I started thinking about the intimacy of datasets during the pandemic. The screen became our primary space of connection. Video and photo apps exploded in popularity. Many of them, like Lensa, were powered by generative AI models trained on vast collections of photographs.

My work as an artist is grounded in photography as a social practice, as something that we participate in. I was struck by the idea that so many of us were already gathered in these datasets, already colliding and collaborating to create the tools sold by these companies.

This became the premise of my project We are already gathered, a series of participatory workshops that I started in 2022. I reached out to more than 200 people whose images I found in LAION-5B, an open source dataset used to train generative AI models like Stable Diffusion and Google Imagen. We exchanged stories about our photographs, our feelings about AI, our doppelgangers and our dreams for tech justice. For the first time in a long time, it felt thrilling to be online.

My gathering was an art project, but there are many ways to get together. Afterall, we’ve already done a lot together. Our faces have been used as the foundation for entirely new images. Together, we have been a part of a technical infrastructure beyond our imaginations. Together, we have had our images used without our consent or knowledge. Together, we’ve made money for someone else. You could think of the other strangers in the dataset as neighbors, colleagues, collaborators. Maybe we want to be co-conspirators, friends, comrades.

Consider this article an invitation to have your own dataset gathering.

Set an intention.

When tech companies scrape the internet to make datasets, they are organizing us. In The Art of Gathering, Priya Parker writes that, “gathering—the conscious bringing together of people for a reason—shapes the way we think, feel and make sense of our world.” AI datasets bring people together for a reason—to make money or build better language models—just not one that typically serves the people contained within them.

Since we are already a part of these billion dollar technologies, how can we hijack the purpose of our gathering for our collective good?

Find yourself and meet strangers.

Let’s focus on datasets that contain images of people (us!) and that are used to train generative AI models. Most of these datasets are not accessible to the public, even if the public is contained within them. For instance, OpenAI’s datasets are private, as are Meta’s and Apple’s. As of this writing, the largest open source dataset of image-text pairs is LAION-5B (an acronym for “Large-scale Artificial Intelligence Open Network,” with “5B” referring to the dataset’s 5.85 billion image-text pairs). Maintained by a German nonprofit organization, the dataset was compiled by Common Crawl, another nonprofit that uses bots to systematically browse the internet and archive its contents.

Peek inside the LAION-5B dataset by visiting their search demo. Try typing your name into the search bar or uploading an image of yourself. If there isn’t an image of you in the dataset, it will show you results that are visually similar—your doppelgängers, so to speak.

Scroll through the search results until you find an image that has an email address in the text description. Common Crawl collected the information in LAION-5b by scraping associated images and texts. Because of this, it’s not uncommon for personal emails or websites to be listed alongside an individual’s image— for instance on a professional portfolio or work website. Take a moment to reflect on how it feels to see all of this information together.

Send an invitation.

By reaching out to someone else in the dataset, you transform it into a social network. Through your invitation, you position yourself and others in the dataset as people who have a relationship to each other and a shared stake in something that you both contribute to.

In my outreach, I introduced myself, explained where I found the recipients’ image, shared their image with them, and invited them to participate in my project.

Share knowledge, build trust, stay flexible, be transparent.

I found myself corresponding with people on every continent besides Antarctica, most of whom had little knowledge of AI. Reactions ranged from anger and dismay to delight and curiosity to confusion and introspection.

Some who replied wanted to participate, and others wanted to ask questions or give input. No one had any idea that their image was in LAION-5B, and most didn’t know what it was. While those who live in Europe had some options to pull their image from the dataset under the GDPR, the rest of us didn’t have formal channels of action available to us.

Have a gathering.

A Workshop

For my gathering, I was interested in using the project to return some kind of agency to the people whose images had been collected into LAION-5B. The project’s image donors were openly invited to participate, and, if they were interested, they were able to set the terms for the use of their image and ask for something in return. Consider what your reason is for bringing people together.

I asked those who wished to take part two things:

1. “Please describe who you would like to wear the mask that has your image”
2. Provide “a prompt of what you would like the group to ‘generate' for you.”

I was riffing off of the idea that when people use generative AI apps, they are, in a sense, producing images that contain very, very tiny pieces of the faces of everyone in the dataset.

During each gathering, strangers came together to choose masks and collaborate to create responses to these prompts. The gatherings became exchanges between the people whose images were in the dataset—dubbed the “image donors”—and those who wore those images. Together, we modeled an alternative kind of AI dataset- one that centers consent, care and exchange.

Potluck

Generative AI datasets take something that might seem small from each of us in order to accumulate something vast that is stored seemingly indefinitely. Similarly, when we bring dishes to a potluck meal, we each contribute something small, and together we make a feast. But then we eat, we nourish each other, and the thing that we brought is consumed and disappears. Invite the people you meet in the dataset to bring a dish to introduce themselves.

Protest

It’s possible that when you connect with others in the dataset, they will be upset—angry, confused, sad, or overwhelmed. Maybe you’ll want to get organized. Maybe you’ll want to learn more about the laws governing AI and the impact of AI on your community. Maybe you’ll talk about the work of tech justice organizers like the Tech Workers Coalition, or maybe you’ll find local coalitions fighting unethical uses of AI, like the Trust Coalition that I worked with in San Diego. Maybe you’ll join one of these groups or start your own.

After your gathering, think about the rest of the iceberg.

In the United States, we typically have no idea what datasets we are a part of, few methods of consenting (or not) to being a part of those datasets, no say in what those datasets generate and dubious ways of benefiting from them. Some companies give us limited means of opting out.

LAION-5B is just one of the countless datasets that we are a part of, and it is one of the very few that we can examine. We are already gathered in datasets that are maintained by employers, schools, governments, doctors, and businesses small and large. The potential of some of these dataset gatherings to build community and reclaim agency is thrilling.

For instance, what could happen if all the people who Target thinks are pregnant got together, or if all the people who police department’s scanners have marked suspicious got organized? These gatherings can, for the most part, only live as thought experiments in our minds. Most of these datasets are inaccessible to us, even if our inclusion in them has a bearing on our lived experience.

In the interim, let’s keep sending invitations.

Acknowledgements: Warm thanks to all the image donors and participants who came together for We are already gathered, to Nicole Kaack, who invited me to put it together as a part of her Body Double series and has been a co-conspirator all along, and to ann haeyoung, for her considered engagement and guidance.