Google said Thursday that its Gemini chatbot's image generation tool created “diverse” images that were not historically or factually accurate, including images of black Vikings, female popes, and Native Americans among the founding fathers. The company announced it would be “temporarily suspended” after it was widely criticized for its actions.
Social media users called Gemini “ridiculously woke” and “useless” after a request to create a representative image of the subject resulted in bizarre revisionist photos he accused.
“We are already working to address recent issues with Gemini's image generation capabilities,” Google said in a statement published on X. “While we do this, we will pause human image generation and re-release an improved version.” We will update it soon. ”
Examples include an AI image of a black man that appears to represent George Washington in a white powdered wig and Continental Army uniform, and an AI image of a black man that appears to represent George Washington in a white powdered wig and a Continental Army uniform, even though all 266 popes in history have been white men. These include an AI image of a Southeast Asian woman dressed as a pope.
In another shocking example revealed by The Verge, Gemini even produced “diverse” representations of Nazi-era German soldiers, including an Asian woman and a black man in 1943 uniforms.
Google hasn't published the parameters that control the Gemini chatbot's behavior, so it's difficult to get a clear explanation as to why the software fabricated different versions of historical figures and events.
“In the name of anti-bias, real bias is being built into the system,” William A. Jacobson, a Cornell University law professor and founder of the watchdog group Equal Protection Project, told the Post. Told.
“This is a concern not only for search results, but also for real-world applications where testing a 'bias-free' algorithm actually builds bias into the system by targeting end results that reach a quota.”
Fabio Motoki, a lecturer at Britain's University of East Anglia and co-author of a paper published last year, said the problem lies in Google's “training process” for the “large language models” that power Gemini's image tools. There may be. Noticeable left-leaning tendencies in ChatGPT.
“Reinforcement learning from human feedback (RLHF) recognizes that humans tell a model what is good and what is bad, and actually shape the model's 'reward' function, or more precisely, its loss function. Don’t forget,” Motoki told the Post.
“So depending on what kind of people Google hires or what instructions Google gives them, this could be an issue.”
This was a major misstep for the search giant, which had just rebranded its main AI chatbot from Bard earlier this month and introduced much-touted new features including image generation.
The blunder also comes days after OpenAI, the company behind the popular ChatGPT, introduced a new AI tool called Sora that creates videos based on users' text prompts.
Google previously acknowledged that it needed to fix the chatbot's erratic behavior.
“We're working to immediately improve this kind of depiction,” Jack Kraczyk, Google's senior director of product management for Gemini experiences, told the Post.
“Gemini's AI image generation generates a wide variety of people. This is generally a good thing since people all over the world are using it. But here it misses the point.”
The Post has reached out to Google for further comment.
When asked by the Post to provide reliability and safety guidelines, Gemini acknowledged that “due to technical complexity and intellectual property considerations, they are not public.”
In its responses to prompts, the chatbot acknowledged “criticisms that Gemini prioritized forced diversity in image production, which may have led to historically inaccurate depictions.” He also acknowledged that.
“The algorithms behind the image generation model are complex and are still being developed,” Gemini said. “They may have a hard time understanding the nuances of historical context and cultural expressions, which can lead to inaccurate output.”