MACHINE YEARNING

I AM THINKING OF A BANANA. I am thinking of Albert Eckhout, the Dutch painter who arrived in Brazil in 1637, and of each of the bananas he painted, reclining against a mound of succulent fruit. Cut for consumption and staged on a parapet before a cloud-stacked sky, they conjure through their placelessness a paradise of cultivated plenty uprooted from the realities of colonial extraction.¹ 

I am thinking of the Brazilian artist Antônio Henrique Amaral, and of his 1972 painting of a swollen banana, its skin split, swaddled in a constricting shroud of black rope. 

If in 2023 you were to have prompted the generative-AI program Midjourney to produce an image of a single banana, it could not have done so. Trained on millions of images of bananas—clusters of supermarket bananas, bunches of kitchen counter bananas, tumescent claws of bananas growing on trees—and beholden to the weight of statistical probability over meaning and context, generative AI did not have the input needed to output just one banana.² It could generate several or very many or even one large shaft with two stems, an unnerving banana monster. But it could not think of a banana, because it doesn’t think; in lay terms, Midjourney and image generators like it produce images by combining the results of calculations based on the likelihood that the verbal prompts it receives are sufficiently similar to the data underlying other images it has previously processed. 

Computer vision works by assigning weighted numerical values to the pixels that make up digitized images. When an image is “classified,” a human identifies and labels what is represented in the image (think about all those reCAPTCHA challenges you’ve taken to prove yourself as a human by clicking pictures of traffic lights—when you do this, you’re classifying images for Google, for free). This is used as training data for a neural network, enabling it to issue a statistical estimation about whether the numerical value of a given image matches those of other images. 

The precursor to today’s computer vision dates back to the late 1950s, when Frank Rosenblatt, a researcher at Cornell Aeronautical Laboratory, unveiled his Perceptron, which programmed a camera to detect the location of a shape on a flash card. While his experiment initially garnered enthusiasm, research into AI stalled in part because there simply was not enough available data on which to train image recognition models, heralding a fifty-year period known as the “AI Winter.” A breakthrough came in the early 2000s when the Stanford computer scientist Fei-Fei Li hired thousands of low-paid, anonymous laborers through Amazon’s Mechanical Turk to tag three million images. She then staged a competition to determine the most accurate algorithm for image recognition based on that training set, the results of which opened the floodgates for the development of computer vision.³ It was around this moment that computer vision began to attract the notice of art historians, critics, and museum professionals. Over the ensuing fifteen years, publications about the theoretical or methodological implications of computer vision for the study of art have proliferated, generating an extensive discourse about its potentials. 

In his 2017 article “The Internet as a Structure of Feeling: 1992–1996,” an investigation of the internet’s nascent years after it became available to the general public, scholar of media and technology Thomas Streeter argues that technologies take on a particular character and achieve certain effects in the world not simply because of what they literally can achieve but because they are informed by what Raymond Williams called “structures of feeling.” For Williams, a structure of feeling encompasses values, outlooks, expressions, and impulses that, while amorphous and broad, nevertheless structure culture and have real impacts in the world. As Streeter writes, it is “the expectations that technologies become intertwined with from early on in their conception” that effect social change as much as the technologies themselves.⁴ The internet has developed to be what it is and caused the social changes that it has because early feelings about it and hopes for what it could achieve shaped the laws regulating it, the uses to which it was put, and the framework governing how we think about its social role. 

Developmentally, generative AI is in a moment similar to the one described by Streeter, when the “cultural expectations of the Internet shaped the development of and thus [became] part of the artefact itself.”⁵ Scholars have focused on what computer vision does (asking how accurate it is, how efficiently it works, and how it might aid art-historical research and teaching) but not on the material conditions upheld or advanced by the rhetoric of its potential use. This is a banana problem: It is as if scholars have been fixated on the piece of fruit itself and not on the worlds that Eckhout and Amaral show us, which were made and destroyed by the harvesting and consumption of that fruit. Instead of focusing on what computer vision can achieve in processing images, I want to offer a historical account of the language and the feelings used to frame its imagined role in art-historical practice. 

Scholars have focused on what computer vision does (asking how accurate it is, how efficiently it works, and how it might aid art-historical research and teaching) but not on the material conditions upheld and advanced by the rhetoric of its potential use.

AN ACCOUNT BY Emily L. Spratt, a historian of Renaissance and Byzantine art, and Ahmed Elgammal, a computer scientist, offers an early glimpse into art historians’ introduction to this technology and the language they would use to describe their hopes for it in the ensuing years. After meeting at a conference in 2013, the two scholars partnered up to build on Elgammal’s work with other computer scientists, which purported to show how computer vision could quantify similarities between works of art. The result of this original paper was a swell of media attention from venues including The Guardian and the Washington Post that cast its proposal as a gauntlet thrown down to humans by machines.⁶ Responding to this media attention and to public reactions from art historians, Spratt and Elgammal subsequently produced a survey that they distributed to both art historians and computer scientists, devised to gauge scholarly sentiments about the meeting of these two fields. 

