AI Is Quietly Deciding Which Art Enters History, Study Warns

Artificial intelligence is increasingly being used to classify, attribute, and evaluate works of art, and a new study argues that this quiet shift is doing something far more consequential than speeding up research: it is deciding which artworks, artists, and traditions become visible in the historical record. In a systematic review published in the journal AI & Society, Uğur Kocager of Istanbul Technical University’s Graduate Program of Art History and sociologist Ebru Yetişkin examined 180 publications drawn from 11 academic databases to map how the intersection of artificial intelligence and art history has developed. Their conclusion is striking. What looks like a neutral technical assistance process is, in fact, a socio-technical reorganization of visibility, authority, and knowledge production in the discipline, one they capture with a new concept: algorithmic canonization.

The idea of a canon, the set of works and figures a field treats as canonical, has always been contested in art history. Scholars such as Donald Preziosi and Eric Fernie documented how the discipline’s methods were shaped by particular institutions, and James Guillory’s work on literary canon formation showed how cultural capital accumulates around certain texts through institutional mechanisms. What Kocager and Yetişkin add is the observation that these mechanisms are now partly encoded in software. When a deep neural network classifies artistic styles, when a computer vision system retrieves visually similar paintings, or when a machine learning model quantifies creativity in an artist’s network, it is not merely processing images. It is reproducing and amplifying judgments about what counts as art, which styles matter, and whose heritage deserves digitization in the first place.

The technical lineage of this transformation stretches back further than the current deep learning boom. The authors trace early computational forays into the humanities to the 1990s, including machine learning systems applied to typologies of archaeological ceramics. By the mid-2010s, convolutional neural networks had transformed the field. Researchers demonstrated that binarized features derived from deep networks could classify artistic styles with remarkable accuracy, and systems were built to search vast databases of paintings by visual similarity. Computer-assisted analysis of brushstrokes in van Gogh’s works showed how digital image processing could extract visible features without human supervision, while studies applying optical techniques to works by Jan van Eyck and Robert Campin hinted at how algorithms could contribute to questions of perspective and attribution that had occupied art historians for generations.

Today, the methodological repertoire is broad. Surveys of computational methods for iconic image analysis document approaches ranging from pattern extraction and recognition in paintings and drawings to deep learning perspectives on beauty, sentiment, and remembrance in art. Some researchers have proposed creative adversarial networks that learn about styles and then deliberately deviate from style norms to generate novel images, prompting fierce debate about whether machines can be creative at all. Others have argued, as computer graphics researcher Aaron Hertzmann does, that computers do not make art, people do, and that the framing of AI systems as autonomous creators is itself a category error. Meanwhile, scholars in the digital humanities have explored how computational methods reshape the transformation of information into knowledge, and how interdisciplinary brokering between computer scientists and humanities scholars actually works in practice.

What makes the new study distinctive is its insistence that these technical developments cannot be understood apart from their social and political context. Drawing on traditions in science and technology studies, including Sheila Jasanoff’s work on the co-production of science and social order and Donna Haraway’s account of situated knowledges, the authors treat AI-assisted art history as a site where classifications and consequences intertwine. Geoffrey Bowker and Susan Leigh Star famously showed that classification systems are never innocent; they sort people and things into categories that carry material effects. Algorithmic canonization extends that insight to the art world: the datasets used to train models, the metadata attached to digitized collections, and the objectives chosen for machine learning systems all embed prior judgments about artistic value, and those judgments are then scaled and standardized in ways no individual curator or historian could achieve.

The empirical heart of the study is a systematic analysis of its 180-publication corpus across three dimensions: historical development, methodological orientations, and patterns of knowledge production. The multistage selection process, managed through a curated bibliographic database, allowed the authors to track who is publishing what, where, and with which conceptual ambitions. The findings reveal what the authors call a significant asymmetry. In the literature they examined, Asia appears mostly in association with data production and technical applications, while institutional structures centered in Europe and North America dominate conceptual framework building and theoretical discussion. In other words, the raw material for AI-driven art history increasingly comes from Asian collections and datasets, but the interpretive frameworks that determine what that material means are being constructed elsewhere.

This division of labor matters because it echoes long-standing critiques of how knowledge is produced in a globalized but unequal academic system. Related work in AI & Society has described the coloniality of experiments in art and AI, arguing that such experiments often proceed as if art history were a universal, placeless enterprise, and has examined cognitive imperialism in AI systems, proposing indigenous epistemologies as counterweights to embedded bias. Community-centered studies of text-to-image models in South Asia have documented how AI’s regimes of representation misrecognize the very populations they depict. Kocager and Yetişkin’s contribution is to show that a parallel dynamic operates in the apparently benign domain of art historical research, where the digitization of Asian heritage feeds computational pipelines whose theoretical framing remains Western.

Crucially, however, the authors resist the temptation to explain everything through a simple West-versus-Asia dichotomy. Their analysis emphasizes that Asia is not a monolith but harbors different centers and spheres of knowledge production, each with its own institutional configurations, research traditions, and technological capacities. Studies comparing Chinese and Western art history through big data approaches, for example, represent a distinct line of inquiry that cannot be reduced to a peripheral role. By positioning Asia not as a comparative object of study but as an analytical lens, the authors aim to reveal the inequalities in the existing knowledge production order from a vantage point that those inequalities would otherwise obscure. This methodological move follows broader arguments in feminist and postcolonial philosophy of science that the view from a marginalized position can expose assumptions invisible from the center.

The stakes extend beyond academia. As the Stanford AI Index documents the accelerating industrial scale of artificial intelligence, museums, auction houses, and cultural institutions are adopting computational tools for cataloging, authentication, and curation. When such tools determine which works are retrieved in a search, which attributions are flagged as plausible, or which styles are deemed similar, they participate in assigning value in the postdigital realm, a process scholars have already linked to canon formation and cultural heritage management. If the training data overrepresents certain regions, periods, or media, the resulting systems will systematically elevate some traditions and marginalize others, not through any explicit decision but through the accumulated weight of what was digitized, labeled, and learned. The authors’ framework gives curators, historians, and developers a vocabulary for interrogating these effects before they harden into a new, algorithmically enforced canon.

The study, which forms part of Kocager’s doctoral research at Istanbul Technical University conducted under Yetişkin’s supervision and supported by the TÜBİTAK 2211-A National PhD Scholarship Program, ultimately calls for a more pluralistic and critical engagement between artificial intelligence and art history. That means, at a minimum, attending to whose datasets train the models, whose categories structure the classifications, and whose institutions hold the authority to interpret results. It also means recognizing that technical choices, from feature extraction methods to similarity metrics, are interpretive commitments in disguise. As AI systems become embedded in the infrastructure of cultural memory, the question is no longer only whether machines can analyze art, but who gets to decide what the analysis means, and whose art history the algorithms will write. The concept of algorithmic canonization is offered as a starting point for answering those questions before the canon closes silently, one training set at a time.

Subject of Research: The role of artificial intelligence in canon formation and knowledge production in art history

Article Title: Algorithmic canonization: AI and art history through an Asian lens

Article References: Kocager, U., & Yetiskin, E. (2026). Algorithmic canonization: AI and art history through an Asian lens. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03361-7

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03361-7

Keywords: artificial intelligence, art history, algorithmic canonization, digital humanities, knowledge production, Asia, computer vision, deep learning, canon formation, science and technology studies, cultural heritage, AI & Society