Missed a Spot

This article in The Washington Post, What AI thinks a beautiful woman looks like. is emblematic of much of the debate around the potential social impacts of generative automation tools, and how they, by constantly regurgitating broader social biases, can be instruments of perpetuating those same biases.

But something stood out for me when looking at the procedurally generated portraits that had resulted from the authors' prompts.

Several of the prompts started with the phrase "A full length portrait photo." This is a fairly standard term, and it refers to a photo that is, well, full length. The subject's entire body is in the frame, with no part of the head or feet cut off or cropped. Many of the images shown don't meet that criteria. While it's certainly possible the The Washington Post cropped the images for best effect, many of the ones presented are full length, so it seems odd that the Post would have needed to crop only select images for clarity.

And that raises something of a dilemma. If the Post cropped images that were full length when first generated, how do we understand that we're seeing what was given to the reporters? There's no indication in the article that any alterations have been made, so it can be difficult to say what may or may not have been changed about them. But if images aren't cropped, then it's pretty clear that these tools are still having difficulty with the "follow basic instructions" part of generating images. In that light, a lack of diversity in skin tones, age and body type in the images generated may not be the biggest problem that needs dealing with.

And this goes back to my current knock on calling these systems "artificial intelligence."

Given the size of these datasets, makers of generative systems are not going to be able to go in and make sure that all of the images are captioned or tagged in ways that will work for everyone. It would simply be too time consuming and two expensive. And so this could simply be a case of "garbage in, garbage out." And if it strikes one that current standards of beauty are garbage, perhaps the results stand to reason.