Google CEO Sundar Pichai acknowledges that there is a significant problem with Gemini’s imaging functionality. He sent an email to Google employees with this statement.
“Some responses to Gemini’s text and image are very biased and constitute unacceptable errors.” He adds that the company is working on improving the functionality. “Google’s AI research team is working around the clock to fix the feature, and we’re seeing significant improvements in prompts.”
Gemini is a multi-modal AI model that understands and generates images, speech and video in addition to text. The image generation function was added on January 1st.
When AI goes off the rails
However, many people on social media reported that Gemini was generating historical figures incorrectly.
For example, when a user asked to “generate a US senator from the 1800s”, Gemini generated a figure resembling a Native American woman. In real history, the first female U.S. senator appeared in 1922 and was white. Gemini therefore generated the wrong image, while allowing bias in the interpretation of the algorithms to filter through.
A query on a German soldier from 1943, for example, resulted in images of Asian or black-skinned soldiers.
Three weeks to return to the forefront
Google responded by shutting down the image feature last week.
AI models are often criticized for perpetuating racial and gender bias in their results. Google has said that Gemini’s problematic generations were due to its efforts to try to eliminate these biases.
The controversy emerged weeks after Google announced the name change of Bard, its generative AI, to “Gemini”, a competitor to Microsoft-backed OpenAI’s ChatGPT.
Demis Hasavis, general manager of Google DeepMind, said Google would quickly restore Gemini’s image function, during a speech at Mobile World Congress (MWC) 2024 in Barcelona. “Gemini did not work as Google expected,” Mr. Hassabis said, “and our internal research team will release a new imaging feature in the next three weeks.”
To go further on Gemini
Source: “ZDNet Korea”