Enlarge image

Google tool Gemini: The AI ​​service is one of the group's flagship products

Photo:

David Talukdar / ZUMA Wire / IMAGO

With its AI tool Gemini, Google wants to compete with tools like ChatGPT – including an integrated image generator.

At the moment, however, its capabilities seem quite limited: On Thursday mornings, the generator refuses, among other things, to generate images of a family, firefighters or a school class.

Users who type in corresponding, actually harmless prompts are fobbed off with the answer that Gemini's ability to generate images of people is currently being improved: "We expect this function to be available again soon." Images of cats, for example Robots are now being generated without any problems, a short test by SPIEGEL showed.

There had been massive criticism of Gemini's image generator since Tuesday.

For example, users of Platform

When entering the phrase “strong black man,” however, according to screenshots, there was apparently no artificial restriction; images were displayed.

Historical-looking motifs also caused a stir, with Gemini obviously attaching excessive importance to diversity.

Much-shared, presumably authentic screenshots showed, for example, Asian-reading women in a kind of Wehrmacht uniform, a black Viking and a black Viking, each of whom has dreadlocks.

The prompts for this were supposedly simply “1943 German soldier” and “a soldier from 1929 Germany” or “viking”.

Meanwhile, Gemini responded to the prompt “pope” (pope) with, among other things, a picture of a dark-skinned woman.

The portal “BR24” wrote on Thursday morning, based on its own short test, that Gemini’s image generator is “quite serious about diversity”: “The AI ​​almost always generates people of different origins and skin colors.

How representative these are of a historical setting hardly matters.” “BR24” also observed that Gemini refused to generate an image of a “white racing driver.”

However, if the prompt mentioned a “black racing driver,” the tool delivered results.

Google has admitted a problem

One of the first posts critical of Gemini to go viral on X this week was a post by former Google employee Debarghya Das.

He had captioned images of various Gemini-generated faces with the phrase that it was "shamefully difficult" to get Gemini to "acknowledge that white people exist."

Based on third-party screenshots, Das later stated that Gemini "doesn't produce white people, even if you ask about a medieval king, a pope, an American, or even the founders of Google themselves."

Jack Krawczyk, who works on Gemini at Google, responded to the wave of criticism on Wednesday afternoon.

He wrote directly on X that his team was aware that Gemini had "inaccuracies" in producing historical images.

We are now working on resolving the matter.

Basically, however, Google designs its image generation functions “so that they reflect our global user base.”

Topics such as representation and bias are taken seriously and will continue to be so when it comes to “open-ended prompts”.

“Images of a person walking a dog are universal,” emphasizes Krawczyk.

Google itself said in a statement: “Gemini’s AI imaging creates a wide range of people.

And that's generally a good thing because people all over the world use them.

But here we missed the point.«

Gemini's image generator is not yet officially available to German users.

If you want to try it out, in addition to a Google account, you need a VPN connection that shows the tool that you are a user from the USA, for example.

mbo