Sundar Pichai admits ‘unacceptable’ errors in Gemini after it produces controversial imagery

In an internal memo published by Semafor, Pichai wrote that “no AI is perfect” but said the company would continue refining Gemini “for however long it takes.” He remained upbeat about Gemini’s future, saying the team has already seen “substantial improvement” in many areas.

The image generation capabilities of Gemini, formerly known as Bard, have been paused since users noticed it was churning out historically inaccurate pictures, like images of Vikings and Founding Fathers depicted as people of colour. This sparked backlash on social media, with many critics decrying the “woke” programming.

Google SVP Prabhakar Raghavan clarified this was not intentional wokeness on the company’s part but rather a series of tuning errors. He revealed in a blog post published February 23 that, essentially, Gemini was calibrated to allow for diversity in generated images but failed to exclude inappropriate contexts for that diversity.

Over time, Raghavan explained, Gemini also became more cautious, sometimes refusing reasonable prompts out of an abundance of sensitivity. This aligns with reports that Gemini declined to generate images of white people in some cases.

Gemini’s image generator will return after fixes, Google says, including “structural changes, updated product guidelines, improved launch processes, robust evals and red-teaming.”