Day 2: What happens if AI eats itself?
This is part of a series where I post every day for 24 days. See Advent of Writing: 24 days of posts.
On the 2nd day, I thought about the impact of AI. A few days ago, a new AI chatbot ChatGPT was released that can do some pretty mind-blowing stuff. This comes only weeks after other huge advancements in text/image generation.
I personally think AI is going to be a huge fundamental shift. It’s a whole new kind of tool that impacts how everyone works. The question is what it will look like.
I think a lot about how creators will be affected. An immediate example is artists: when someone uses Stable Diffusion to generate 100 incredible images, isn’t that scary?
Things are gonna change, but I don’t think everyone is going to lose their job; it’ll be another powerful tool to use. But still, it sucks that your skill with oil painting is diminished, right?
Maybe that won’t happen, at least for a while. Here’s one outcome: AI produced images actually get worse over time. As that happens, the value of original skilled creators actually goes up as business seek out people who don’t use AI.
Why would it get worse?
- AI works by training a model on a bunch of content from the internet
- The output of AI is mostly good, but there’s still small artifacts. Especially if the prompt doesn’t play 100% nicely with the model. Most people aren’t using perfect prompts, so most output is going to have small artifacts. Let’s give it a number: the output of AI is 95% the quality of an original artwork.
- Millions of people use AI to generate art and upload it to the internet.
- AI continues to be trained on images from the internet.
At each stage of training, the set of images it uses gets worse because it’s including all the AI generated images as well.
This probably won’t happen this clearly, but it’s interesting to think about. As AI starts being trained on itself, what will the output be?