As background, recall that the origin of Western art history as a discipline is typically dated to the publication of Giorgio Vasari’s The Lives of the Most Excellent Painters, Sculptors, and Architects (1550), but the field entered the academy through the so-called Vienna School (c.1847), thanks to Rudolf Eitelberger von Edelberg and colleagues. Justification for this transition were arguments that art history need not be “mere” subjective interpretations but should embrace an ethos of truth finding. Thus, the objective (albeit difficult) questions of attribution and authentication were of paramount importance in the early years of academic art history. As Wasielewski points out, “objective” problems such as identification are “…not the end goal of most scholarship in the field today” (p. 94) and goes further to question “…whether truth should have any place at all in humanities research today” (p. 115).

The computer scientists entering art history come with disciplinary ethos, reward system, technical skills, and interests in deducing truth rather than presenting multiple interpretations. As such, the rise of computer-assisted connoisseurship places art scholarship at a crossroads. Whether the majority of art historians—particularly academic art historians—care about finding truth in their discipline (aided by computational methods and ethos) rather than producing sequences of inherently biased interpretations is yet to be answered.