A cultural object called dataset

In an era where images are treated as data and datasets become predictive tools, the talk Archivi e curatela incontrano l’IA (Archives and curatorship meet AI)—curated by Sineglossa for Creators Day, organised by Delizia Media, held on 13 June in Bologna—opened up an urgent space for reflection on what artificial intelligence reveals, and what it leaves out of frame. From curatorship in the digital realm as a political act to data corruption as artistic practice: researcher Gaia Tedone and visual artist Paolo Bufalini are questioning the dataset as both cultural and ethical object. Here is how.

On 13 June, within the industrial atmosphere of Spazio Bianco at DumBO in Bologna, the talk Archivi e curatela incontrano l’AI, curated by Sineglossa for Creators Day¹, lasted around thirty minutes—yet it was enough to spark a chain of reflections that later took shape in two interviews: one with researcher, curator and digital humanist Gaia Tedone, the other with visual artist Paolo Bufalini. Two distinct paths, but a shared meeting point: a deep attention to the image, what it says, what it hides, how it is classified in AI systems, and the dust that settles on it, when time passes or is allowed to seep into the dataset.

What are images within AI systems? Who decides what they represent, and for what purpose they are collected and organised? And what is left of the image once its resemblance fades? When the face disappears, the data becomes corrupted, and what is left is only a vibration? These are the questions at the heart of Tedone’s research, and they resonate in Bufalini’s artistic process too. Here is how.

What is an image in AI systems

During the talk, as slides of futuristic installations and AI- or machine-based art practices played, Gaia Tedone said something that echoed long after: «The dataset is a cultural object». This might seem obvious to those familiar with these discussions, but for outsiders, especially when it comes to images, it flips the dominant logic entirely.

Beginning with her project This image is not available in your country²—a simple t-shirt bearing the familiar error message displayed when images are blocked due to geographic restrictions—«I began to realise that technology operates in very specific ways», says Gaia Tedone. «The circulation of images online is tied to dynamics of power that are often invisible», just like the images themselves, which have an opaque, “numeric” side that remains hidden. «That project became the device that allowed me, during my PhD, to refine the concept of the networked image», she explains. «An image that has a double nature: on one side a visual interface—what we see and interpret—on the other a computational object made of data, metadata and signals that circulate, accumulate and are read by machines».

From this perspective, an image is no longer merely a representation but an infrastructure: a node within a network that transforms and measures it. In fact, Artificial Intelligence as we know it today could not exist without this dual nature of the digital image. In AI systems, the image becomes a visual form stripped of ambiguity, segmented and rigidly classified to be read and processed by statistical models³.

This is because the meaning in AI systems no longer emerges from cultural or discursive space, but is determined by those who build, label and use datasets: engineers, computer scientists, corporations.

«Labelling, the process of assigning semantic tags to images, is never neutral», Tedone emphasises, «and the development of generative AI has relied on vast amounts of data and images—many of which were taken from the Internet, carrying with them the full weight of our cultural stereotypes».

In this sense, the image is governed by an “algorithmic” power structure that decides what is visible, what is relevant, and what is left out.

The artist as data craftsman

In a system where the AI industry treats images as raw material for value extraction, reducing them to codes and categories within an economic and predictive logic⁴ curatorship can become an act of resistance and re-signification.

According to Tedone, curatorship in the digital field «is a method for organising visual and semantic meaning—that is, of languages and codes» that enables us to read and question the cultural mechanisms behind images.

During the talk, she presented two emblematic examples: the exhibitions Data/Set/Match at The Photographers’ Gallery in London⁵, and Training Humans at Fondazione Prada⁶. In both, datasets originally created for scientific purposes and reinterpreted by artists are not simply exhibited but critically analysed, amplified and pushed to extremes in order to expose the inner workings of automatic classification. These are projects that stretch the logic of visual technologies to their limits, revealing their aesthetic, political and historical frameworks.

Here enters the figure of the “artist-craftsman”, as Tedone calls it—someone who uses artificial intelligence not merely as a tool, but as a material to be shaped from the data itself, constructing and questioning it.

«Artists working with datasets become craftspeople of their own data foundations», Tedone explains. «They become authors of the work by engaging in phases of the process that, as users, we rarely touch. They build their own archives, their own images, their own labels. For example, by photographing at least one hundred and fifty tulips⁷for a particular project, they are able to classify them and create their own dataset». In this way, it becomes clear that every generated image rests on specific classification choices—with aesthetic and semantic implications—and that there is always human labour behind it. «It is a craft gesture, but also a political one», and it is precisely this approach we find in Paolo Bufalini’s project Argo, in which he trained an AI model using his own family archive. 1,300 photographs of his mother, father and sister were used to explore the dataset as a cultural object and as a gateway to the dreamlike, producing a synthographic⁸ series of sleeping figures that appear real but are not. If Tedone studies artists who build archives to reframe images, Bufalini embodies the figure who, as he says, «Building a personal dataset means choosing what to preserve, what to show, and what to forget».

Image stereotyping: towards collapse

«The model tended to generate extremely harmonious, polished faces—lips, noses, mouths», Bufalini recalls. Even when trying to introduce variation into the image generation process, the faces that emerged were always, if subtly, “harmonised”. «There was an internal pressure within the model towards stereotyping», despite fine-tuning⁹ on a base model already pre-trained with images resembling real photography—thus without graphic or 3D-style elements. Bufalini chose to let a trace of this “algorithmic smoothing” remain: both as evidence of the tool’s predictive logic and as a disturbing mirror of how portraiture itself functions. «What emerges is an interesting, structural tension. You can work to counter it, but only up to a point».

