Linda Dounia: Amplifying the Self

As new ways of making unfold, Notes on Creation reflects on the self-devised processes used by today’s most exciting generative artists.

The medium is the message, as the theory goes. For Linda Dounia, this should be applied to new technologies too.

Linda is fascinated by visual communication – how the means through which ideas are transmitted impacts the stories being told. Therefore, she applies a much-needed critical lens to artificial intelligence tools. Her primary issue is that the biased datasets that most AI tools are trained on reproduce the same bias in their output – whether that is because of the programmer’s prejudice, the breadth of the training data, or the way that they are created. The result is that “these tools are very bad at representing a majority of the world because they’re not bringing those unrepresented people and their contexts into the rooms where the technologies are being developed.”

Instead of that bias, Linda makes art that exposes those flaws. She works with a system called a generative adversarial network (GAN). A GAN is a type of machine learning system that can be trained to generate images based on the data it has been trained on. The level of control over the output is determined by three factors. Firstly, the size and nature of the inputted data set. Secondly, how many training steps there are included. Thirdly, the coded parameters the programmer sets to explore the map of outputs. This is important since she is “not totally seduced by the randomness of what a GAN can produce.” For her, generative systems are simply the vehicle to explore her own creativity.

Linda likens GANs to mini AI. Whereas AI tools like Midjourney have datasets in the billions, GANs work to the scale of thousands through transfer learning. For Linda, it is important that the data is personal, “whenever a GAN is present in a body of work it ties to something very personal to me.” The process of creating these inputs is therefore very labour intensive – for one project Linda handpainted two thousand vignettes. In this way, she inserts herself not just into the individual GAN but into the broader narrative of digital art.

If I'm using photographs from my family, it's an amplification of memory. If I'm using a data set of flowers that I grew up seeing and wonder where they are now. That's an amplification of my environment. So it's an extension of who I am when I use a GAN.

This amplification of the self has profound significance for Linda as a Senegalese woman, given that it is often African peoples and their knowledge that are erased from archives and datasets. By engaging with the GAN, Linda can recreate the world in her image. However she also feels changed by the process. “The first time you approach this tool you come as flesh, human, and soft, but they have their own intricacies and their own personality” – over time you adapt to the ways the machine learns.

Faux Letters

Faux Letters is an offshoot from a project Linda started in 2022. It began with the observation that most AI image generators are particularly bad at lettering. Linda also questioned the absence of African writing systems from the archive, as well as the privileging of Roman scripts. The project required extensive research into different writing systems across the world.

There’s so much beauty across the continent and language systems I didn’t know about that I found through this project and was inspired by. I was grateful to go to the edge of that curiosity and to have a GAN push me there.

Linda transformed this information into data by handwriting approximations of the many scripts – playing close attention to the geometry and flow of the letters. The final work represents the myriad ways humans have found meaning through shapes.