Democracy Dies in Data: Why Use AI in Museums?

AI companies have been partnering with art museums and image libraries left and right, attracting much media attention. ARTnews recently reported on how the Harvard Art Museums’ marketing team tested an effort to bring in new audiences with the use of generative AI. The Boston Globe related how museums around Boston are using AI to “bring history to life.” And several media outlets have echoed a press release from the Prado, which claims that an AI-powered video reimagining works in their collection under the lighting conditions of a solar eclipse constitutes a form of outreach to audiences who do not typically visit their museum.

Common to all of these collaborations is the argument that AI will make art more “accessible;” that it will “democratize” art. But how, exactly, does AI achieve either of these things? What do “accessible” and “democratize” even mean here? And why has this become the go-to frame for promoting the use of AI in museums?

A misunderstanding of whose interests are being served has made it difficult for people to see the answer to these questions. Writing for the Financial Times, Naomi Rea suggested, for example, that the will to deploy AI is a recent development coming from within museums themselves, desperate to find new ways to cope with falling attendance and budget shortfalls.

Taking the longer view makes it clear that it is the tech industry that has been cultivating AI usage within cultural institutions and not the other way around. Over 15 years ago, AI developers were hungry for more data on which to train their image recognition models. Museums hold just such a trove. What’s more, recognizing patterns among complex paintings, with their extremely subtle gradations of intermingled color, offered an intriguing challenge to computer vision specialists and promised to improve the models trained on these artifacts. Academic computer scientists and engineers had in fact been exploring the use of works of art for training computer vision since the early 2000s (you can find on Github, for example, lists of datasets comprising open access reproductions of art for this purpose). When Google inaugurated its Arts & Culture Platform, it was in part with this sort of training in mind. Access to already-labeled datasets of images in museum collections could be a boon for development. Indeed, in 2022, Google reorganized Google Labs to encourage people to do the work of finding new applications for algorithmic models trained on its Arts & Culture data.

BUT HOW EXACTLY has the “data” from historical images been useful to this industry? As 6 Nature authors point out in the very title of their 2025 paper, “Computer-vision research powers surveillance technology.” Data from a wide array of sources—and that includes digital reproductions of art—have been used to develop models that bear on visual policing, such as facial and body-part recognition. In 2025, the Prado collaborated with the AI firm Sherpa.ai to count crowds of people in paintings. The Prado claims that this will help people “better understand” these works of art, but it is patently difficult to see why, and they certainly don’t offer any details as to how. More likely, the collaboration is either helping to hone the use of computer vision in quantifying the size of, say, a protest; or, simply habituating people to this as a seemingly benign activity.

Critically, the determinations made about criminality, personality, and pathology by AI are obscured to the public and are outside people’s capacity to audit. This itself is fundamentally undemocratic and grounds for objecting to the suffusion of this industry into museums. Nevertheless, research in computer vision is being enlisted to assist in spurious forms of medical diagnosis based on images, and even for vetting job candidates’ personalities or individuals’ emotional states based on “affect recognition.”

But the promises to the data industry are much greater. Waymo and Uber don’t simply make money by creating a system for ride hailing; their platforms and vehicles function as closed-entity spaces for data collection, including of customers’ demographic information, in-app and in-vehicle audio recordings, and user reviews, just to name a few sources. Viewed through the eyes of tech companies, museums are a lot like Waymo: Contained spaces in which masses of people enter with a defined but nevertheless diverse set of needs, desires, and ways of satisfying them.

The sheer volume of data that can be captured and sold from our actions in a museum and from the works of art in museum collections would make any AI firm salivate. In 2021 Italian museums were partnering with tech firms, for example, to gauge how much time visitors spend in front of paintings and the expressions and postures they make while looking at them. Currently, a raft of museums is likewise commissioning “bespoke” chatbot docents to give interactive tours based on that museum’s collections. But no matter how much supposed input from museum professionals is used to develop these puppets, ultimately their puppeteers are the firms whose main objectives are not educational but mercenary: whether that means convincing the public that they, as opposed to qualified curators and museum educators, can provide an enriching museum-going experience; or simply maximizing the conditions for data capture.

NEVERTHELESS, LET’S TAKE these promotions on their own terms: They claim that AI can democratize art itself. Frankly, this is nonsense. Strictly speaking, democracy—which comes from the ancient Greek for demos (the people) and kratos (rule or strength)—refers to the extension of governing to the people. This is obviously not what is meant by the slogan that AI can democratize art.

More loosely, what they mean is that AI will make art accessible to more people. But this, too, fails to bear up to even the slightest scrutiny. A simple dictionary definition from Merriam-Webster has accessible meaning, “capable of being understood or appreciated.” With this definition in mind, we might say that increasing access to works of art would entail investing in genuine art historical education to aid people in understanding and becoming interested in everything that the visual artifacts of our past can convey. This requires paying more museum educators and not laying off curators, collections managers, and educators as the Museum of Fine Arts in Boston did this past year. It would mean extending opportunities to new artists to engage with historical art in museum collections, and loaning works of art to smaller or remote museums. It would mean doing what the Harvard Art Museums have already done: make admission free. All this is what makes art more accessible.

Instead, AI is being used to erect barriers between people and art. Referring to the examples with which I opened, this includes creating the illusion that an eighteenth-century woman can interact with beholders—with a cringey caricature of a French accent to boot. It involves an AI start-up necromancing a chatbot Frederick Douglass to narrate his own life at the Museum of African American History in Boston. And it embraces the Prado paying a media and technology studio to create a video showing their paintings in lighting conditions that they were not painted to portray. These gimmicks are all counterfactuals that actively misrepresent historical artifacts and people.

And they are gimmicks. But it is wrong to dismiss them as gimmicks alone. These AI stunts and larks exploit culture itself in order to advertise AI, particularly its capacity to educate and even to reshape education itself. They are conditioning us to see this project as safe to trust in the hands of private technology companies.

Whether it’s paying marketing firms to generate slop to entice more audiences or taking money from Big AI for the use of their collections, it is the AI industry museums are granting greater access to, not people. This is not democratization; this is the transformation of cultural institutions into quarries for the extraction and production of capitalist value.

Still, alongside “access” and “democratization” another buzzword we often hear aligned with the use of AI in museums is “empowerment.” Learning about and understanding history is one path toward empowerment, but instead of giving power to more people to learn about the past, museums are empowering Big Tech to exploit art and our activities in museums as raw material for their own purposes. Given the outsize influence this industry now has on our government, this is a dangerous proposition for our democracy.