AI will eat itself

For the Dungeons & Dragons group I run, I’ve started using Midjourney to create images illustrating the weird monsters the players encounter in their adventures. I also use it to create images of NPCs (non player characters) who the group meets along the way. Two of the people in our group queried why the female NPCs our party encounters are always beautiful, full breasted and often as not, wearing slightly revealing clothing. I was at a loss to give a reason. I didn’t request the images to be of attractive, buxom or scantily clad women. The only thing I can think is that the images poured into AI art tools like Midjourney have a preponderance of images that are all of the above. Similarly unless you specify the ethnic background of the person you want a picture of, I’ve found it will almost always give you a picture of a white person.

So how does this bias come to exist? Is it that the images fed to programmes like Midjourney are somehow chosen by people with a bias, either conscious or unconscious, for certain types of image? Or do these images tend to be more common in the world, therefore creating a numerical bias?

Either way as AI images, and for that matter written words, start to dominate the online world, AI bots are inevitably going to start feeding on the things they have created. This thought-provoking short video illustrates what might happen when AI images start to eat themselves. The point is that the data will eventually start to aggregate, alighting on a median average of everything it’s been trained on. So anything further away from the median will get lost, or just drop off. What pernicious effects this could potentially have on society as AI generated words and images become more prevalent are anyone’s guess. I feel like text and image curation needs to be part of the discussion as nations grapple with tech companies to keep AI tools in line.

I read about a brilliant image-to-text-to-image feedback experiment a while back in which someone asked an AI to create an image of the perfect pizza. Then describe what was in the image. Then asked it to create an image based on the description. Then asked it to describe the image. And so on and so on, until eventually a man’s face started to appear in the pizza. What strange chemistry is at work here when a human face begins to appear in your quattro stagioni? This seems like something of the ‘black box’ effect where AI just does its own weird thing and how it gets to the result is fairly opaque.

Speaking of pizzas, if anyone is worried that AI is ready to replace humans at making TV commercials, rest assured, as this post from May demonstrates.