Living with data set images

When we encounter data sets, we typically see them from a distance.They are large collections of images. Thousands, millions, now billions of images put together in a grid. Each image becomes a pixel in a larger picture.Their scale escapes our comprehension.

Contemporary AI systems are driven by an instance of what Hanna & Park (2020) term "scale thinking", an "unwavering commitment to scalability". The belief is that with ever-growing collections of images, machines could better understand, and now also make, images. More data, more compute. The assumption is that scaling up is always better. (Below: Trevor Paglen's From Apple to Anomaly makes visible, and critiques, the scale of data sets by exhibiting 30 thousand images from the ImageNet dataset as a large, unbroken mosaic at the Barbican’s Curve Gallery.)

But what can an image do? What if we look at these images not as drops in an ocean, but as interesting, meaningful elements of an algorithmic visual culture? What could we learn by zooming in?

Together with Bruno Moreschi, we developed a practice of "living with images." It started in 2020, during the pandemic, when we began taking images out of large-scale computer vision data sets and printing them as postcards. We sent sets of three images to over a hundred people around the world — art historians, engineers, museum guards, artists, programmers, microworkers, activists — and asked them to live with the images for at least a week, then respond however they wanted.

Printing them proved to be quite important to scale down. These are images that exist as data: scraped from the web, stripped of their context, tagged or annotated by underpaid microworkers, and fed into AI training pipelines. They are operational images, in Farocki's (2000) sense: not made for human eyes but for machine processes. Printing brought these images back into conversation. A printed image has weight, it has edges. You can pin it to your kitchen wall, use as a bookmark, take it on a walk. You can also burn it, as one of our participants did.

Data sets take situated images — someone's photograph, taken somewhere, for some reason — and remediate them into interchangeable data points. Printing doesn't recover their "original" meaning (that's not the intention after all), but it restores its conditions for singular encounters. The image becomes a thing in the world again, open to accumulating new meanings through proximity and time. The responses we received from people revealed how much meaning these images carry once they are returned to everyday life — something we discuss in detail in our article. What interested us next was whether this could work as a teaching practice.

We adapted this practice of living with images to our teaching.Inspired by critical pedagogy's idea that learning happens through reflective action (a praxis of critical conscientization), the idea is not just to talk about data sets in class, but to get students to experience them. 

The structure is simple, and we detail it in full in our teaching guide. In a first session, students are introduced to data sets. Then, they receive printed data set images and write down their first impressions. Next they take the images home and are asked to live with them: place them somewhere visible, carry them around, look at them over several days, and take notes as their impressions shift. They are also asked to prepare a "response" in any (creative) format they wish.

Some guiding questions for the students: What do you think is the story of your image(s)? How would you tell them to your friends and family? Does this story change as you look more and more at this image?How do you think your three images relate? What images from your life would you like to put next to them? What relations emerge from the connection between these images and new images?How do you think these images have been used before? Where do you think they come from, and how do you think they are to be used by algorithms? How do you think an algorithm would understand it, and why?

Students can also explore how their perception of these images is different from an algorithm. As a student explained: "I took pictures of each physical print and asked Gemini: “what do you see in this picture?” The results were surprising to me because AI actually described the images in ways similar to how I would when I first saw them, but it also indicates details that I did not notice, such as camera angle and composition. Yet, something interesting pops up is that when I look at these photos, I tend to imagine colors or even moving images, even though they are printed in black and white. However, for every output, Gemini started with something like: “this is a black-and-white image” or “a grainy, black-and-white photograph”. Perhaps this shows that AI systems cannot imagine pictures in the same way as humans, much as explained by Offert and Phan (2024); instead, it identifies features associated with textual descriptions and generates outputs based on statistical probability."

After spending some time with the three images, they may start to connect and become a new "data set". Below, a poem created by a student XYZ  that explored this:  camel is chilling in the sun,triangular ancient pyramid looking fun.Giza, Saqarra or pyramid of the sun, honestly i cant tell which one.flashing the light, feeling crazy,party night is waiting on this lazy.moutwash to feel fresh, quick selfie to remember this mess.still, dead, no breath, fishies waiting on their fate.cold as ice, full of silence, fishies never swim on islands.  

In a second session, students share their experiences and responses. Some have cut up their images and scribbled on them. Others have written poems or lost the images altogether.The conversation then opens onto bigger questions about how data sets are built, what gets lost when images are annotated, and who these images belonged to before they became training data.This can be used to introduce a longer lecture on data sets or the  networked/digital image. Or else tied to a longer discussion on readings they may have done on this topic (suggestions included in the teaching guide).

What I've found in using this teaching exercise is that students arrive at complex insights through their own experience.Some discover how personal their relationship with photographs actually is, how images carry affect and memory in ways that resist easy categorization.Others get curious about provenance: where did their image come from? Who took it? Did they know it ended up here?The exercise creates conditions for interesting questions to surface, while supporting students' agency and creativity in exploring different facets of the issue.

To support this practice, I coded a small, open-source tool. It does one simple thing: picks random images from a selection of eight computer vision data sets and generates a printable PDF. You choose how many pages, and it puts them together — ready to print and hand out.

The tool is deliberately modest. In a field obsessed with ever-larger models and ever-bigger data sets, there's something worthwhile about making a small thing that just helps you slow down and zoom in. It doesn't analyze or classify or generate images. It just randomly places them on pages and hands the interpretation back to people.

It's also designed to be easy. No technical setup or infrastructure are needed. Just input a number, download a PDF, print it, and distribute the images. The friction of printing and physically handing out the pages is the point: it's what takes the image from data back to material.

The tool is open source on Github. Feel free to remix it, expand it, or build something entirely new with it.

Data sets work by making images interchangeable — any beach can be "beach," any face can be "person." Scale thinking crucially depends on the individual being indistinguishable as seen through the whole.By living with images we realize how they contain multitudes: their multifacted meanings can't be generalized or aggregated.

The pedagogical goal is not rejecting data sets or proposing to go back to a pre-digital way of seeing. Rather, our hope is to create productive pedagogical frictions that make visible what the scale of algorithmic seeing seeks to hide.Every image in a data set was once someone's photograph. The passage from personal image to training data is not a neutral shift. Every form of data collection and interpretation implies consequential choices about what matters, what gets kept, and what gets flattened away.

Ultimately, if critical pedagogy teaches us anything, it's that how we teach matters as much as what we teach. The same is true for algorithms. Current AI systems are trained to see through a flat, mechanical, top-down process: images are collected, tagged, and seldom questioned.Living with these data set images is also, then, a speculative exercise: could the slow, careful, situated ways we teach students to see images help them imagine entirely different forms of algorithmic seeing?

Hanna, A., & Park, T. M. (2020). Against Scale: Provocations and Resistances to Scale Thinking (arXiv:2010.08850). arXiv. http://arxiv.org/abs/2010.08850Offert, F., & Phan, T. (2024). A sign that spells: Machinic concepts and the racial politics of generative AI. Journal of Digital Social Research, 6(4). https://doi.org/10.33621/jdsr.v6i440462