Gemini AI is introducing a new image generation model called Imagen 3, months after Google disabled its previous model due to its controversial output. The search giant announced on The Keyword yesterday that “Imagen 3 offers advanced image generation capabilities with built-in safeguards and adherence to our product design principles.”
Google's Gemini image-generating feature, Imagen 2, drew attention in February for its wildly inaccurate portrayals of historical figures. The company had hoped that its AI image-generating tool would avoid the pitfalls of bias and portraying real people. But in an effort to be more inclusive, the tool it created ended up gravitating too far to the other end of the spectrum, becoming biased against lighter-skinned people.
The issue was made public in a post by X-user “End Wokeness,” who showed that a search for “portraits of America's Founding Fathers” returned a Native American man, an Asian man, and two dark-skinned men. The image generator was so inaccurate that it ended up depicting people of color as Nazi soldiers.
Google AI on the Founding Fathers, Vikings, and the Pope: pic.twitter.com/lw4aIKLwkpFebruary 21, 2024
New AI technologies must grapple with these issues, especially since these programs lack understanding of human history and ethics. Because of this, AI tools such as Microsoft's Copilot Designer can generate images that are unintentionally stigmatizing. Google attempted to solve this problem by making its image-generating tools avoid these cultural stereotypes and biases. However, one unintended consequence of this was that its Imagen 2 tool became too inclusive, historically inaccurate, and sometimes downright offensive. Due to this controversy, Google had to remove Imagen 2 three weeks after its release for a fix.
The company said it would relaunch its image generation tool within a few weeks, nearly 20 weeks after it was pulled. According to a blog post on The Keyword, Imagen 3 “does not support the generation of realistic, identifiable people, depictions of minors, or scenes that are excessively cruel, violent, or sexual.” That said, while acknowledging that its generative AI tool can make mistakes, Google said it will continue to improve it through user feedback.
Imagen 3 will be rolling out to Gemini users in the coming days, with English Gemini Advanced, Business and Enterprise users having first access. With technical improvements, better datasets and clear product principles after thorough testing, we hope to deliver better, more accurate and less jarring results.