But as they state at the outset, “Although these surveys were intended to assess perceptions of both art historians and computer scientists, the focus of our study was on the receptivity of art historians to computer vision technology.”⁷ Throughout the discussion of the survey’s results, art historians are on the back foot, this condition ascribed to our “unease” and “anxiety” about technology rather than to the unintelligibility of the survey questions. If I were asked whether I thought computer vision could “aesthetically comprehend and judge beauty,” I would wonder, first, what they think an art historian does, before even considering the question of whether a convolutional neural network could do it. The paper proceeds to paste together a transhistorical collage of names (Aristotle, Carl Linnaeus, Giovanni Morelli, Giorgio Vasari) in the service of the idea that partnership between the arts and science (both undefined) has an illustrious genealogy. The conclusion is clear: Art historians and computer scientists should collaborate; and since art history is an aesthetic enterprise involving judgment and a concern with “beauty, style, dating, and relative influence,” computer vision’s capacity to quantify these things results in an epistemological challenge to art historians that provokes greater self-awareness about the limitations of our methods.⁸

An article of faith in Spratt and Elgammal’s paper, and one that is regularly professed in the literature on computer vision and art, is the benefit of collaboration, whether cross-disciplinary or between humans and machines.⁹ In both cases, the limit on individual human capacity is a lamentable impediment to progress, which we can overcome either by pooling the intellectual resources of specialists in a wide array of fields, or by assimilating ourselves to transhuman assemblages.

COLLABORATION WAS THE ORDER of the day at an April 2018 symposium hosted by the Frick, a milestone event that brought together computer scientists, tech entrepreneurs (including representatives of Facebook and Google), and financiers alongside art historians and librarians to forecast future applications of computer vision. In a lightning-round talk, Elizabeth Honig, a historian of early modern art, offered a snapshot of a project she had initiated in 2012, which involved a student devising an image recognition system to assist research on the repetition of objects and compositions in the workshop of the seventeenth-century artist Jan Brueghel. Over two days of talks, it was one of only three examples that used computer vision to answer a question formulated from within art history itself.¹⁰

But near the talk’s conclusion, something interesting happened. Honig described how, after presenting this project at Berkeley a few years later, she was approached by the computer scientists in attendance; “we love your dataset,” they said to her.¹¹ From the vantage of 2025, it’s an ominous moment. At the time, the audience’s laughter and Honig’s ebullience obscured beneath the timbre of jollity that the largest contingent at the event was there to advance the development of and receptivity to surveillance capitalism’s latest technology, while the much smaller contingent of art historians, curators, and librarians were enthusiastic about a tool that might assist them in answering questions about art. For the computer scientists, an enterprise that demands ever more labeled data to devour would of course find it convenient to have a freely available dataset, already labeled through others’ hard work and ready to process (especially at that moment in the early teens when training sets largely comprised photographs and not the trickier objects like paintings and drawings).¹² Look at all those bananas! 

It seems impossible to challenge the idea of collaboration as inherently good for knowledge production; two heads are better than one. It’s just common sense. But common sense is not a value-free confrontation with material conditions; it is a set of values that through repetition and social reproduction has achieved the status of neutrality: a positionless position that resists refutation because it has no position to refute.¹³ What emerged from so many of the Frick symposium’s talks was a celebration of collaboration as a value in itself, regardless of whether its outcomes are equally valuable to all collaborators. 

The historiographical incoherence at the heart of many of the Frick talks persists in more recent efforts to legitimize computer vision’s place in art history by means of lineage. On the one hand, we are told that art historians are resistant to technology or are slow adopters; on the other, that technology—and photography in particular—was a precondition for the institutionalization of art history as an academic discipline in Germany at the end of the nineteenth century because it allowed scholars to examine and compare seemingly countless works of art. 

Casting errors as valuable for self-reflection without providing concrete examples of how that reflection has been beneficial does little more than create a permission structure for the failures of this technology.

Whether or not it is true that art history as a discipline is defined by the perceptual protocols demanded by a certain kind of imaging technology, claiming as much paves the way for arguments endorsing the application of computer vision to the study of art. Recently, Emily Pugh and Tracy Stuber, scholars at the Getty Research Institute, have argued that computer vision “should be regarded as one link in a longer historic chain of machinic vision that began in the nineteenth century with photography.”¹⁴ The imperative to unite these things in a sequence, to create a genealogy, is a choice; it is not a natural state of affairs.¹⁵ The appeal to tradition issues an implicit argument for inevitability. It is also to claim—as some have done explicitly—that the same debates about the study of art through photography are most relevant to the application of computer vision to art. 

