You might be aware that artificial intelligence algorithms work from a human-fed collection of information — a dataset, in other words — to either generate an artwork or identify ones with no or contested attribution. While combing through those sets, AI often uses deep learning, through which it is able to infer and analyze complex data and patterns. Beyond that, however, we don’t truly comprehend how its “thought process” works — what Amanda Wasielewski calls the “black box” of AI mechanics in her seriously titled new book, Computational Formalism: Art History and Machine Learning (MIT Press, 2023).
Despite our little to complete-lack-of comprehension of their inner workings, AI has rapidly become the controversial talk of the art world. AI-assisted technologies such as Midjourney and Stable Diffusion have instigated hot debates over their remodeled Boticellis and Vermeers. To the chagrin of many, we have decided to bestow AI with prizes over ourselves in art contests. Now, scholars ponder if its tools can also aid with the academic study of art history.
Wasielewski aims to answer that question in this jargon-free but not-the-most organized study of recent experiments; the book’s most critical aspect surrounds the pros and cons of using AI to identify art images and resolve attribution. She considers subjects ranging from recent collectives such as Obvious and AICAN, scientific methods for object analysis or “technical art history,” and the use of iconography to categorize images, to understand if AI can be pragmatic for art historians.
Wasielewski points out that datasets making use of deep learning can be problematic since they are typically fed information from the Western art canon, which is subject to human-led academic biases. Art historians, for instance, traditionally grant sole attribution per artwork. Even when a particular work was partially made by apprentices or artisans, it is credited to a single “master” artist (e.g. “Workshop of Bernini”). Such an approach creates datasets unfitting for AI analysis of collaborative products of multiple artists existing historically in non Western regions (for instance, Mughal ateliers).
Deep learning may also affect notions of authenticity in the art market. Later in the book, Wasielewski explores the infamous case of Leonardo’s Salvator Mundi (c. 1499–1510). Though the work’s attribution confused several (human) experts in the past, the author explains that forgeries are even tougher to identify via AI since a dataset determines an artist’s style via formal and visual characteristics of artworks. Style, methods, and materials, however, often change within an artist’s trajectory.
In more straightforward cases of known attribution and singular styles, AI can sort images efficiently. But can it interpret art? The answer is a fairly straightforward “no.” AI lacks human researchers’ ability to engage in primary study of techniques, and the capacity to contextualize an artwork within history. Quite simply: AI cannot reason why artworks look the way they do. Her thesis that AI is not suited for a humanistic pursuit of art history comes through strongly, but it could have been supplemented by further information in the form of more case studies. Presently, AI technology faces ethical concerns of copyright infringement and its use in replacing the labor of scholars, artists, and others, rendering its future uncertain. Wasielewski maintains, however, that ignoring available tools may be counterproductive in cases such as sorting material in line with its datasets. A rational balance between embracing and repudiating AI in the study of art sounds desirable — though this may be a complicated feat until we prioritize demystifying the enigmatic black boxes behind these technologies.
Computational Formalism: Art History and Machine Learning by Amanda Wasielewski (2023) is published by MIT Press and available online.