Newsletter

The Ghost Stays in the Picture, Part 1

Archives, Datasets, and AI Infrastructures

Eryk Salvaggio

Eryk Salvaggio

— 1 min read
The Ghost Stays in the Picture, Part 1
Two black and white photographs of shadows mingling with sidewalks. On the left is a human photograph taken from Flickr, on the right, an image generated with Stable Diffusion 1.6. (Flickr image, “Life Goes On” by sinkdd, CC BY-NC-ND 2.0).

Today I’m sharing a link to my first blog post written as part of my Research Fellowship with the Flickr Foundation. The two posts look at the YFCC100M dataset, a collection of 99.2 million images released by Yahoo! (then the parent company of Flickr) for researchers in 2014.

That dataset, I argue, is a case study not only in how images in a dataset haunt the products of image generation tools like Stable Diffusion or Midjourney. But by serving as benchmarks and tools for calibration, datasets can also haunt infrastructures of Artificial Intelligence.

How did this dataset shape the automated decision making processes that were then included in longer, more complex systems of image generation? How does a case study like this reflect on our relationships with datasets, as opposed to our relationships with archives, and the ways both exist in unique tensions with the images and texts that we want to preserve?

You can read the first of two pieces today on Flickr.org.

Read It

Things I Am Doing This Week

Though I am not listed on the itinerary, on June 3 I’ll be part of a panel at ACM FAccT (remotely) called Generative AI Labor Impacts: A Time Capsule of Workers' Stories organized by Alex Hanna (DAIR), Tamara Kneese (Data & Society) Nataliya Nedzhvetskaya (UC Berkeley; Collective Action in Tech) Clarissa Redwine and Kristen Sheets (Collective Action in Tech) and Xiaowei Wang (UCLA/Center on Race and Digital Justice).

Read more