Agencies Encouraged by Innocean USA to Advocate for Diversified Artificial Intelligence Training Datasets

Agencies Encouraged by Innocean USA to Advocate for Diversified Artificial Intelligence Training Datasets

Rewritten Article:

A big problem that creative director David Mesfin encountered while working on his documentary, "Wade in the Water: A Journey into Black Surfing and Aquatic Culture," was the inaccurate AI results for images of Black surfers on Midjourney.

"Every time I made a query or a prompt, I'd end up with a white surfer, but with a dark skin tone. The data just wasn't right," Mesfin told ADWEEK. "Representation isn't just about inclusivity. Here, it's about accuracy."

Mesfin's experience led INNOCEAN USA to join Breaking Bias, an initiative launched by diverse stock image database POC Stock to tackle AI bias by creating datasets that authentically represent BIPOC (Black, Indigenous, and People of Color) communities.

With INNOCEAN USA as a driving force, more than 22 agencies, including FCB Global and David&Goliath, have united under the cause, eager to rectify AI's lack of authentic representation of BIPOC communities.

"AI is only as good as the data it's trained on," Steve Jun, CEO of INNOCEAN USA, told ADWEEK. "When they explained that a simple action could help resolve a complex issue, we knew backing this project was the right move."

As part of the Breaking Bias initiative, participating agencies are using POC Stock images alongside comprehensive metadata to help retrain AI models, ensuring a more accurate representation. The images are meticulously reviewed by a human curator, and the metadata is then fine-tuned to maintain authenticity and prevent biased outcomes.

"This includes annotating race, ethnicity, gender, and cultural fluency," commented DeSean Brown, chief relationship officer of POC Stock. He added that the process includes multiple rounds of review prior to the images being approved for AI training.

David Angelo, founder and creative chairman of David&Goliath, emphasized that when it comes to accurate representation of BIPOC communities, AI "still lacks authenticity and often plays into stereotypes," but he believes that collective action can help AI self-correct.

AI bias has long been a concern in the advertising industry, with organizations like Lerma and Alma sounding alarms about stereotyping in AI tools like Midjourney in 2023. In 2024, The Brandtech Group introduced Bias Breaker, a tool designed to correct identity factors like race, gender, and age in AI content. That same year, Mastercard launched Mastercard Small Business AI, a generative AI chatbot addressing the needs and perspectives of multicultural and minority business owners.

The Breaking Bias coalition plans to keep expanding its dataset, focusing on underrepresented groups in the coming months. "We're looking beyond the U.S.," Mesfin explained. "AI is used all over the world, and it needs to reflect that diversity."

As the Breaking Bias initiative grows, Mesfin hopes it will extend beyond the current group of agencies to more industries and communities.

"This effort will require more than just the few agencies we currently have, and even including AI platforms like Midjourney and DALL·E," he said. "Ideally, they would actually partner with us and help us assist them."

"We need more agencies to tap into their passions, join this effort, and help this evolving tool mirror the real world," Jun added.

Beyond Breaking Bias, other efforts like Microsoft's Inclusive Design Toolkit, which incorporates AI to understand diverse user perspectives and recognize various forms of bias, are helping combat AI bias at the system design stage. Google researchers have developed tools like What-If, allowing product teams to test AI models against hypothetical scenarios to identify biases related to feature importance and fairness. Ultimately, industries are becoming increasingly aware of the unique challenges AI bias poses, with strategies that emphasize retraining AI with diverse, representative datasets, human oversight, metadata accuracy, transparent model evaluation, and collaborative efforts across organizations to share resources and knowledge.

1. David Mesfin, recognizing the inaccurate AI results for images of Black surfers, initiated a conversation about AI bias with INNOCEAN USA.
2. INNOCEAN USA, driven by the need for authentic representation, joined Breaking Bias, an initiative to tackle AI bias, producing datasets that authentically represent BIPOC communities.
3. David Angelo, emphasizing the need for diversity and inclusion in AI, believes collective action can help AI self-correct, eliminating stereotypes and ensuring a more accurate representation.
4. Beyond Breaking Bias, Microsoft's Inclusive Design Toolkit incorporates AI to understand diverse user perspectives and Google researchers have developed tools like What-If to identify biases in AI models, promoting diversity at the system design stage.
5. As the Breaking Bias initiative expands, it aims to collaborate with more industries, agencies, and AI platforms like Midjourney and DALL·E to create an AI technology that reflects the diverse world.

Read also: