Google’s AI, barely renamed, provokes a first controversy

When artificial intelligence loses control: on February 20, for several hours, chatGPT, OpenAI’s AI, experienced a major bug, providing incomprehensible responses. And for several days, Gemini, Google’s AI which has just changed its name, has been blamed for having generated several realistic images that are out of step with historical reality.

When artificial intelligence loses its marbles… Tuesday February 20, for several hours, chatGPT experienced a big bug, and since Thursday February 22, Google has deactivated the generation of images containing men or women, with its AI renamed Gemini , after the publication of certain visuals.

Google quickly apologized for these “inaccuracies” In “Gemini-generated representations of certain historical images”. In this case: a pope, Vikings and Nazi soldiers, all with black skin. “We missed the objective,” recognized the Mountain View giant on the X network. And yet, Google’s intention was laudable: since the beginnings of image-generative AI, ethnic diversity and different skin colors were abnormally under-represented.

Google, with Gemini which is only 16 days old, has therefore tried to compensate. Blessed bread for some of the most conservative Americans. They accuse Google, although known for its defense of diversity, of wanting “eliminate the blanks” images generated by Gemini. A not insignificant ideological and political trial, nine months before the American presidential election.

How can we explain the sometimes gross errors of these AIs?

By remembering, first of all, that artificial intelligence is not that intelligent. The results are, by nature, perfectible. We saw it with the beginnings of generative AI, and these hands with six fingers or worse. Google has also undertaken to rectify the situation immediately, but this controversy is obviously the business of competitors such as Firefly from Adobe, MidJourney or even DALL-E from Open AI.

The concern, with Gemini and its laudable effort in favor of diversity, is to have done it without nuance and not to use historical reality as a safeguard. For example, when asked for an image of elected officials in the US Senate in the 1800s, Gemini should not have produced images of black female senators, since the election of the first of them, Carole Mauseley Braun in Illinois, dates from almost 200 years later: it was in 1993.

Historical reality as a safeguard

With this controversy, Google is getting further bogged down in AI. Gemini has just succeeded Bard. New name, but same technology. Remember the launch of Bard a year ago, presented as a competitor to chatGPT. Google’s AI claimed, on the day of its official announcement, that the James Webb telescope had taken the very first photo of an exoplanet, with supporting animation. This was false, since the very first photo dates from 2004.

Since then, we suspected that Bard, the name, was doomed. Hence “Gemini”: new facade with – incidentally – a paid premium version at $19.99 per month, which was to embody a new beginning.