Google apologizes for ‘missing the mark’ after Gemini generated racially diverse Nazis::Google says it’s aware of historically inaccurate results for its Gemini AI image generator, following criticism that it depicted historically white groups as people of color.
Well, the tech is of course still young. And there’s a distinct difference between:
A) User error: a prompt that isn’t as good as it can be, with the user understanding for example the ‘order of operations’ that the AI model likes to work in.
B) The tech flubbing things because it’s new and constantly in development
C) The owners behind the tech injecting their own modifiers into the AI model in order to get a more diverse result.
For example, in this case I understand the issue: the original prompt was ‘image of an American Founding Father riding a dinosaur, while eating a cheeseburger, in Paris.’ Doing it in one long sentence with several comma’s makes it harder for the AI to pin down the ‘main theme’ from my experience. Basically, it first thinks ‘George on a dinosaur’ with the burger and Paris as afterthoughts. But if you change the prompt around a bit to ‘An American Founding Father is eating a cheeseburger. He is riding on a dinosaur. In the background of the image, we see Paris, France.’, you end up with the correct result:
Basically the same input, but by simply swapping around the wording it got the correct result. Other ‘inaccuracies’ are of course to be expected, since I didn’t really specify anything for the AI to go of. I didn’t give it a timeframe for one, so it wouldn’t ‘know’ not to have the Eiffel Tower and a modern handgun in it. Or that that flag would be completely wrong.
The problem is with C) where you simply have no say in the modifiers that they inject into any prompt you send. Especially when the companies state that they are doing it on purpose so the AI will offer a more diverse result in general. You can write the best, most descriptive prompt and there will still be an unexpected outcome if it injects their modifiers in the right place of your prompt. That’s the issue.
C is just a work around for B and the fact that the technology has no way to identify and overcome harmful biases in its data set and model. This kind of behind the scenes prompt engineering isn’t even unique to diversifying image output, either. It’s a necessity to creating a product that is usable by the general consumer, at least until the technology evolves enough that it can incorporate those lessons directly into the model.
And so my point is, there’s a boatload of problems that stem from the fact that this is early technology and the solutions to those problems haven’t been fully developed yet. But while we are rightfully not upset that the system doesn’t understand that lettuce doesn’t go on the bottom of a burger, we’re for some reason wildly upset that it tries to give our fantasy quasi-historical figures darker skin.