Computer vision, however, is not an extension of photography. It does not show us a banana. It shows us that a digitized image is highly likely to represent a banana, based on the similarities between the way its data is configured and the way that of millions of other digitized images is arranged. Computer vision is better described as a form of information management and prediction, as configurations of data with which probabilistic applications have been made, based on human instructions of varying degrees of clarity or specificity. A more productive way to think about lineages would be to consider a commensurate historiography that recognizes the stakes of information management as the essence of machine learning.¹⁶ A historicist approach to the issue would reject the simplicity of a single narrative that weds the development of art history to the reproductive tools we use to represent works of visual art. Instead, it would examine the material and ideological conditions within which a given imaging technology emerges, is deployed, joins others, and works within information management systems to map canon formation and inform the methods we use to investigate works of art. It would also encompass issues of ownership and power: Who owns these tools and systems, and what tangible or intangible benefits result from their use to frame art-historical inquiry?

Instead, however, under the influence of the metaphoric fallacy that computer vision is a kind of sight, some have proposed that, like photography, it might offer an enlightening supplement to the human eye. Insights gained from computer vision’s erroneous conclusions form another touchstone in writings from about 2020 onward. It was around this time that widely read publications like Ruha Benjamin’s Race After Technology (2019)and high-profile projects like Trevor Paglen and Kate Crawford’s ImageNet Roulette, 2019, and Adam Harvey and Jules LaPlace’s MegaPixels, 2017–20, revealed how biases encoded in computer vision make it an unreliable system for extrapolating useful information from digitized images. The idea that scholars adopted in response is that computer vision’s errors might be revelatory for exposing the prejudices suffusing the online data on which it is trained, or that its infelicities might cause us to see something that our conventional ways of looking at the world obscured from view. For example, while no human eye has seen every one of the roughly sixty-five thousand Ed Ruscha photographs processed by computer vision at the Getty Research Institute in Los Angeles, maybe the confident statistical work of a prediction machine can incite new thoughts—even when it is manifestly incorrect, as Nathaniel Deines of Getty Digital suggests in an article on the J. Paul Getty Trust’s website.¹⁷ He remarks that the program that processed the Ruscha photographs erroneously classified over a thousand of them as containing snow; what we are meant to make of this error beyond its capacity to “[provoke] reflection” is left unsaid. 

Advocates of computer vision have adopted the position that, even in error—especially in error—computer vision can revise the way we see the objects of our study. “I have become more attentive to my own habits of looking, with a sharpened awareness that my perception is inflected by both disciplinary and personal biases,” art historian Elizabeth C. Mansfield concluded in her CV-assisted analysis of John Constable’s cloud studies, published in 2022. “Do I believe computer vision has the potential to serve as a kind of prosthetic ‘period eye’? Not exactly. But I do think there’s potential for computer vision to aid art historians in gaining a better understanding of perception as well as vision.”¹⁸ This framing posits the algorithm as a form of alterity, an otherness that aids in self-knowledge. 

Casting errors as valuable for self-reflection without providing concrete examples of how that reflection has been beneficial does little more than create a per­mission structure for the failures of this technology. In the case of the Ruscha photographs, computer vision was simply wrong. In the case of Constable’s cloud studies, it could not be used to answer the art historian’s original questions and therefore required her to shift her queries to match what the technology can do.¹⁹ Questioning ingrained assumptions, denaturalizing our habitual ways of looking, noticing our blind spots—these are commendable goals. But when the rhetoric stops short of articulating how and what is different about what we are seeing with the aid of computer vision and why it matters, then all we have are slogans in support of computer vision itself. I question the validity of suggesting, as Mansfield and others have done, that machine vision avails a new epistemology, or that we can divest ourselves of our disciplined perception with the aid of commercial software or coded operations that are sui generis to the technology and not, instead, what they actually are: an automation of the epistemology of the humans who coded it. Computer vision doesn’t hallucinate because of an inscrutable deus inside the machina. It “hallucinates”—such a groovy word for being wrong with confidence—because that is the ideology with which it was encoded. 

