- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
Google is coming in for sharp criticism after video went viral of the Google Nest assistant refusing to answer basic questions about the Holocaust — but having no problem answer questions about the Nakba.
So if someone wrote a prompt to make an image of a black woman as a pope, would you expect the model to only return historical popes?
If the model is supposed to be able to make both historically accurate and possibilities, why would the expectation for a vague prompt to be historical instead of possible?
If the model is supposed to default to historical accuracy, how would it handle a request for a red dragon? Just the painting named Red Dragon, dragons from mythology, or popular media?
Yes, there is could be something that promotes diversity or it could just be that the default behavior doesn’t have context for what content ‘should’ be historically accurate and what is just a randomized combination of position/race/gender.
Of course it will draw black female pope if you request, but if you do not - it would not. As a gross approximation, ANN is an interpolator of known data-points (with some noise), and if you ask simply a pope, it will interpolate between the images it learned of popes. Since all of them are white male it is highly unlikely for ANN to produce black female (the noise should be very high). If you ask black female pope, it would start to interpolate between the images of popes and black females. You have to tune the model so that when you ask just for pope, something else pushes the model to consider otherwise irrelevant images.
Would expect a lot of models to struggle with making the pope female, making the pope black, or making a black female a pope unless they build in some kind of technique to make replacements. Thing is, a neural net reproduces what you put into it, and I assume the bias is largely towards old white men since those images are way more readily found.
Even targeted prompts, like a zebra with rainbow colored stripes, had very limited results 6 monts ago where there would be at least 50% non black and white stripes. I had to generate multiple times with a lot of negative terms just to get close. Currently, the first generation of copilot matches my idea behind the prompt.
Clearly the step made was a big one, and I imagine tuning was done to ensure models capable of returning more diverse results rather just what is in the data set. It just has more unexpected results and less historically accurate images for these kind of prompts. And some that might be quite painful. Still, being always underrepresented in data sets is also quite painful. Hard to get to a perfect product quickly, but there should be a feature somewhere on their backlog to by default prevent some substitutions. Black, female popes when requesting a generated pope? To me that is a horizon broadening feature. Black, female nazis when requesting nazis? Let that not be a default result.