«It is not just about data: it is about how the model works. Its statistical nature means it will always tend to generalise—and therefore marginalise whatever lies beyond the centre».

Bufalini’s response is not rejection, but subtraction. The stereotype is acknowledged, displayed, then hollowed out—and this direction is embodied in his new image series Rückfiguren. Part of his PhD project at the Academy of Fine Arts in Naples, it began in parallel with Argo¹⁰ and uses the same generative process. Rückfiguren—a title that references the Romantic motif of the figure seen from behind, contemplating a landscape, and alludes to a kind of technological Sublime—figures with their backs turned mark a threshold. «By denying the face, I focus on the surface of the image, its grain. And there something more pure emerges, closer to the essence of the generative process», Bufalini explains. His work thus becomes a dig into that generative core.

«I would also like to experiment with dataset corruption: to collapse the model in on itself and observe what remains».

This process, which involves repeatedly retraining a generative model on the very images it has produced, «is a way of pushing AI to its limits, but also of revealing its internal workings, its blind spots, its fractures».

«It is fascinating to see what happens when a model learns from itself», because it leads to a progressive detachment from the original structure. «This opens up a new space—not a technical failure, but a new way of thinking about the image».

In line with Tedone’s thinking, Bufalini too is questioning what happens to images in the digital realm. «However faithful it may be, an image is never reality», he asserts. «There is always a gap—and this is true even for photography, because what it shows is always the result of choices: framing, exposure time, point of view. Within that gap, the image gains its own substance: it becomes reality in its own right, and an agent in the world». Artificial intelligence, trained on stereotyped images and labels, «does not reinterpret reality, but interprets interpretations», says Bufalini.

«It does not work with the world, but with its filtered representations. It is a second-degree form of indexicality¹¹»

A condition in which the image loses its link to the real—not a trace of it anymore, but an abstract synthesis of pre-existing data, the result of a prediction shaped by earlier representations. «That is why, in preparing the dataset, I scan images in high resolution, but do not clean off the dust, which is absorbed during fine-tuning and becomes a trace of reality in images that have lost their connection to their original context», Bufalini recounts.

Ethical guidelines for datasets

Talking about datasets today means talking about archives, but also about memory, exclusions and desires. It means crossing the image like a threshold, letting the dust settle rather than wiping it away. The dialogue between Tedone and Bufalini marks an important convergence: on one side, the curator who reads data as a cultural object; on the other, the artist who becomes curator of his own generative process. Both demonstrate that we can no longer avoid thinking of the dataset as an aesthetic and ethical device. Both already have concrete proposals in mind: a set of ethical guidelines for artists working with generative AI, based on principles such as transparency, declaration of data provenance and awareness of stereotyping effects. More than that, as Tedone puts it: «We can no longer afford to leave dataset creation solely in the hands of technicians and big tech», she says. «We need an interdisciplinary approach: artists, theorists, archivists, interaction designers, curators. Each with a different perspective on data to interrupt extractive dynamics and propose new forms of representation».

---

1. The open event, organised by Delizia Media, connects cultural organisations with the world of new media and content creation. The 2025 edition took place at the Spazio Bianco in Bologna’s Dumbo and welcomed around 600 visitors. Read more: https://www.deliziamedia.com/creatorsday ↩︎
2. Learn about the project: Tedone, G. (2017). Tracing networked images: an emerging method for online curation. Journal of Media Practice, 18(1), 51–62. https://doi.org/10.1080/14682753.2017.1305843 ↩︎
3. Read more: Tedone, G. (2019, November 29). From spectacle to extraction. and all over again. Unthinking.Photography. https://unthinking.photography/articles/from-spectacle-to-extraction-and-all-over-again ↩︎
4. An economy in which meanings (images, texts, symbols, data) are produced, distributed, controlled, and exchanged within a system that determines who can say what, how, and with what social or material effects. Read more: Tedone, G. (2019, November 29). From spectacle to extraction. and all over again. Unthinking.Photography. https://unthinking.photography/articles/from-spectacle-to-extraction-and-all-over-again ↩︎
5. Discover the exhibition: https://thephotographersgallery.org.uk/data-set-match ↩︎
6. Discover the exhibition: https://www.fondazioneprada.org/project/training-humans/ ↩︎
7. The reference is to Anna Ridler’s work: http://annaridler.com/myriad-tulips ↩︎
8. A synthetically generated image through the use of Artificial Intelligence (AI) and text-to-image models. It differs from traditional photography in that it is not based on light or the recording of a real-world image, but on algorithms that create the image based on textual input or other existing images. ↩︎
9. Fine-tuning is the re-training of an existing Artificial Intelligence model on a smaller, more specific dataset to adapt it to a particular style, content or context. ↩︎
10. We have already talked about it here: https://mangrovia.info/passato-performance-paolo-bufalini-argo/ ↩︎
11. In photographic theory, the idea of index, developed by Charles Sanders Peirce and taken up by Rosalind Krauss, defines the photograph as an index mark, i.e. a trace left by an actual event: something has been in front of the camera, and light has imprinted that presence on the film or sensor. Even if what is photographed is staged or constructed, what matters is that something physically happened in front of the lens. With the image generated by AI, this link is broken. Hence, the idea of indessicality. For more: Krauss, R. (1977). Notes on the index: Seventies Art in America. October, 3, 68–81. https://doi.org/10.2307/778437https://www.jstor.org/stable/778437 ↩︎