Google is providing additional details about the situation that led to the temporary halt in Gemini’s image generation of individuals.
This week, Jack Krawczyk, the Senior Director of Product at Google, mentioned on X that the Gemini team is focused on fine-tuning the AI model to deliver accurate historical results.
Following numerous user complaints, Gemini posted historical images showing people of different races in situations where they may not have been present. As an example, Gemini presented an image of a founding father who was African American and a Native American man and Indian woman to symbolize an 1820s-era German couple.
The solution ended up being more complex than the team initially anticipated, and on Thursday morning, Google disabled the feature that allowed generating images of individuals with the tool.
In a recent blog post, Google’s Senior Vice President Prabhakar Raghavan provided additional insights into the issues with Gemini. Raghavan mentioned that the tuning to ensure diversity in Gemini did not account for certain cases that should not show diversity. He also mentioned that over time, the model became excessively cautious and started refusing to answer certain prompts, even misinterpreting harmless prompts as sensitive.
Put simply, Gemini had been taught to represent various races, but was unaware that there were situations where it might not be suitable. It also had a different understanding of what could be considered inappropriate than was originally meant.
In the end, the model ended up “overcompensating in some situations and being overly cautious in others, resulting in images that were embarrassing and inaccurate,” he stated.
Raghavan mentioned that Google intends to implement significant changes before reactivating it, with a focus on an improvement process that will involve thorough testing.