There’s been much hilarity this week about Google’s Gemini and its allergy to images of white people. They’ve had to pause marketing of the image generator.

Someone asked Gemini to explain how it works and it did. Basically, when you ask Gemini to make a picture, your prompt first goes to the language Model, which is instructed to insert words like “diversity” into your phrase before passing it to the image generator.

This is an extremely dumb and crude method of getting the result they thought they wanted. There’s no attempt to distinguish between plausible diversity and historical impossibilities, which leads to hilarious images like a racially diverse Waffen SS. And my personal favorite: a 17th Century English king eating watermelon (Gemini couldn’t recognize the racial stereotype because the model itself had injected race).

I can’t get over how crude this bodge is. The irony is that if they hadn’t messed with it, it looks like they had a pretty good image generator on their hands. They have damaged their product, perhaps fatally.