AI generators say there are no Black surfers. This group is out to change that

When David Mesfin was producing his documentary on Black surfing culture, Wade in the Water, back in 2023, he had a problem. Like millions of other people since ChatGPT and other GenAI tools emerged in late 2022, Mesfin was experimenting and using these tools to generate imagery for the film. “But the results were always the same: white surfers with darkened skin,” says Mesfin, a creative director at ad agency Innocean. “It was a clear sign that these systems weren’t built with us in mind. That moment made it impossible to ignore how deeply bias is embedded in the technology.”

This week, sparked by that moment, Mesfin and his colleagues have launched “Breaking Bias,” a collaboration with stock image firm Pocstock, and more than ​22 agency partners. As part of the project, 16 photographers donated their time to capture more inclusive images to help create what they’re calling the first-ever Ethical AI Guidebook—a framework that helps creators and developers make AI more like the real world.

Only 2.3% of AI-generated images of dentists featured Asians, despite the fact they make up 22% of the profession. About 9.6% of nurses are Latino, but 0% of AI-generated images reflected them. As more and more marketers utilize AI to create content, it’s crucial to build and maintain data sets that represent reality as accurately as possible. 

“The goal of the Ethical AI Guidebook is to spark a mindset shift—encouraging content creators, agencies, and AI developers to make inclusivity a core part of how they generate and use imagery,” says Mesfin. “This initiative is about more than just fixing flawed outputs; it’s about creating long-term behavioral change within the AI ecosystem. By offering practical guidance and real-world examples, we aim to help the industry build technology that reflects the full spectrum of humanity.”

Training challenge

So far, the project has created more than 96,000 images for Pocstock’s inventory, and the images will be accessible to companies that utilize AI image generation like Adobe, Amazon, Canva, Google, Microsoft, OpenAI, and Shutterstock.

The biggest challenge for Mesfin and project collaborators was the vetting process to make sure they selected photographers who not only had technical skill but also understood and were connected to the communities they were aiming to represent. 

“We weren’t looking for models; we were looking for real people, and that meant being intentional about casting talent that reflected the full spectrum of the demographic,” says Mesfin. “For example, when capturing Black surfers, we made it a point to include both men and women, with a range of skin tones from dark to light. We also prioritized having both male and female photographers involved to bring different perspectives to the work.”

Strategic approach

The challenge of tackling bias in AI models has been an ongoing issue for years. Pocstock cofounder and chief relationship officer DeSean Brown says that if the goal is for AI to produce outcomes that would appear more authentic and inclusive to humans there needs to be a finely curated and strategic approach to capturing excessive amounts of images and data. 

“We need to target specific people, communities, and actions, [then] capture and label the images accurately with cultural and technical nuance,” says Brown. “Many content creators and collectors may not have the experience, process, or resources to assess negative stereotypes and bias or the ability to label data in a culturally accurate way. Continued partnership among tech companies, content creators, stock and data collection companies can help the industry get to where it needs to be.”

Given the scale of the task at hand, Breaking Bias is an ongoing project that is always looking for new content partners to join who want to help build more representative data sets. “If you’re part of an AI company, we welcome your collaboration, too,” says Mesfin. “This is a collective effort, and the more voices we bring to the table, the stronger and more accurate the technology becomes.”

View the full article