'Black Joseon Dynasty General' in Hanbok... Google Faces Humiliation in Less Than a Month [AI One-Bite News]

Google Gemini Halts Faulty Image Generation
Excessive Diversity Consideration... "Unacceptable Error"
Korean Startup Meeting Zuckerberg "Proposes Multilingual Versions"
Enhancing Language Diversity and Accuracy Can Differentiate Models

A Joseon Dynasty general with a Black appearance wearing hanbok, and an 1800s American senator resembling a Native American woman.

Google has temporarily suspended the image generation feature of its artificial intelligence (AI) model 'Gemini' due to errors that generated historical figures as people of color incorrectly. Google added the image generation feature to Gemini on the 1st of this month, but faced embarrassment in less than a month.

Sundar Pichai, Google’s CEO, seems to have taken it personally as well. He sent an email to all employees stating, "Our generative AI caused discomfort to users and highlighted biases," adding, "This is an absolutely unacceptable error and our mistake," strongly reprimanding the issue.

[Image source=X]

In fact, this is not the first time a generative AI model has made such errors. Since the data AI learns from is biased, the outputs inevitably reflect that bias. Most big tech companies leading generative AI technology are American, so it is natural that the training focuses on English or English-speaking cultures. Even if other languages available on the internet are used for training, it is unlikely that bias can be completely eliminated. There are minority ethnic groups or cultures that do not exchange information via the internet.

Reflecting reality as is also problematic. If AI reproduces existing stereotypes in the world, those stereotypes only become more entrenched. For example, if you ask a generative AI to draw a person holding coffee and wearing sunglasses, it will depict a white woman; if you ask for a person selling food at a street stall, it will depict an Asian woman.

However, the nature of the errors has now changed somewhat. Errors have appeared where people of color or women are mechanically inserted without considering historical facts. When asked to draw an "1800s American senator," Gemini generated a figure resembling a Native American woman. The first female senator in the U.S. appeared in 1922 and was white, but the image was fabricated without considering historical facts. Even in cases where the image of a white male should clearly be prioritized, such as Vikings or Nazi German soldiers in World War II, the AI depicted East Asian women or people of color.

Google’s explanation is as follows: it was the result of pursuing maximum diversity. While this generally improved the quality of responses, sometimes it created errors by excessively prioritizing diversity over historical context.

Amid this, a domestic startup made an interesting proposal to big tech. On the 28th, Mark Zuckerberg, Meta’s CEO, visited Korea and met with startups including Upstage. Kim Seong-hoon, CEO of Upstage, who attended the meeting, proposed the idea of creating versions of LLaMA 3 in various languages. The approach involves selecting representative companies or developers by country to collaborate. Zuckerberg reportedly responded positively to considering this proposal.

If LLaMA 3 is released in multiple language versions, it will have a distinctive advantage in dominating the global AI ecosystem. Instead of mechanically injecting diversity, diversifying the training data itself could reduce errors. Improving response accuracy and reducing bias would itself become a competitive edge.

Meta’s next-generation language model LLaMA 3 is expected to be unveiled in July. The new version is said to provide improved answers to controversial questions. Whether Upstage’s proposal will materialize is unknown, but there is keen interest in whether an even more evolved model will emerge.