Illustration by TIME; reference image courtesy of Linda Dounia Rebeiz

When the Senegalese artist Linda Dounia Rebeiz types “buildings in Dakar” into OpenAI’s text-to-image model DALL-E, it returns squat, decrepit low-rises covered in dirt and peeling paint. They are unrecognizable to Rebeiz compared with the vibrant architecture she sees every day that fills the Senegalese capital.

This is just one example of why Rebeiz, 29, rarely ever uses large-scale models like DALL-E or Midjourney. She finds them inflexible, rudimentary, and riddled with biases that reinforce stereotypes or misconceptions, especially when it comes to images of the Global South. “With DALL-E, it seems impossible to get around the bias and the issues,” she says. “You just had to endlessly reprompt, and it wasn’t working.”

Instead, Rebeiz mostly creates her art with generative adversarial networks (GANs), neural-net architectures that allow her to train AI carefully on her own datasets. Rebeiz has taken pictures of hundreds of images of Senegalese flowers and historical buildings and scanned many more from national archives that had yet to appear anywhere online, and made them into public datasets for others to build upon. Using these novel yet historical datasets, she has created several striking projects, including Once Upon a Flower, which simulates how humans might perceive floral images once global warming has killed off actual flowers.

Rebeiz has also taken a leadership role in encouraging other Black artists to use GANs and participate in a new frontier that currently gets so much wrong about their culture. This summer, she curated a group show on the digital art gallery Feral File featuring 10 Black artists working in AI. “I have to put my drop in the ocean,” she says. “Even if it’s an infinitesimal difference, there’s still a sliver of hope that we can figure out ways of making the data something we can interrogate and change.”

More Must-Reads from TIME

Contact us at

Sougwen Chung
Yann LeCun
Abeba Birhane
Audrey Tang