Google apologizes for “missing the mark” after Gemini generated racially diverse Nazis
Google apologizes for “missing the mark” after Gemini generated racially diverse Nazis
Google says it's temporarily suspended the ability of Gemini, its flagship generative AI suite of models, to generate images of people while it works on Google says it's temporarily suspended the ability of Gemini, its flagship generative AI suite of models, to generate images of people while it wor...
I read the article yesterday and have been exposed to it a dozen times since, but I still laugh every time I see the phrase “racially diverse nazis”
What’s funny about this is that it IS, in fact, at least somewhat historically accurate: https://en.wikipedia.org/wiki/Free_Arabian_Legion
While they did wear separate uniforms from the Nazis, there were in fact both Blacks and Arabs fighting on the side of the Nazi regime in WW2. Asians too, of course, since Japan was allied with Germany.
Pretty sure these tools are often seeded with prompts that enforce diversity. Bing does the same or similar. I'm more amused by this, as the process isn't aware and can't actively enable or disable these settings.
To actively fit a historical prompt, it would need to not only consider images from the period, but also properly synthesize historical data to go with the prompt.
Yes, I saw some talk and a screenshot somewhere that showed that apparently in its current state, Gemini can (or could) be asked to output the prompt enhancements it used along with the generated images.
The screenshot showed someone asking for images of fruit, and the enhanced prompt included "racially diverse groups of people". Now if they're inserting something like that even for images containing no people at, it stands to reason that this is just a default enhancement they ALWAYS apply, no matter the prompt, which would explain the racially diverse Nazis (and all the other brouhahahas we've seen from them).
That's really what I'm expecting. My guess is that the training data is skewed, and the prompt cannot adjust.
Either the machine will need to understand what is expected, or the company will need to address this and allow people to enable or disable diversity.
The first option may be impossible to attain at this stage. The second can lead to inappropriate images.
I feel some variant of Conway’s Law comes to play with AI and these biases in training sets and that there’ll be no way to address it without first addressing the biases in society.
Racially diverse nazis
That tracks for google
This is the best summary I could come up with:
Google has apologized for what it describes as “inaccuracies in some historical image generation depictions” with its Gemini AI tool, saying its attempts at creating a “wide range” of results missed the mark.
The statement follows criticism that it depicted specific white figures (like the US Founding Fathers) or groups like Nazi-era German soldiers as people of color, possibly as an overcorrection to long-standing racial bias problems in AI.
Over the past few days, however, social media posts have questioned whether it fails to produce historically accurate results in an attempt at racial and gender diversity.
The criticism was taken up by right-wing accounts that requested images of historical groups or figures like the Founding Fathers and purportedly got overwhelmingly non-white AI-generated people as results.
Image generators are trained on large corpuses of pictures and written captions to produce the “best” fit for a given prompt, which means they’re often prone to amplifying stereotypes.
“The stupid move here is Gemini isn’t doing it in a nuanced way.” And while entirely white-dominated results for something like “a 1943 German soldier” would make historical sense, that’s much less true for prompts like “an American woman,” where the question is how to represent a diverse real-life group in a small batch of made-up portraits.
The original article contains 766 words, the summary contains 211 words. Saved 72%. I'm a bot and I'm open source!
What a bunch of ever-identical bullshit...