Google’s artificial intelligence (AI) model, Gemini, is facing widespread criticism on social media for inaccurate depictions of historical figures. For the Mountain View firm, it is essential to correct the flaws in its system.
With Gemini, the company wanted to catch up with OpenAI and Microsoft. The conversational agent, deployed a few days ago, is capable of generating images like its competitors, although this functionality is, for the moment, not available within the European Union. It is precisely this tool that is singled out by many Internet users.
By wanting to do well, the company created a new problem
Several examples of Gemini’s historical inaccuracies have been shared on X, formerly Twitter. For example, generative AI represented a “ German soldier in 1943 » by a black man. It also generated visuals of the Founding Fathers featuring white men, but also black or Native American…
These errors are mainly denounced by far-right sympathizers, who insinuate that Google wants to “ to erase » white people. The reality is much more nuanced. For several years, experts have been warning about discriminatory biases, whether racist or sexist, exhibited by AI. Studies have also highlighted this phenomenon, pushing publishers to act to counter it.
Some denounce the fact that Gemini generates black or Asian women to represent “ an American woman “. On the contrary, it is a real illustration of the diversity present in the United States, proving that Google has tried to find a solution. In an effort to prevent its model from amplifying racial bias, the company instead seems to have created another problem. By generating, for example, the image of a black woman to respond to a prompt asking her to draw a “ 1800s US senator ”, AI erases a real history of racial and sexual discrimination. The first female senator in the United States was white, and served from 1922.
Google makes its mea culpa
Google announced its error in a publication on X. “ We are working to immediately improve this kind of representation. Gemini’s AI image generation produces a wide variety of people, which is generally a good thing because people all over the world use it. But she misses the mark here », writes the company.
She made the decision to “ suspend the generation of people images ”, and is working on an improved version of his model.
Source : The Verge
23