Google CEO Sundar Pichai addressed the company’s recent issues with its AI-powered Gemini image generation tool after it started overcorrecting for diversity in historical images. He called the turn of events “unacceptable” and said that the company’s “working around the clock” on a fix, according to an internal employee memo published by Semafor.

“No AI is perfect, especially at this emerging stage of the industry’s development, but we know the bar is high for us and we will keep at it for however long it takes,” Pichai wrote to staffers. “And we’ll review what happened and make sure we fix it at scale.”

Pichai remains optimistic regarding the future of the Gemini chatbot, formerly called Bard, noting that the team has already “seen substantial improvement on a wide range of prompts.” The image generation aspect of Gemini will remain paused until a fix is fully worked out.

This started when Gemini users began noticing that the generator began cranking out historically inaccurate images, like pictures of Nazis and America’s Founding Fathers as people of color. This quickly became a big thing on social media, with the word “woke” being thrown around a whole lot.

Prabhakar Raghavan, Google’s senior vice president for knowledge and information, did not lay the blame on wokeness, but rather a series of tuning errors. Basically, the model was fine-tuned to allow for diverse groups of people in pictures, but “failed to account for cases that should clearly not show a range.” This led to controversial images like people of color showing up as Vikings and Native American Catholic Popes.

Raghavan also said that the model became more cautious over time, occasionally refusing to answer certain prompts after wrongly interpreting them as sensitive. This accounts for reports that the model refused to generate images of white people.

It sounds like the company was trying to both please a global audience and ensure the model didn’t fall into some of the traps of rival products, like creating sexually explicit images or depictions of real people. Tuning these AI models is extremely delicate work and the software can easily be led to make ridiculous errors. It’s what they do. In any event, I’d prefer a historically inaccurate Catholic Pope over unexpected violent imagery any day of the week. Chalk this up as yet another reminder that AI still has a long way to go. 

As for Gemini, the company promises the image generator will return in the near future, but it still requires a suite of fixes and tests to make sure this never happens again, including “structural changes, updated product guidelines, improved launch processes, robust evals and red-teaming and technical recommendations.”

Google CEO Sundar Pichai addressed the company’s recent issues with its AI-powered Gemini image generation tool after it started overcorrecting for diversity in historical images. He called the turn of events “unacceptable” and said that the company’s “working around the clock” on a fix, according to an internal employee memo published by Semafor.

“No AI is perfect, especially at this emerging stage of the industry’s development, but we know the bar is high for us and we will keep at it for however long it takes,” Pichai wrote to staffers. “And we’ll review what happened and make sure we fix it at scale.”

Pichai remains optimistic regarding the future of the Gemini chatbot, formerly called Bard, noting that the team has already “seen substantial improvement on a wide range of prompts.” The image generation aspect of Gemini will remain paused until a fix is fully worked out.

This started when Gemini users began noticing that the generator began cranking out historically inaccurate images, like pictures of Nazis and America’s Founding Fathers as people of color. This quickly became a big thing on social media, with the word “woke” being thrown around a whole lot.

Prabhakar Raghavan, Google’s senior vice president for knowledge and information, did not lay the blame on wokeness, but rather a series of tuning errors. Basically, the model was fine-tuned to allow for diverse groups of people in pictures, but “failed to account for cases that should clearly not show a range.” This led to controversial images like people of color showing up as Vikings and Native American Catholic Popes.

Raghavan also said that the model became more cautious over time, occasionally refusing to answer certain prompts after wrongly interpreting them as sensitive. This accounts for reports that the model refused to generate images of white people.

It sounds like the company was trying to both please a global audience and ensure the model didn’t fall into some of the traps of rival products, like creating sexually explicit images or depictions of real people. Tuning these AI models is extremely delicate work and the software can easily be led to make ridiculous errors. It’s what they do. In any event, I’d prefer a historically inaccurate Catholic Pope over unexpected violent imagery any day of the week. Chalk this up as yet another reminder that AI still has a long way to go. 

As for Gemini, the company promises the image generator will return in the near future, but it still requires a suite of fixes and tests to make sure this never happens again, including “structural changes, updated product guidelines, improved launch processes, robust evals and red-teaming and technical recommendations.”