THE MOMENT A PRODUCT begins to spit out errors with self-assurance should be the moment that it is pulled off the factory line. That is not what is happening. Instead, these errors have been grist for the mill for those who are making hay of AI while the sun shines on its promise. As digital-humanities scholar Leonardo Impett has remarked, as of this moment there is only “a small trickle of work which attempts to cross machine vision with art history,” and that work “has struggled to make any real impact on the rest of the discipline.”²⁰ Publications on this topic by scholars in art history have exhibited rudimentary use cases, including identifying the subject matter of a painting and attempting to show trends over time in the composition of paintings or disposition of figures in space. Others have used computer vision with some success in aiding material analysis of pigments and texture; more mercenary applications by museums and collectors include its potential to authenticate works of art. The cluster of articles devoted to the topic that appeared alongside Impett’s piece in a special section of Art Bulletin in 2024, “Art History After Computer Vision,” prove his point: Out of eight articles, only two actually use computer vision to analyze works of art. (Unsurprisingly, both projects were concerned with photography.) 

Nevertheless, new certificates in art and AI, publications, conferences, symposiums, and workshops are cropping up prophesying the value of computer vision in the field of art history rather than demonstrating its value.²¹ My point here is to critique not these individual projects, but the larger dialogue they comprise: Conversations about what AI does, which are prolific, can easily become a form of misdirection, hustling attention away from what the hopeful rhetoric of its potential authorizes.

That hopeful rhetoric has surfaced in rather a counterintuitive way. The special section of Art Bulletin—which, as the discipline’s flagship journal, constitutes a semiofficial statement about AI’s importance in the field—devoted the majority of its space to critical commentary about the application of computer vision to works of art. Many of its contributions marshal examples of naive uses of AI, the antithesis of what those who are knowledgeable could do with this technology. They issue well-informed and insightful caveats about the dire ethical implications and harmful consequences of computer vision’s use but nevertheless, for the most part, conclude by prospecting some vague potential for the technology to shape the future of the field. 

And so, when these caveats are leveraged as prefatory to the proposition that we, too, should implement machine learning in our work, then it is a performance akin to awareness raising. We are shown an example of all the harms that can occur through the deployments of this technology, so as to assure us that the authors know precisely what they are not endorsing. Here, self-awareness displaces action. The solution, of course, is to do something about it rather than retconning its deficiencies as failing forward. But art historians are not in a position to fix the problems with machine learning; we can only recuse ourselves from its support. Postmortems of biases and errors, recognition of the limitations and misalignment of CV’s capabilities with matters of relevance to current art-historical research and teaching raise the very question that few people seem to be asking: What good is this? Or, more importantly, for whom is it good? Cui bono?

Big tech has primed the pump, and in response many have embraced its start-up logic. Imagine a product (computer vision art history), make a pitch, get the funding, figure everything out later. There is a kind of speed we are being asked to commit to, in line with the demands of tech itself. Governed by a growth mindset devoted to the logic of disruption, moving fast and breaking things, a flourishing structure of feeling about AI and art history has sprung up on the grounds of very little AI-and-art-history to have feelings about. The few projects that do apply computer vision to studying works of art seem to exist only to justify funding, which then perpetuates the rhetoric of machine learning’s promise. In many cases, once the project period has lapsed and new funding has been applied for, there is no stewardship, because no one was invested in the historical dimension of the research in the first place. Incorrect metadata, broken links, and wasteland webpages are the digital detritus littering this research ecosystem.²³

As Daniel Greene has shown with respect to public libraries and secondary schools, programmatic training and education in digital skills ensure the survival (but not the flourishing) of institutions whose public support has been continually eroded. He writes, “After defunding and delegitimizing institutions of social reproduction based on their purported unfitness for the information economy, the state and capital then coerce these institutions into solving problems that they didn’t create and don’t have the tools to address.”²⁴ The practice of art history does not serve the citizenry in the same way that secondary schools or public libraries do. But as a field in the humanities it has been subject to the same processes that Greene has dissected so trenchantly. While art historians struggle to get sufficient funding to support the fundamental elements of our research (travel comes to mind), universities have shown ready largesse when it comes to funding collaborative projects that twin the humanities with computer science.²⁵ This is the context in which collaboration has become the only viable path toward institutional success and even survival. 

It is 2025, and seemingly everyone wants us in the humanities to do stuff “with AI,” informed not by what the technology avails but by the hopes it encodes. Surveying the literature produced in the small corner of art-historical studies and computer vision over the past decade reveals the contours of these aspirations: the generation of a field nurtured by transdisciplinary collaboration with shared resources representing a new chapter in the evolving realm of art-historical investigation, affording us novel ways of seeing that liberate us from restrictive convention, so long as we remain self-aware. Whether or not computer vision, or AI more broadly, is the key to achieving any of this, what we are left with is a structure of feeling that is now a part of AI as an artifact, the aspirations pinned to its potentials, which are directing its adoption into spheres of intellectual inquiry where it may have little or even nothing to offer. It doesn’t matter whether AI can deliver on the promises that its boosters make regarding art-historical inquiry; all that matters is that there are enough people who believe that it can.