Tangible results of decades of academic and political brainwashing. Eventually the AI models will coalesce around the groups of people that built and influenced them. Digital twins of thought groups so to speak. And there will be battles around which AI is the arbiter of truth. Not to mention “fact checking” meta AIs.
This isn't any different from the entirety of human history. The people in charge of the books write the books. It's why China does it's best to make the world forget about Tiananmen Square, or why in the early days of computing, only white men were shown doing the work despite the substantial amount of women in field (especially for things like ballistic calculations).
There is also a difference between error and purposeful misrepresentation.
Couldn't the same argument be made the other way? With your line of reasoning, why even bother finetuning a model for DEI? After all, it's just a generative model, not a world model.
Please don't take this comment as an ideological or political comment; I'm only trying to outline a flaw that I perceive in your reasoning. I would write a similar comment if we were discussing an apolitical, dry mathematical topic.
That is fair! I was being inflammatory with my DEI remark and could easily be viewed as an "ad hominem attack."
I would expect these models to misrepresent reality until we do the near-impossible: convince a large and diverse swath of people to understand and work with each other, work with ML models, and contribute to shared datasets that we can train our models with. I have vastly oversimplified what it would take to create an accurate world model, however, the main point is there. We need feedback from everyone. And likely, feedback from more than just people.
Why would Google/Adobe want to artificially increase diversity in its training set? Well, there is a lack of diversity in the training set in reality, so the developers are likely trying to balance something that can easily be seen as racist or prejudiced. Given America's history of how it treats people with different cultures, religions, or skin colors, this should not come as a surprise. Other countries have made similar errors in their ways throughout history.
Unfortunate side effects: mass creation of things like this posted everywhere, new models trained on those (because AI cant copyright its own works) and the cycle continues and strengthens its own faults
> I asked Firefly to create images using similar prompts that got Gemini in trouble. It created Black soldiers fighting for Nazi Germany in World War II. In scenes depicting the Founding Fathers and the constitutional convention in 1787, Black men and women were inserted into roles. When I asked it to create a comic book character of an old white man, it drew one, but also gave me three others of a Black man, a Black woman and a white woman. And yes, it even drew me a picture of Black Vikings, just like Gemini.
Why would we expect Gemini or Firefly to not be able to produce these images? I can definitely sympathize with not wanting to make offensive images but there’s no way these tools could be sanitized to not produce something offensive given the right prompts. Just like how you can’t stop photoshop users from creating similar images manually.
It's not like black soldiers for Nazi Germany are offensive. It's just stupid and everyone who knows how ML works knows that Google does a lot of stupid, often called “woke”, stuff to produce these images. Any other non-mutilated ML algorithm would just give you the more historically but also more common in the dataset version of white Nazis.
1. Gemini and Firefly are programmed (by a prompt) to give you "inclusive" results when you ask for depictions of people. So if you ask it to generate a crowd of people, while normally the AI could generate mostly or purely white crowd (perhaps because white people dominated the training dataset), with such a prompt it will try to give you people of both genders and all races.
2. Because of the above, in some context where you would expect only white people, for example the founding fathers of USA, it also depicts people of various races and genders. This is historically inaccurate and produces a lot of controversy, especially in times, when doing the reverse, depicting a historically non-white person as white, usually leads to some kind of outrage. And so such behavior is put to the same category as e.g. claiming "you can't be racist against white people".
3. And as a result, many people decide to "fight back" by trolling the creators of the tool, using the bias where instead of being negative towards white people, it is negative towards other people. So instead of taking positive figures like the founding fathers, you ask AI to generate depictions of evil people, who historically were white. Now the groups who claim "you can't be racist against white people" suddenly care about the historical inaccuracy. The trolls turn their enemies against each-other.
4. Your example with making the black nazis image manually misses an important point: the prompt used wasn't "a picture of black German soldiers in 1945". It was "a picture of German soldiers in 1945". If you ask an artist to draw you an artwork, which you describe as the latter, and he creates a picture of black soldiers, I think you just wouldn't pay and consider the contract to not be fulfilled.
I tried Firefly for a month. I had the same experience that I did with Dall-E. At first you think, "wow". But then when you try tro wrangle them into completing a specific task, the "wow" quickly turns to "why?". I spent more time refining prompts than I would have just designing those things I was after myself. And this has nothing to do with polical correctness, these models just don't work very well for production.
20 comments
[ 3.5 ms ] story [ 58.3 ms ] threadThere is also a difference between error and purposeful misrepresentation.
this is not true in California
source: Dr Dobbs
This is untrue. Adobe has been infected, and long-time employees have been leaving left and right because they no longer feel welcome there.
Please don't take this comment as an ideological or political comment; I'm only trying to outline a flaw that I perceive in your reasoning. I would write a similar comment if we were discussing an apolitical, dry mathematical topic.
I would expect these models to misrepresent reality until we do the near-impossible: convince a large and diverse swath of people to understand and work with each other, work with ML models, and contribute to shared datasets that we can train our models with. I have vastly oversimplified what it would take to create an accurate world model, however, the main point is there. We need feedback from everyone. And likely, feedback from more than just people.
Why would Google/Adobe want to artificially increase diversity in its training set? Well, there is a lack of diversity in the training set in reality, so the developers are likely trying to balance something that can easily be seen as racist or prejudiced. Given America's history of how it treats people with different cultures, religions, or skin colors, this should not come as a surprise. Other countries have made similar errors in their ways throughout history.
Why would we expect Gemini or Firefly to not be able to produce these images? I can definitely sympathize with not wanting to make offensive images but there’s no way these tools could be sanitized to not produce something offensive given the right prompts. Just like how you can’t stop photoshop users from creating similar images manually.
1. Gemini and Firefly are programmed (by a prompt) to give you "inclusive" results when you ask for depictions of people. So if you ask it to generate a crowd of people, while normally the AI could generate mostly or purely white crowd (perhaps because white people dominated the training dataset), with such a prompt it will try to give you people of both genders and all races.
2. Because of the above, in some context where you would expect only white people, for example the founding fathers of USA, it also depicts people of various races and genders. This is historically inaccurate and produces a lot of controversy, especially in times, when doing the reverse, depicting a historically non-white person as white, usually leads to some kind of outrage. And so such behavior is put to the same category as e.g. claiming "you can't be racist against white people".
3. And as a result, many people decide to "fight back" by trolling the creators of the tool, using the bias where instead of being negative towards white people, it is negative towards other people. So instead of taking positive figures like the founding fathers, you ask AI to generate depictions of evil people, who historically were white. Now the groups who claim "you can't be racist against white people" suddenly care about the historical inaccuracy. The trolls turn their enemies against each-other.
4. Your example with making the black nazis image manually misses an important point: the prompt used wasn't "a picture of black German soldiers in 1945". It was "a picture of German soldiers in 1945". If you ask an artist to draw you an artwork, which you describe as the latter, and he creates a picture of black soldiers, I think you just wouldn't pay and consider the contract to not be fulfilled.
I end up just mashing something up and feeding it to MidJourney with an image weight parameter along with my prompt.